station-master/CHANGELOG.md

Changelog

Detail behind each commit. Commit messages stay high level; the reasoning, the measurements, and the things that turned out to be wrong live here.

Measured figures are 100 solitaire Standard games with the developer bot unless stated otherwise. The target is 20 Revenue over 5 Days.

Versions

Every commit carries a bump, decided with Jesse rather than assumed:

• third digit — bug fixes to what is already there.
• second digit — a new set of features.
• 1.0 — the first release we think is solid enough to call one.

The number lives in package.json and nowhere else; scripts/build-web.ts stamps it into every page as v0.1.0 · <sha> · <date>, so what is deployed can always be identified from the page itself.

---

Unreleased

0.4.0 — 2026-08-13

Phases 0 and 1 of the multiplayer plan — the seams, not the server

docs/architecture/multiplayer.md is the plan; this is its first two phases. Nothing a player can see changed, and that is the exit criterion: the whole point is that solitaire is byte-for-byte the same game while the code underneath it stops assuming there is only ever one of you.

Seat and player are now different things. They were the same integer everywhere, which is correct today and wrong the moment Employee Rotation moves someone to a different chair. Offices and districts are keyed by SEAT; hands, Revenue and the Fedora belong to the PLAYER. seating[], seatOf and playerAtSeat make the mapping explicit — it is still the identity mapping, so nothing moves yet.

The split immediately found a real bug: awardDeparture was indexing s.players with a seat. Correct under the identity mapping, silently paying the wrong railroad under rotation, and unfalsifiable in solitaire. That is what the change is for.

Per-player turn state. turn: TurnState became turns: Map<PlayerIndex, TurnState>, one per player at phase entry, read through turnOf(s, player) at ~50 sites. Behaviour-neutral by construction — the phase cursor still walks one player at a time, and every existing test passed unchanged. It is done now rather than in Phase 2 because it is a wide, mechanical change to the engine that costs almost nothing today and would be a rewrite once a wire format depends on the shape. What it buys later is players acting in parallel where the rules allow it, which is where the latency in a turn-based game over the internet actually lives.

The page renders from Frame + Menu alone. main.ts had eleven reads through into GameState — the deck, the RNG, other players' hands — each one a thing a server would never send. Frame grew option, status, outcome, players[] and handCount to cover them. A test now reads main.ts and fails if a reader comes back, because the cheap moment to catch that is now and not in Phase 2.

A Session between the page and the game. LocalSession runs the engine in the browser exactly as before; a RemoteSession will hold no authoritative state at all — it cannot, having neither the deck order nor the other hands. So the interface is deliberately the smaller of the two, and what only a local session can do (undo, local save, dealing a new game) is declared in capabilities, which the page reads to hide those controls rather than calling them and failing. submit is async even though the local one answers immediately: a page written against a synchronous submit would have to be rewritten for the server.

test/session.test.ts is the proof — 13 tests, the first of which plays seed 77 to a finish through both routes and asserts the boards and the histories are identical.

Docs. protocol.md was written from the rules before the engine existed and had become a second, drifting copy of intents.ts and events.ts; it now points at them and keeps only what the types cannot say — what is deliberately not an intent, the two loops a client must not flatten, and the redaction surface. Three architecture documents still said WebSocket where D5 chose SSE + POST; fixed.

Then a review found the seat/player audit was half done, and it was right. Two flagged sites, and sweeping for the pattern found six more. All of them are correct today and all of them break the moment Employee Rotation lands, which is exactly the failure mode the split was supposed to end.

Keyed an Office Area by player when officeAreas is keyed by seat: refusesThisOffice (apply.ts — the Crack Limited's terminals-only rule), spendDispatchBonus (advance.ts — the Fedora is held by a PLAYER, the Telegraph is installed in a SEAT), impediments (narrate.ts), and two district lookups in the bot. adTrackCount(s, player) in narrate.ts was passing a player into a SeatIndex parameter — the type said seat and the caller said player, and TypeScript cannot tell two number aliases apart, which is precisely why this needs a test rather than a type.

Read seat 0 regardless of the viewer: the bot's takingRank ranked a face-up Office card against seat 0's Office tier for every player, so in a multi-player bot game everyone chased player 0's upgrade. stats.ts and save-replay.ts also index seat 0, which is correct — they are solitaire summaries — and now say so through areaAtSeat instead of looking like the same bug.

And Frame was seat-scoped in the board and the hand but not in four fields beside them. revenue, moves, blocked and the pace note all still reported player 0. A player would have been shown someone else's score, someone else's Moves remaining, and someone else's jammed facilities — that last one a list of squares that are not on the board they are looking at. Worse, the crew lookup was keyed by row,col alone while every district shares an origin, so a crew standing at (0, 1) in one Office Area was drawn at (0, 1) on every player's board.

Five new tests, all of which fail against the previous commit: per-seat Revenue/pace/impediments, crews staying on their own board, a full 3-player game played with seating rotated before the first move, impediments following a player to their new chair, and a source-level guard that fails if anything outside state.ts/apply.ts touches officeAreas directly — the class, rather than the eight instances of it found by hand.

Bot unchanged at 7.0 mean over 200 solitaire Standard games (the bot fixes are behaviour-neutral with one player), and all four published replays still play to the end — which, for a change of this width, is the whole report.

0.3.1 — 2026-08-13

Groundwork for multiplayer

Two steps that hold whichever way the hotseat-or-server question goes, plus a deck change.

The 22 opponent-directed cards are out of every deck, not just solitaire's. Q6 removed them from solitaire because they have no legal target with one player; they are out of the competitive deck now too, because checkPlay answers both categories NOT_IMPLEMENTED and dealing them would make ~9% of draws reject outright. buildDeck returns 213 in both modes. DECK_SIZE (235) is now documented as the CATALOGUE rather than the size of any deck in play — the two had quietly become different numbers. Three Enhancements (Facing Point Locks, Water Column, Overpass) exist only to answer these cards and stay dormant until they return; recorded in TODO.md as multiplayer work.

A net under the multi-player engine paths — test/multiplayer.test.ts, 13 tests. Everything else in the suite is solitaire, so these had been running unwatched. They cover 2/3/4-player games playing to a finish; per-seat Office Areas and Crew Tray counts; the Fedora passing to the next seat, only on a shift boundary, reaching every seat over a game; Subdivisions splitting at the Offices that have upgraded and not at the ones that have not; one player's oncoming train barring another's departure, and no longer barring it once a Control Point puts them in different Subdivisions; and that awardDeparture pays the Office that ran the train rather than seat 0 — which with one player was unfalsifiable.

Two failed when first written and both times the TEST was wrong, not the engine: the Superintendent cases drove the clock by calling advance in a loop, which never moves it — advance stops and asks for input, so the game has to actually be played. The engine's multi-player paths passed everything on the first honest run, which is the useful result here.

snapshot and actionMenu take a seat. They were hardcoded to player 0 in eight places across view.ts, game.ts and main.ts — correct with one player, and a quiet disaster with more: every seat would have been shown player 0's railroad including player 0's hand, which is the one thing the state model calls secret. All eight now take a viewer, defaulting to 0 so solitaire and every replay are untouched, and describeIntent now describes an intent against the acting player's district rather than seat 0's. A test shows three seats getting three different boards and three different hands, so the parameter is exercised rather than merely present.

Bot unchanged at 7.0 — solitaire never dealt the opponent cards, so none of this moves it.

0.3.0 — 2026-08-12

Sharp curves are out of the deck

Eight cards, dealt zero copies. The only thing that made a sharp curve different from an ordinary one was moveCost: 2, and nothing ever charged it — every switching move costs exactly 1, hard-coded — so they were geometric duplicates taking eight draws from a deck the rebalance already considers too diluted. track.test.ts even had a case called "a sharp curve differs from a curve only in the Moves it costs" which asserted the two were identical, describing a difference that did not exist.

Jesse's call: take them out rather than build the Move cost, since a per-card movement cost is a change to the Move model and the rebalance can wait. The rows stay in the catalogue at zero, exactly as Poling does, so the design is still visible and the geometry still works if they are ever dealt again. Deck 243 → 235, track 104 → 96, solitaire 221 → 213.

Four tests broke on the new deck composition and none of them was wrong about the game — all four were resting on a single seed or a threshold that had drifted, so all four are now measured properly:

• The oscillation detector ran on one seed. Measured across sixteen: thirteen games show no aimless shuttling at all and three reach a run of five. So it is a minority behaviour, not the every-game waste the test was written for — it now asserts a rate (most games clean) plus a ceiling.
• Interlocking reaches the board in 7/60 games, down from 15/60. The bot has been pulled toward other work by departure Revenue and spends its opening on the track it was dealt. Floor lowered to match, and recorded as drift rather than quietly restated.
• The action list reaches 9 buttons, up from 8 — all of them Switching, because the opening deal grew districts from 17.9 to 20.3 cards and a crew has more squares it can legally reach. The list getting longer for a good reason, not the cross-products that test was written to kill.
• A test hoped seed 111 would deal an Office upgrade. It now puts one in hand.

Bot after all of it: 7.0 mean, median 5.0, wins 5/200, districts 18.8 cards.

The special-train rules are enforced

All nine. They had been declared on TrainRules and read by nothing, so a Crack Limited could shunt an industry and a Military train could be worked by Porters — the restrictions are what make a special train special, and none of them applied. Each is checked where the act happens, and a LOCAL CREW (trainNumber: null) is exempt from every one of them: it has no card, so it has no rules.

• noSwitching (both expresses, Light Engine, Campaign, Circus, Military) — moving, setting out and sorting are all refused. All three are switching.
• dropOnly (X13 Appleseed) and pickUpEmptiesOnly (X22 Pee-Dee) — enforced on the MOVE, not on a coupling action, because coupling is mandatory (§A.4): there is no running over cars and leaving them, so the restriction has to bite on the move that would pick them up or it cannot bite at all.
• oneFreightPerLocation (trains 3/4 Express) — a budget per SQUARE rather than per turn, which is what "one freight car at every location" says: the Express works a car here, moves on, works another there. Dropping and picking up share the one budget, because the card says "drop OR pick up". Kept in turn.freightWorked, keyed by tray and square, cleared with the turn.
• noPassengerWork (Military, Director's car) and terminalsOnly (Crack Limited) — Porters refuse the train. canBoard/canDetrain now filter the trains at the Office rather than taking any of them, so the rule actually bites; a new passengerRefusal works out whether a card is the reason and says which, because "no train at the Office" was both wrong and unhelpful when there was a train standing right there.
• stopThenExpedite (X17 Campaign) — the speeches happen at the first Office it reaches: that arrival is an ordinary stop, and every arrival after it is expedited. A speechMade flag on the tray, alongside stopPointClaimed, for the same reason — it is the TRAIN that stops, and an Extra runs once.

coachStaysOnStationTrack could not be built as intended, and the reason is worth recording. The chosen reading was "the coach may only be set out at the Office", but §A.4 refuses the Office square to every drop — "the Office track is Operational Rail, but Rolling Stock may not be left there" — so "only at the Office" and "nowhere at all" are the same rule. It is enforced as the effect that survives: the Local may shunt its freight car around the district and may not abandon its coach doing it. If the station track is meant to become a real place to leave a coach, that is a change to §A.4 rather than to this card; noted in TODO.md.

copiesNextScheduled was deleted rather than implemented. No train card ever carried it: a Second Section is a Maneuver card played on a train due out, with its own intent (newTrain.secondSection) that has worked all along. It was an unreachable second description of a mechanic that already existed, and it had been sitting in the "nine rules read by nothing" count as the one entry that needed removing.

Cost, measured over 200 games: revenue 8.0 → 7.2 and wins 14/200 → 6/200. Restrictions make the game harder, which is the point — but it is worth knowing that enforcing §7 took back about a sixth of what the departure-Revenue rule gave. The train tooltip no longer says "NOT YET ENFORCED BY THE ENGINE"; it says what each rule does, because the restriction is the character of the card.

Two rule changes, both provisional, both measured

The opening deal is now 3 track + 3 other, from two separately shuffled piles. Track is shuffled apart from the rest, each player is dealt three of each, and the leftover track is shuffled back in for the rest of the game — the split is an opening-deal device only. A player opens holding six against a hand limit of three, so the first turn is spent choosing which district they can afford; §6.2 already permits "any or all" cards to be played in a turn and puts no cap on discards, so the reduction can always be made and the engine needed no special first-turn case.

This is the answer to the run-around problem, which was measured five ways and is a SUPPLY problem: a turnout and a matching-hand curve were in hand together on 0.2% of turns, about once every eight games. It worked, modestly — run-arounds 4/60 → 7/60, districts 17.9 → 20.3 cards — with revenue unmoved on its own (−0.1, inside noise). Nowhere near the 91/100 of the private-supply era, so the supply question is softened rather than answered.

One Revenue for every train that clears your section. Paid when a train is highballed out of your Office, once per train, and not again when it later runs off the end of the Division — that far Division Point belongs to whoever is seated at it. A local switching crew has no train number and earns nothing. Worth exactly +5.40 a game: departures run at 5.40 and each pays 1. The bot goes 2.7 → 8.0 mean, median 0 → 6.5, wins 2/200 → 14/200. Unlike freight and passengers it pays for traffic the player does not have to work, which is the point — keeping the line clear is the Superintendent's job and nothing paid for doing it well. It also makes the victory target live: 20 over 5 Days is now reachable largely on traffic.

A regression that came with the deal, recorded rather than hidden. Forced to shed on turn one, the bot lays pieces it should discard: track butting a card that cannot accept it went from 7% to 15%. A player would simply discard the ones with nowhere good to go. The regression floor in sim.test.ts was raised from 0.15 to 0.25 with the reasoning written in, and it is in TODO.md as bot work for after the rebalance. It also means part of that district-size gain is padding.

Found while fixing the tests this broke. Three of them were passing on luck rather than checking anything, and the deal change exposed all three at once: the commodity test asserted a tank car is set out somewhere in 30 games when it happens in 3% of games; the replay-sound test asserted six cue kinds from a single seed when the bot couples in only 20% of games and sets cars out in 43%; and two web tests depended on a seed happening to deal a particular card. All four are now decisive — pooled seeds, a sample sized to the base rate, or the card put in hand deliberately. Separately, moveTrain's branch for departing an Office straight onto a Division Point turns out to be unreachable: buildDivision always lays DP · Mainline · Office · Mainline · … · DP, so an Office is never adjacent to one. It is kept correct rather than deleted, and setup.test.ts now asserts the flanking invariant that makes it dead.

All five published replays were re-recorded: the deal changes what every seed deals, so the previous recordings died at 2 intents of ~350.

Nine things playtesting found

Eight fixes and one new rule, all from one session at the table. Two of the nine turned out not to be bugs at all, and saying why is most of the work in those.

Left and right were on the wrong diagonal — for turnouts and for curves, the same way. The engine called a card left when a train entering at its points saw the diverging route leave to its right: {stem:'w', through:'e', diverge:'s'} heads east, and south is the driver's right. Curves inherited the same inversion, because a curve takes its hand from the turnout whose leg it continues. Every track card a player has ever held was labelled as its mirror. The artwork was never wrong — board-svg.ts draws rails from connectionsFor geometry alone — so this is words only, and the fix flips the hand field on the TRACK_CARDS rows together with CURVE_VARIANTS and TURNOUT_VARIANTS. The row ORDER is deliberately left alone: setup.ts builds the deck by walking that array, so a row's position decides which physical card a seed deals, and reordering to look tidy would silently re-deal every published replay. Verified by fingerprinting seed 1234's whole track deck before and after — 101 cards, byte-identical, and the three replays that survived the rest of this work prove it in practice.

A turnout can now be laid on top of a card already down. Reported as: a district can only hang off a turnout, so a player who laid a straight along the main and then wanted to branch there had no move at all — the piece had to have been a turnout when it went down. A turnout may now upgrade a straight at any rotation, or a curve whose arc matches its own diverging leg. Both are strict port supersets of what they replace, so an upgrade can never sever a join a neighbour relies on and needs no connection test; the new leg may reach nothing, which is the point. Blocked on a standing car (UPGRADE_OCCUPIED) and on a built Interlocking or Telegraph (UPGRADE_ENHANCED) — both about the card being in use rather than its shape. The rule is stated in ARCS rather than hands so the flip above cannot reach it. The bot found and played it 40 times across 40 games without being taught anything.

A Depot showed three MEN | AT | WORK boxes it has no Laborer to work. An Office is a Passenger Facility and the sign is printed "For Freight Facilities" (§9.1), but every facility got a three-slot array of nulls and the renderers loop it — the green and red rows either side were guarded and the MAW row was not, in all three renderers. Fixed in the MODEL rather than the renderers: menAtWork is now null on a passenger facility, which is what the "Freight only" comment on that field has claimed all along. The type flushed out every reader, freight work is now refused on the shape of the facility rather than on it happening to hold 0 Laborers, and a fourth renderer cannot reintroduce the bug. Two more artifacts of the same "an Office is a Facility" modelling went with it: a Depot read SHIPS + RECEIVES, an industry's flow word, and was drawn a siding square although its industry track has length 0.

The side panel painted inbound loads green. Its shared box helper used one class for every filled box, so the red row rendered green while the board SVG on the same screen drew it red — two views of one card disagreeing about the colour code at the same moment. Green is outbound everywhere now, red inbound, grey for the siding, which is a place rather than a direction.

The card just drawn now sits first in the hand, badged NEW. The engine pushes onto the end, which with a wrapping row put the card you just turned over wherever the eye is least likely to be, among two others that look exactly like it. Reversed in the DISPLAY — actionMenu for the play page and snapshot for both replay viewers — and deliberately not in the engine: the bot iterates its hand to generate options, so moving the stored order would reshuffle its tie-breaks and invalidate every revenue figure in TODO.md. Confirmed: 2.8 before, 2.7 after, over 200 games.

Two Ice Houses could be built in one district. Industries have been barred from doubling up since Q4, but a Modifier is a different card kind and had no such check at all. One of a kind per Office Area now. Enhancements are deliberately left alone — an Interlocking is a plant at one junction, so a second on another straight is a different installation. This is the change that killed a published replay: seed-4894942 had recorded a game that played two Waiting Areas, so it was a recording of illegal play and has been re-recorded.

NOT A BUG — "the Ice House added the laborer but not the outbound slot". The card prints +1 outbound; a Grocer's Warehouse is flow: 'inbound', so allows.outbound is false and the capacity was being raised on a direction that can never be drawn or stocked. The laborer landed because Laborers have no direction. An industry's printed flow stays absolute — no modifier turns a receiver into a shipper — so the grant is now dropped rather than credited, and the reason is said out loud in both directions: the card in hand names which of its printed hosts cannot use which half, and the facility panel reports a suppressed bonus instead of quietly showing a number that did not move. The same trap catches Truck Dock and Forklifts on a Grocer's, and Waiting Area, Restaurant and Hotel on a Whistle Post. Two things fell out of this: TrackCard.modifiers was initialised everywhere and appended nowhere, so the panel's "Modifier cards standing beside this industry" row had always been empty; and a Waiting Area was lengthening the Office's industryTrack, which drew a Depot a siding to spot cars on — the phantom siding above, arriving by another door. That one was found by rendering 12,000 frames of real games, not by a fixture, and it is why the smoke pass exists.

An Interlocking on the board was a bare label. It now carries hover text from the card catalogue, and the one card nothing reads says so. Seven of the ten resolve in play: the dispatch chain (Telegraph, Telephone, Radio) plus Interlocking, which holds an arrival at the Limits when the Office is full; Yard Office, which diverts a coachless train; Small Yard, which re-orders a consist; and ABS Signals, stored on the Mainline node rather than in enhancements[]. Facing Point Locks and the Water Column are wired and read but answer opponent-directed cards a solitaire deck omits. Overpass alone has no code path anywhere, and is the only one that carries the warning.

Got this badly wrong first time, and it is worth recording how. The table was filled in by grepping for four helper function names — dispatchBonus, hasDistrictEnhancement, isProtectedFromDerail, watertowersRemovable — and reading "no match" as "no implementation". But Interlocking, Yard Office and Small Yard are read by key in advance.ts and apply.ts, and ABS Signals through node.absSignals, so five of ten statuses were wrong and the tooltip told players that four working cards did nothing — worse than the bare label it replaced. enhancements.test.ts had covering tests for all four, passing, the whole time. The test written to guard the table made it worse rather than catching it: it asserted effect === 'live' if and only if the rule carried a dispatchBonus, which restates the assumption that produced the error instead of checking it. It now asserts the three sets by name, and every row cites the file that reads it, because a status without a citation beside it is a claim nobody checked.

Three rules read back from the sheet

Jesse's follow-up after the above. Two were already right, which is worth recording so nobody re-investigates them; one was a real gap that had been sitting in plain sight.

Clearance asked about the next CARD, not the next Subdivision. §8.1 is explicit — "if there is a train in the next Subdivision moving towards the considered train, the considered train will not depart" — and evaluateClearance only ever inspected node.transits, the single card being entered. A train ran headlong into a Subdivision an opposing train was two cards deep in and was stopped only on the Stage they actually met. subdivisions() had been sitting in state.ts for this the whole time, called by nothing outside a test. Red Flags and ABS Signals now read the card the OTHER train stands on rather than the card being entered; those were the same card while this only looked one ahead, and the protection belongs where the train it protects actually is. Worth knowing how much this bites: every Office starts as a Whistle Post, so the whole railroad is ONE Subdivision until someone upgrades, and each upgrade to a Control Point splits one in two and buys capacity back. The bot is unmoved (2.7 both sides, 200 games) because solitaire schedules only ~1.3 trains a game — this is a rule that earns its keep with 2–5 players.

Straight-placed enhancements replace the straight, so they no longer stack. An Interlocking's printed placement is "any Running Track Straight": it goes down IN PLACE OF the straight, and what stands there afterwards is an Interlocking, not a straight carrying one. Nothing checked it — an enhancement only pushes a string onto enhancements[] and leaves the geometry alone — so a single straight could hold Interlocking, Telegraph and a Water Column at once. The Telegraph → Telephone → Radio chain is untouched and is not an exception: Telephone prints "on Telegraph" and Radio "on Telephone", so those target a named card rather than a straight, which is exactly why they still stack. The printed placements settle it; requiresOnSameCard already enforced it.

ALREADY CORRECT, twice. Control Points and Subdivisions were modelled properly — isControlPoint is false only for a Whistle Post, and subdivisions() splits between them — so only the clearance use of them was missing. And §8.2's "moved immediately to the Office, where it stops for orders" was already the behaviour: arriveAtOffice returns moved, and the caller adds the train to movedThisPhase. The one exception is deliberate and recovered — Q3, an expedited train departs the Stage it arrives, and gets a second moveTrain still subject to §8.1.

A missing §8.1 condition, restored to the ruleset. Jesse read the highball conditions back off the source sheet and the first one — "the train did not just initially arrive from a mainline card at the Office this Stage" — was absent from rules-v0.2.md entirely. The BEHAVIOUR was right, but only as a side effect: the Mainline Phase visits each train once per Stage in numeric order, so a train that spends its visit arriving has no visit left to depart with. Nothing anywhere stated the rule, which makes it precisely the sort of property a later change to that loop breaks in silence. Now written into §8.1 and pinned by tests, together with the observation that this is the condition the printed Expedite rule exists to override — without it, Expedite was a special case with no general rule to be special against, which is why Q1/Q3 found it so hard to place. All six conditions now have coverage, including the Gap 2b carve-out: a train standing clear on Secondary Track in a Whistle Post district must not hold up the Subdivision, which the new subdivision-wide scan had to be careful not to break.

Verification. 438 tests (up from 416) and a clean tsc. Beyond the unit tests: the track deck fingerprint before and after the flip; all published replays replayed end to end; the bot baseline re-measured against a worktree at HEAD; and 40 real games rendered frame by frame through the same functions the page calls — 12,169 frames, which is where the Waiting Area siding surfaced. Not done: nobody has clicked through this in a browser.

0.2.0 — 2026-08-09

The bot plays twice as well, and a rule it was never following

Adopted, after measuring: cap the trains at the Office's A/D capacity, and operate before drawing. Together +1.52 ± 0.17 (t = 9.16) over 1600 paired seeds — revenue 1.19 → 2.72, wins 11 → 21 in 1600, collisions 0.38 → 0.02. Both are now default play; the flags that carried them are gone, and what remains in BotTweaks are two ABLATIONS that turn them off, because the question a measured heuristic needs later is "is this still true?" rather than "does this help?".

The six candidates that measured neutral are deleted rather than left switched off. Adding all of them on top of the two winners was worth −0.08: +1.44 against +1.52 without.

Jesse's idea — build first, then switch — was right about the axis and wrong about the half that mattered. Reserving Day 1 for development measures +0.52; reserving none and simply preferring the work to the draw measures +0.48 with a far better spread (317 seeds better against 77, where the build phase gives 329 against 144). With the cap alongside, no build phase beats one Day beats two. So the keeper is "operate rather than draw", and the building phase — the intuitive part — is inert: the draw-priority ordering was already developing the district well enough.

It works through the constraint the funnel had been pointing at all along: green-stocked Cargo phases +60%, freight revenue +39%.

And then the same tooling proved that constraint was the wrong one. Stocking green boxes speculatively — filling them before a car is spotted — raises stocked phases five-fold, from 5.6 to 28.8 of 60, and moves freight revenue by nothing at all (1.16 → 1.15, and −0.10 revenue overall, t = −3.62). The green box was never the gate. The gate is the spotted car, and the 8% figure was low because stocking is conditional on a car being there — it was measuring the car supply all along. Switching quality, not Freight Agent turns, is where the freight economy is won.

The engine was not following two rules it prints

Chasing the Rolling Stock census found the cause, and it is not a modelling ambiguity after all. §9.3 says an unload requires "an empty car of that type in the Division Yard" and §9.2 says de-training requires "a white empty coach in the Division Yard". Neither requirement was checked, and both reducers conjured the replacement car instead of taking it — so every unload and every de-training minted a car, 1.29 a game against a supply of 80.

Both now take the car from the Division Yard, as the rules say. The census is conserved: censused after every batch across 80 games, nothing changes it but a collision, where before it drifted to 119 against 80. A test does that check now, because this class of bug produces no symptom until a supply number is tuned against it.

It costs the bot about 0.09 revenue — unloading is meant to consume supply — and that is the point: ROLLING_STOCK_SUPPLY can be tuned against reality now.

Where the game actually stands

With a bot that no longer wastes revenue, the balance question resolves. Revenue is linear in trains scheduled at about 1.9 a train, and trains are capped by A/D capacity, which is the Office tier, which is a card you have to draw. 37% of games never leave the Whistle Post; 53% earn nothing at all; the median game scores 0 and the best of 800 scored 26.

Twenty Revenue would need roughly eleven trains and eleven A/D tracks. A Terminal has four. The target is not missed, it is unreachable — and that is now a deck question with numbers behind it rather than a suspicion. TODO.md carries the three ways out.

Five ways to teach the bot to plan a siding, and why none of them can work

Asked directly whether the bot could think across turns and stop building dead-end stubs. It can be taught to; it does not help, and finding out why was worth more than the attempts.

attempt	result
hold ALL track for the siding	−0.70 (t = −3.27)
hold only CURVES — the only piece that can climb back to the main	−0.26 (t = −3.22)
finish an open run before cutting another way down	0.00 — 398/400 identical
treat a second turnout as the piece that closes the loop	0.00 — 400/400 identical
spend a curve only on a square that actually closes a run	−0.11, 15 games in 400 differ

The pieces never meet. Over 12,000 Local Operations turns, a turnout and a curve are in hand together on 0.3% of them, and a turnout with a MATCHING-hand curve on 0.2% — about once every eight games. A run-around needs five specific pieces of the right hands arriving in a usable order, and the bot does not reach the two-piece prerequisite, let alone the fifth.

It is not hand pressure, which is what the first two attempts assumed. The hand is full: mean 2.66 cards, at the three-card limit on 78% of turns. The bot plays 11.4 track cards a game against 1.5 discarded, spending each piece as it arrives because a piece that builds something now outscores holding one that might build more later — and the measurements say it is right to.

So the dead-end stubs are not a planning failure. They are what a five-piece structure looks like when the pieces are drawn one at a time from a 243-card deck: 91 run-arounds per 100 games when track was a private supply the player chose from, 29/100 once track was drawn, 4/60 today. TODO.md carries the options, and all of them are deck changes rather than bot changes.

Ten ways to make the bot better, and one of them works

Every heuristic below was measured paired over 400+ seeds, and confirmed at 1600 before being believed. Nothing is adopted yet — every flag is off, so the bot plays exactly as it did.

The only thing that moves revenue is refusing to schedule a train the Office cannot hold. trainCapSlack=0 — never more committed trains than A/D tracks — is +1.09 ± 0.16 (t = 6.79), taking the bot from 1.19 to 2.28 mean. Every other reordering of the bot's preferences measured inside the noise. Adding all six of the small ones on top of the cap is worth a further 0.10.

Three failures worth more than the success.

Refusing to bury the engine costs 0.35 a game (t = −2.98). The tweak works perfectly — burial falls from 8.4 decisions a game to 0.03 — and freight halves along with it. Coupling is mandatory (§A.4), so the moves that bury the engine are the moves that pick cars up. Burial is the price of collecting, not a mistake to be coached out. The refinement that only refuses to ARRIVE at the Office buried is +0.16 and inside the noise.

Reserving Moves to get home costs 0.55 (t = −2.32), even though 62 of the 120 trains still on the board at game end were stranded in the district, unable to depart from anywhere but the Office. The switching work is worth more than the departures it forfeits.

Granting clearance when the train ahead has one Stage left costs 0.98 (t = −5.24) — a clean rejection of a rule that looked safe. Q13 collides on catching up, so a follower let onto a card whose occupant is leaving "cannot" catch it — except trains move in numeric order, so the follower can enter the region the leader still occupies before the leader has moved. "About to leave" is not "gone".

And three exact no-ops, each for a different reason. Playing Interlocking ahead of a train card changed nothing because the two sit in hand together 0.04 decisions a game. Stocking the Office platform first changed nothing because the bot already chooses it 79% of the time. Spending an idle turn on the Freight Agent changed nothing because it asked the same predicate a branch three steps earlier already acts on — a tautology, and 400/400 identical games is what one looks like.

The funnel explains the pattern: 8% of Cargo phases have a stocked green box, and the bot already takes 42% of the turns where stocking is productive. There is nothing to prioritise better. What remains is the economy itself, which is a deck question rather than a bot one.

A conservation audit, and the bug it found

Clearing an inbound box mints a car — 1.29 a game against a supply of 80. TODO.md had this as an open question ("no way to tell a load from a car"); it is duplication, and the two directions are not symmetrical:

• Outbound is paid for. stockToOutbound splices a loaded car OUT of the Division Yard to become the load, and loadCompleted banks the emptied car as the loaded one takes its place.
• Inbound is not. unloadBegan turns one loaded car into an empty car plus a load on MEN|AT|WORK, passengersDetrained does the same to a coach — and inboundCleared then pushes that load into the Classification Yard as a car, while the car it came out of is already back in service.

Over 200 games, cars-at-the-end minus 80 plus collision losses equals the inboundCleared count exactly in 71/200 games and 258 against 279 overall; the remainder is loads still in flight at the final whistle. Since ROLLING_STOCK_SUPPLY is the number that is supposed to set supply pressure, no supply figure can be tuned until this is decided. Not fixed here — it is a modelling decision.

Also found: state = fold(events) is not literally true. Replaying the event log onto a fresh state throws, because the phase driver mutates state directly and emits a descriptive event afterwards. Replay works by re-applying intents, not by folding events. Nothing is broken today, but the README claims the property and reconnection would rest on it.

The replays were all dead, and now they cannot be

All three published replays managed 2 intents of roughly 400 — the site was serving three recordings of nothing, exactly as TODO.md predicted would happen silently. save-replay.ts records bot games as saves, and verifies every one round-trips before writing it: same revenue, same Day, same intent count. harness.test.ts fails if any published replay stops short of its own history.

Six replays published, 18 to 23 Revenue, against a target of 20 and a bot median of 0 — including one game with four collisions that still cleared the target, and two with none.

0.1.1 — 2026-08-08

A way to tell whether a change to the bot helped

The bot averages 1.4 Revenue against a target of 20, and the obvious next step is to teach it to play better. That step could not be taken honestly, because there was no way to tell whether a heuristic had helped: revenue has σ ≈ 9 across games, so two runs of the identical bot differ by about a point through nothing but the deal. Every single-change claim in this changelog before the Interlocking work is inside that noise, and TODO.md has said so for a while.

node src/sim/compare.ts 1600 trainCapSlack=1 runs the current bot and one variant over the same deals and reports the per-seed difference. Giving both sides the same seed takes the deal out of the comparison: σ drops from ~9 on the level to 5.3 on the difference, and 1600 seeds puts the standard error at ±0.13 — in 1m45s. The noise floor moves from ±1.0 to about ±0.15, which makes every heuristic in the training plan resolvable. No parallelism needed; 100 games take 4.8s.

The report prints the better/worse/identical split beside the mean, because they are different claims. The first tweak measured is the case in point — trainCapSlack=1 over 1600 paired seeds:

REVENUE DELTA  +0.77 ± 0.13   (t = 5.97, σ of the paired difference 5.13)
seeds better 146 · worse 105 · identical 1349
best seeds: +75, +50, +46      worst seeds: -26, -20, -13

84% of games are untouched. It does not make the bot play better; it removes a rare catastrophe, and the mean rides on a handful of rescued games. Reporting that as "revenue up 65%" would be arithmetically true and misleading about what changed. At 400 seeds the same tweak read t = 2.57 — "not proven" — which is exactly the verdict it deserved there.

The tweak is measured but not adopted: the flag stays off, so this commit changes no bot behaviour. Turning it on is a change to how the bot plays and belongs in its own reviewed commit.

And it prints the funnel for both sides, because revenue can rise two ways: a channel started working, or an expensive channel was abandoned for a cheap one. A strict train cap raises revenue and cuts freight events nearly in half, and that has to be visible rather than inferred.

The funnel is new — GameStats.funnel, sampled live rather than recovered from the log, because the interesting gates are conditions rather than occurrences. "Was a green box stocked while a car was spotted" is not a thing that happens; it is true or false at a moment. It needed a per-decision hook on playGame (TurnObserver), separate from the existing event observer, so a phase is sampled once instead of once per event in the batch. What it says about the current bot:

• Passengers: of 7.4 arrivals a game, 70% reach a Passenger Facility, 25% carry the empty coach boarding requires, 21% the loaded coach detraining requires.
• Freight: of 60 Cargo phases, a green box is stocked in 8% and all three requirements meet at one industry in 7%. Freight is gated almost entirely on stocking.
• Stuck: 8.4 decisions a game are taken with a train's engine buried mid-consist, and on 3% of them is there a legal way to set the nose cars out — because it happens at the Office, where Rolling Stock may not be left.

developerBot is now makeDeveloperBot({}), byte-identical to what came before (asserted on three seeds by full event-stream fingerprint, and 1.4400 mean on 100 games either way). Variants exist so two policies can be compared in one process rather than by editing the bot between runs — which is how you end up comparing two things you cannot reproduce. Tweaks are temporary: a flag that measures well becomes the default and is deleted in the same commit.

The first tweak found the bug this tooling exists to find, twice over. trainCapSlack caps committed trains against the Office's A/D tracks. Written first as a gate on the two branches whose comments say they exist to play a train card, it measured exactly zero difference over 400 paired seeds — because followThrough ends with a generic "play what is in hand" fallback that played the card anyway. A cap has to remove the option, not guard the branches that reach for it.

Then the test for it was wrong in the same shape: it asserted only that some decision changed somewhere, and passed against the broken bot, because a tight cap does change which Local Operations option gets chosen — it just fails to stop the card being played two steps later. It now asserts the contract (committed trains never exceed what the Office can hold) and is verified to fail against the broken version. "Something moved" is not the promise.

Also new: a determinism self-check. The same policy on the same seed must produce an identical event stream, since the paired method rests on it and the failure mode is silent.

Tests 403 → 413.

0.1.0 — 2026-08-08

The first numbered build. Everything below the "second playtest pass" heading was made under 0.0.1, across twenty commits, which is exactly the problem the version convention above exists to fix: "the current build" was only ever answerable by a commit hash.

A third playtest pass — and three of the "display problems" were engine bugs

Fifteen items came back from two playtest sessions. Several of the ones that read like drawing faults were not: the board was reporting the game accurately and the game was wrong.

Cars could be added to a train that was not being made up. newTrain.placeCar accepted any tray with room in its consist, while the New Train Phase's own idea of the train it is waiting on — trainNeedingCars — requires the tray to be at a Division Point and short of its card. So during any New Train Phase the Division Yard would hand cars to a train standing on a siding in your own district, or one halfway across the Division. Measured before the fix: 50 such offers across 8 solitaire games, including Train 9 mid-crossing with three cars already aboard. That is the "cars magically appeared on my train" report, and the answer to it.

The two questions — "is this the train being assembled?" and "may this car be added?" — now come from one predicate (isBeingMadeUp, in apply.ts beside check), which is what stopped them disagreeing. newTrain.passCar takes the same guard.

The make-up panel merged two trains into one. Two trains can be built in the same Stage — a timetabled train and a Second Section, or an Extra — and the panel collected every option from every tray, titling the result with whichever tray came first out of the map. Reproduced at seed 99, Day 3 Stage 12: eighteen car chips under "Making up Train 8", covering two different trains. Worse than the wrong caption: the Division Yard chip binds to the FIRST matching option, so clicking a hopper could couple it to the other train. The panel is now scoped to the one tray the engine is waiting on. This was reported as the history announcing Train 9 while the panel said Train 10.

Upgrading the Office deleted Modifier bonuses. officeUpgraded wrote the new tier's printed numbers straight over the facility, so a Restaurant beside a Depot — +1 passenger out, +1 porter — was silently erased by the next upgrade, after the card had been spent. The tier is applied as a difference now, so the upgrade raises the Office by exactly what it is worth and leaves what is standing beside it alone.

A sentinel inside a coordinate's own value range. ABS Signals goes on a Mainline card, and the node index travelled as the fake coordinate { row: -1, col: node } — but row −1 is an ordinary district row, the first one below the Running Track, where most districts start. So ABS Signals highlighted whichever district card sat at that column, and an ordinary Enhancement laid one row down was described as being "out on the Mainline, node −2". card.play now carries node as its own field and a Mainline placement has no coordinate at all; the board simply does not light up for it. This was the other half of "why is my Depot highlighted?" — the first half being that Enhancements legitimately attach to played cards, which nothing on screen said.

The train, drawn the way it stands

consist is ordered nose first, and both renderers drew index 0 leftmost whatever the train was doing. A westbound train therefore came out right and an eastbound one came out mirrored — reported from seed 270861860, Train 10 running east and drawn engine-first at the WEST end, which reads as an engine shoving its whole train ahead of it. The board is a map, so the drawing obeys the map: west on the left, nose toward the way the engine faces. A crew on a north-south spur keeps nose-left and lets its ▲/▼ say the rest.

The Division chip draws the train now. It was a name and a figure, and the figure was Stages left to cross — read once as the car count, and once it was labelled · 2⧗, read as redundant beside the position the card already draws. It is: the region position is DERIVED from that countdown. So the chip carries the consist instead, in the same vocabulary the district card uses, and the countdown moved into the tooltip where there is room to say what it is. Stages are not regions — a card is two regions of fixed distance, and how many Stages a train needs over them depends on the card speed and the train (crossingStages); they coincide only for a 30 card and a normal train.

Loaded or empty, told the same way in both places. An occupied slot on a district card took a generic blue fill and only a LOADED car overrode it, so at 26×15px empty read as blue-ish and loaded as brown-ish — while the tray beside it made the same distinction unmistakable. Reported as "I left a car in a siding and now I can't remember if it was loaded". The card now matches the tray, and every slot and every car in a consist carries its own words as a tooltip.

Saying what the game already knew

Why a switching move is not offered. The switching game was played off a list of coordinates — move to (0, -2) as a button, nothing on the board, and no account at all of the squares missing from the list. exploreMoves returns the walk's REJECTIONS alongside its destinations, out of the same traversal, so a reason on screen is the rule that actually refused the square rather than a second guess at it. Five conditions: the rails do not meet, another train is on the card, the industry is locked by MEN AT WORK (§9.3 — no train may enter it or cross it), the coupling would overfill the consist, or it is a turnout you may run through but not stop on.

The board now shows where the crew is, where it may go, and marks the rest with its reason on the card's own tooltip. A turnout is drawn differently from an obstruction, because it is not one — and it is kept out of the Blocked panel, where every turnout in the district would otherwise appear.

Mandatory coupling is named on the button. "Move to (0, -2)" becomes "move to (0, -2) — couples loaded hopper, caboose on the way (onto the nose)". Coupling is compulsory (§A.4) and was announced only in the history panel, which is the one place a player is not looking while switching. 70 move buttons in five games now say what they will pick up.

The workers are on the card. "How do I see the number of laborers in an industry card?" — you could not; the number that decides every Cargo phase was in a side panel. Porters were worse: porters had been on the view-model all along and no renderer had ever drawn it, which is why a Restaurant's +1 porter had "no indication anywhere". Both are on the card now as free-over-total, and the Facilities panel shows porters beside laborers.

Ships-out and receives, in words. A Refinery and a Grocer's Warehouse were distinguished by the stroke colour of one 13×12px box and the direction of a 10px chevron. Rendered both and diffed the SVG to check the complaint: identical shape, identical sign, same box count in the same place. The card says SHIPS OUT or RECEIVES now.

And the Office told the truth about itself. baseOf returned zeros for a passenger facility despite a comment claiming it read the Office tier, so a plain Depot displayed out 1 +1 — crediting a Modifier that was never played. The card tooltip had the mirror-image bug, reading the printed tier rather than the Office as it stands, so a Modifier's effect never showed there either.

Undo, and the chrome around it

Undo, as far back as you like. The save is the seed plus the intents, so undo is a replay without the last one: no undo stack, no inverse of each action, and no way to reach a position the rules could not have produced. Solitaire only. It is a deliberate take-back rather than a rewind — the RNG advances with the replay, so a train card re-rolls the same Stage, but you can see that roll and then spend the turn differently. TODO.md carries the question of whether a Stage boundary should become a commit point.

Both Division end labels were clipped. They hung off the outside of the end cells and needed padding wider than the caption to survive, which the board did not have — and giving it to them would have spent that width on two captions instead of on the map. Centred under their own Division Point they cost nothing, and PAD drops from 92 to 22.

The turn chart scrolls away no more. The title bar was sticky and the row under it — Day, Stage, phase, waiting-on — was not, so scrolling the board took away the thing most worth glancing at. They are one sticky block now.

The timetable has a key. Blue is a train due, violet is the Stage you are in, dimmed is gone, and the green is a flash marking the slot the die just filled — which is exactly why it needed saying.

And the rule that shapes every district is finally written down. A district only grows outwards: the Limits signs are the growth point on the Running Track and move outward with the card, and no card anywhere can be inserted between two cards already down. Enforced since the beginning, stated nowhere. It now sits beside the district, and a legal square that already carries a card says what playing there would do — EXTEND THE RUNNING TRACK HERE, ATTACH TO THIS CARD — instead of lighting up blue and saying nothing.

Measurements

The two engine fixes take things away, so both were measured rather than assumed: 200 paired games before and after come to revenue 1.4 either way, cards played 24.4 → 24.2. Neither was paying for anything the bot relied on.

Worth stating plainly, since the changelog above still carries older numbers: the developer bot now averages 1.4 revenue against a target of 20, with 1 win in 200. That is not a regression from this work — it is the same at f4c0f49 — and it is consistent with what TODO.md already records about barring curves from the Running Track and making industries stub-only. The bot has not been retaught since those rules tightened.

Tests 386 → 402.

A second playtest pass

The arrival message told you to do work you could not do. Train 4 is the Express, which carries expedite: true — Q3 departs an expedited train in the SAME Stage it arrives. So it correctly arrived and highballed in one Mainline Phase and correctly did not wait at the Depot. But trainArrived narrated "it will highball again next Mainline Phase, so any work must happen now" to every arrival, which for an expedited train is exactly backwards, and the log then contradicted itself two lines later. The event carries expedited now and the message says which case it is.

TX14 (2) was Stages remaining to cross that Mainline card, not the car count — a bare figure beside a train's name, read as the consist twice by the same player because that is the obvious guess. Now TX14 · 2⧗.

"stock a coach at (0, 0)" reads as putting a car on the track, and was reported as exactly that confusion: no train at the Depot, so how is a coach being stocked there? It is a load taken from the Division Yard into the green Loading box, waiting for a train that can carry it — for a passenger facility, people on the platform. It now says so.

"trains enter" / "trains leave" claimed the Division ran one way. Odd trains run west and even run east, so both Division Points are a way on AND a way off; the buffer stops only mean it is a line rather than a loop. Both ends now read in and out.

The discard targets were "a very tiny highlight". Choosing a card to PLAY lights big ghost squares on the board; choosing one to discard lit a 2px outline on a small tile in a panel you might not be looking at. The piles now grow, brighten, say DISCARD HERE, pulse once, and dim the Salvage Yard beside them so the three that are choices read as a choice.

The timetable, and watching the die land

Playing a train card rolls 1D12 for the Stage it departs, and the card simply left the hand — the answer arrived as one line in the history panel, among many. The twelve slots have been in the Frame all along, and only the standalone replay ever drew them.

The play page now has a Timetable panel: twelve Stages across, the train due out at each, the current Stage lit, Stages already gone dimmed. Playing a train card flashes the slot the die just filled, sounds a die-and-chime cue, and says it in words above the actions — "Train 4 is scheduled to depart at Stage 11 — see the Timetable". The flash marks the moment rather than the state: it is gone by the next render, which is what makes it read as "that just happened".

No acknowledge-click. It would stop the game to say something the highlighted timetable says better, and would be clicked through by the third time.

The load pipeline is drawn the way freight actually flows

Reported after unloading at a warehouse: "it went W, A, M, and then to the red box at the far right."

The engine was right. §9.3 is explicit — "The first Laborer replaces the load with an empty car of that type on the industry's track and places the load on WORK. Additional Laborers move it to AT then MEN. The last one places the load on a red Unloading box." An unload ends in red, and it did.

The drawing was wrong. The pipeline was laid out green → MEN|AT|WORK → red, left to right, with both arrows pointing right. But loading runs Green → MEN → AT → WORK → car, and unloading runs the other way entirely: car → WORK → AT → MEN → red. Green and red therefore both belong beside MEN, and the car belongs beside WORK — so putting red at the far right put it exactly where the car is, and an unload appeared to run backwards across the whole row and land on the end it had just left.

Now green and red both sit before the sign with their arrows pointing opposite ways, and only the boxes an industry actually uses are drawn: a Grocer's Warehouse has no green box and a Mine Tipple no red one, where before every industry showed both and half of them were dead squares.

The Facilities panel had the same left-to-right assumption in its row labels. green and red are now waiting to load and cleared inbound, each with the direction its loads travel spelled out, and a row an industry cannot use is omitted rather than shown empty.

A playtest report, worked through

Backing up turned the train around. facing is which way the ENGINE points, and it was reset to the direction of travel on every move — so one reverse move silently spun the train about, everything read "forward" again, and a run-around became pointless: you could change ends for free by backing up twice. Running forward the engine leads and points the way it went; backing up it trails, still pointing the way it came, which is the port it arrived through. Both hold around a curve.

"drop 1 car(s)" hid an option entirely. It never said which car, and it read identically for a nose drop and a tail drop — so the action list's duplicate-label filter discarded one outright, and setting out from the front of the train could not be chosen at all. The same failure as the turnout rotation earlier, in a feature added two commits ago to fix the make-up deadlock. Now: "set out the caboose off the back".

ABS Signals offered the Office Area. The card says "any Mainline card", and checkEnhancementPlacement reads placement.col as a Division NODE index for it — while the candidate list handed it occupied grid cells. So "(0, 1)" was accepted because node 1 happened to be a Mainline card: the label and the meaning were different things. It now offers the Mainline cards by name, "on the Uncontrolled Siding, out on the Mainline".

Two T10 chips on one Office. A train standing at the Office is on the Office grid card AND on an A/D track, so it arrived in both lists and was drawn twice.

The train was invisible on the board, which is what made the switching game unplayable: every decision is about car ORDER — which car comes off next, which end a cut couples onto — and the card showed a name badge. The train is now drawn as it sits in the tray: engine in its place with an arrow for which way it points, cars in order, loaded solid and empty hollow. That arrow is also what makes "(reverse)" mean something.

Modifier effects were applied and invisible. The Ice House and Local Small Groceries both worked — capacity.outbound and laborers went up — but the card draws Math.max(1, greenCap) green boxes, so 0 → 1 looked identical, and laborers were never drawn at all. The Facilities panel now shows laborers, outbound and inbound as numbers with what a Modifier added (2 +1), names the Modifier cards beside the industry, and says in the tooltip what the card itself prints. Laborers had been in a hover tooltip only, which is the wrong place for the number that decides every Cargo phase.

Coupling was silent and left no trace. A car simply vanished from the board with only a line of history to say where it went. There are now two new sounds — a knuckle-coupler clank for coupling and a quieter one for setting out — alongside the whistle, bell and conductor.

Moves left were reported only in the history panel, which is the one place a player is not looking while switching. Now beside the buttons, and struck red at zero.

A/D tracks were a tooltip. "3 A/D tracks" with nothing on the card — the number that decides whether the next arrival is an automatic collision. Drawn as pips, filled for taken; there is no room on the card for more rails and the count is what matters.

The phase changed under you. Local Operations ends the moment the last Move is spent and the automatic phases then run themselves, so the page could change between two clicks with no notice. A banner now names the phase it moved to.

What was already right, from the same report

The car type and colour on the board (read as "cab in red" without being told), the Blocked panel explaining a full industry track, the Facilities panel's accepted car types and spotted cars, and the Division chip's T10 (2) with its consist on hover. All four were reported as useful, and none of them changed.

Softlock: a train being made up with no visible way to make it up

Reported at Stage 10 of seed 775569289 — Train 10 at the West Division Point, history saying "now taking cars", and nothing at all under Your Move.

Moving train make-up onto the Division Yard chips took the "Making up …" group out of the action list, and everything that was not a car went with it: the heading naming the train and what its card calls for, and the "no more cars" button. The nine cars the train could take WERE clickable on the yard chips the whole time — nothing on screen said so, and the panel a player looks at was empty.

The panel is back, and now says where to click: "Click a car in the Division Yard below to add it — 9 kinds it may take are highlighted there." It carries the pass button when passing is legal, which §7 allows only when the Division Yard is bare — "must make every effort to find a suitable car".

The engine was never at fault. 2065 New Train decisions across 60 campaign games, and not one offered zero options; the pass/place pair covers the phase. The panel still words the third case honestly rather than implying a button that is not there.

Two guards, because the obvious one would not have caught it. A menu-level invariant — every option the engine offers must be reachable through something the menu exposes — passes on this bug, because makeUp.pass was in the menu and correct all along. What failed was the page never reading it. So there is also a coarse check that main.ts references every field the menu offers: a field nothing reads is either dead or a control that has gone missing. Verified by putting the regression back and watching it fail.

Four reports from playing seed 775569289

A curve was described as a turnout. Both read "east-west track with a 45° leg", which is a turnout — a road straight across the card plus a leg off it. A curve has ONE road: in from the east or west edge, along the centre line to the frog, out at 45° through the middle of a north or south edge, and nothing runs past it. Which is precisely why it may not be laid in the Running Track, so describing it as though it had a through track contradicted the rule that stops you.

A curve in hand showed no preview. The shapes were read off the card's legal PLACEMENTS, so a card with nowhere legal to go had nothing to draw — and that is exactly when a player most wants to see what the piece is. They now come from the card itself, which is where they belong: what a piece looks like does not depend on whether there is currently a square for it.

"Realignment on Mainline card 3" named a raw node index. It said nothing about which stretch of the Division it meant or what it would do, and it had no tooltip either — because the action list attaches one only when a label happens to contain an em-dash, which this one did not. It now reads

Realignment on the Uncontrolled Siding — the second Mainline card west to east; converts it to Double Track

and any action naming a card falls back to that card's own description when its label carries no explanation of its own. That was the actual complaint: the same card explained itself perfectly in hand and said nothing in the action list.

It offered only one Mainline card, and that was correct. REALIGNMENTS converts Plains, Curves, Uncontrolled Siding and Trestle; the Division on that seed is a Heavy Grade and an Uncontrolled Siding, so only the second could be converted. The engine was right and the label was hiding it — "Mainline card 3" gave no way to tell a considered restriction from a bug. Naming the card fixes the report without changing the rule.

The hand is the action surface, and the yard makes up the train

The action list reached 22 buttons, and most of it was a cross-product. A card appeared in two panels under two different models: as a subject under "Play a card from my hand", which then highlighted squares on the board, and as one flat button per Department under "Discard a card from my hand". Four cards times three Departments was twelve buttons repeating the same three choices four times, about 290px of the list. Separately, making up a train offered up to ten buttons reading "add loaded hopper", "add empty boxcar" — while the Division Yard sat on screen already showing exactly those cars by type and load state.

Both are now on the objects already being looked at, using the pattern board placement always had: pick the thing, then pick where it goes.

• Every card in hand carries its own verbs. play highlights the squares it may go on, exactly as before; a card needing no square (an Office upgrade, a train, a maneuver) goes down in one click. discard lights up the three Department piles as targets — they already show their top card and their depth, which is precisely what you choose between.
• A make-up car is picked off the Division Yard chip that shows it. The loaded and empty counts are separate targets, because a car of a type and a load state is exactly what the choice is.
• The action list keeps what is not about a card or a car: the Local Operations choice, drawing, switching moves, the Freight Agent, and finishing.

Measured over a full game of seed 430: the widest action list went from 22 buttons to 5. A test now walks the same game and fails if it climbs back above 8.

The rotation step stays where it was and is now the only thing the placement panel shows — the card is picked in the hand and the square on the board, so a rotation is the one question neither of those can ask. Its hover previews, added earlier, are unchanged.

A note on the two replay viewers

TODO.md now carries an item to decide between them. The standalone node src/sim/replay.ts writes a self-contained HTML file that nothing links to and that .gitignore excludes; the site reads JSON saves from public/replays/. The standalone one carries the bot's decision trace and a timetable panel, which is debugging material rather than something a player wants. No action taken.

A replay now looks like the game it is a replay of

"Extra slow" was there and did nothing. The site's replay viewer builds its interval with whatever the speed select held when play started, and nothing re-read it — so changing pace mid-replay had no effect at all and the pace looked stuck. The standalone replay had always restarted its timer on change; this viewer was missed. Both offer the same five paces, extra slow through very fast, and a test now asserts they stay in step.

The turn chart lived on one screen out of three. Where you are in the Day — the five phases with the violet "you are here" — was in main.ts alone, so both replays reported the Day and the phase as two plain strings. The same position looked like a different game depending on which screen you were on. It is now sim/turnchart.ts, shared exactly as the board renderers are: the playable page and the site viewer import it, and the standalone replay embeds it by Function.toString() because it is a single file with an inline script and cannot import anything.

And the side panels were three against eight. The viewer showed the Division, the Office Area and the log; the play page shows those plus cards in hand, the Department decks, the yards, the blockers and the facilities. A replay could not answer "why is nothing moving?" — which is most of what a replay is for. web/panels.ts now renders all of them for both pages, and the duplicated CSS is gone from play.html.

Three tests hold it there: both screens must call the same panel renderers, all three must use the shared turn chart and none may keep a private copy of the phase table, and the two viewers must offer the same paces.

You could not rotate a turnout at all, and now you can see what you are laying

The rotation was being thrown away before it reached the menu. actionGroups drops duplicate labels, and describeIntent for a card play said only play right-hand turnout at (0, 1) — no rotation in it. So a turnout's two orientations produced the same label and the second was silently discarded. The "choose a rotation" step existed and worked; it was never given more than one rotation to choose between. The label now names the orientation, and both survive.

And the buttons are pictures now. Hovering a rotation draws the piece as it will land on the board, and hovering the card itself draws every shape it could be laid as — which is how a player sees a turnout has two orientations before picking a square at all. Rendered by officeSvg, the board's own renderer, on a one-card board: the preview and the board cannot disagree about what the piece looks like, and the rails come from the engine's connectionsFor, so a preview cannot promise a shape the placement will not produce.

Placeable.spots carries the links for this. data-tip-html on the tooltip renders a figure above the caption; it is only ever set from markup this app builds.

Turnouts say what they do, and a curve may not break the Running Track

"Right-hand turnout, stem east, through west, diverges north at 45°" is three pieces of jargon and a compass reading, and none of it answers the only question being asked: if my train comes in from over there, where can it go? Turnouts now read

allows traffic from the east to travel west or turn to the south

and curves

carries traffic from the west round to the south

everywhere they appear — on the card in hand, on the placement it would make, and on the card once it is down. §A.1's rule falls out of the same sentence rather than needing a shouty clause of its own: a train coming the other way, from the through end or the diverging one, may only leave by the stem, so the two roads never join.

A curve laid in the Running Track dead-ends the main. A curve has ONE road — from an east or west edge round to its 45° leg — so a card of it standing in the running row stops the through route at that square and cuts the Office off from its own Limits. Nothing forbade it, and the bot did it: an early trace shows it laying a ne curve at (0,−1), turning the west end of its own Running Track into a stub. Every card that may stand in that row now has to carry the road across it — carriesThroughTrack — which straights, turnouts, Limits signs, the Office and industries all do. Reported as BREAKS_RUNNING_TRACK rather than NOT_CONNECTED, because it is a different mistake: the card would join perfectly well and would still leave the main stopping dead at it.

The crossover is confirmed, both hands. A turnout that "allows traffic from the east to travel west or turn to the south" joins the one directly beneath it that "allows traffic from the west to travel east or turn to the north" — the two 45° legs are one continuous rail across the card edge. It already worked; it now has a test that says so in those words, along with its mirror built from the other hand, the mismatched pair that must NOT join, and a crew actually running down through it onto the parallel track.

Measured, and worth flagging: barring curves from the running row cost the bot a good deal — districts 28.0 → 19.7 cards, facilities 2.23 → 1.85, revenue about 2.0 → 0.8. The rule is right and the bot was partly living off an illegal placement. Two tests needed widening rather than weakening as a result: the unload regression sampled three deals and now samples twelve (unloads still happen, 34 across a 40-deal sweep, just not on those three), and the canvas bounds test now holds a straight as well as a curve, since a curve alone is offered only one square once the running row is closed to it.

The Crew Tray is a train, and a train must be made up to leave

§8.2 was not checked at all. A train could highball onto the Mainline engine-last with its caboose in the middle. Now a train held at the Office is held until it is made up: the engine at an end of the tray — pulling or pushing, both are real — and the caboose at the far end from it. The check is deliberately direction-free, because what a train may not be is broken-backed, with the engine buried among its own cars and some ahead of it and some behind. That state is only reachable through switching: a train arrives made up and comes apart because the player took a cut onto the nose or picked cars up in a run-around.

engineAt was written and never maintained. Cars taken onto the nose go AHEAD of the engine — Appendix A: "a train can pick up two cars and add them to the Crew Tray in order that they were in, pushing them into the Facility" — so the engine stops leading, and its recorded index did not follow. It does now, and drops adjust it the other way.

Setting out from the nose was missing entirely. Appendix A uses the move in its own worked example — "Back up and drop off everything on the nose of your train (red and blue) on Card B" — and switch.dropCars could only ever take from the tail. Without it, cars taken onto the nose could never come off, so an engine buried in its own train had no way back to an end: the first version of the make-up rule stranded a train permanently in 6 games of 40, holding an A/D track for the rest of the game. fromNose fixes it, and a cut is now guarded to come off an OUTER end only — lifting cars from beside the engine would leave the far end of the train coupled to nothing.

The bot learned both remedies: dig the engine out when it is buried, and shed a misplaced caboose when it is at a reachable end. Measured over 40 games: 106 make-up holds across 6 games → 13 across 1, revenue 1.55 → 2.05. The one that remains is a caboose stuck mid-train, which genuinely needs a run-around or a Small Yard — the game working, not a defect.

A knock-on worth recording: with cars now set out properly, the freight pipeline runs clean — 55 loads started and 54 completed across the 40-game sweep, with zero jams of either kind. The regression test that asserted the bot clears a MEN|AT|WORK jam had become untestable, so §6.3's unjam is now asserted directly against a constructed jam instead of hoping the bot stumbles into one. It also checks that clearing the jam reopens the track, which is the point of it.

You can see which car is where. The board printed the first three characters of each car's label, which for "loaded hopper" and "loaded boxcar" alike is loa — every car on the map looked identical. The switching game is entirely about getting the RIGHT car to the right industry, so type now reads by colour and by three letters (box hop tnk rfr cch cab), and loaded shows as a filled slot against an empty one's outline — the same distinction the printed game makes with coloured tokens.

Operational Rail — a locked industry could be driven straight over

Checked the whole rule against Appendix A of StationMasterPrototypeRules.pdf, which defines it outright: "Operational Rail is any track card that a train can stop and leave Rolling Stock (uncouple) on. Operational Rail is any track card with a train wheel icon on it."

Most of it was already right, and the wheel icons in the rules diagrams confirm which cards carry one. Straights, curves, industries and the Limits cards all do. The turnout does not, and the page says so in words — "Since there is no Operational Rail wheel icon, the train may not stop on this card" — so a train runs through one and may not stop or uncouple on it. Both already held.

The Office is Operational Rail. The Depot card in the diagrams carries a wheel icon and is drawn as one of the green squares a Crew Tray may move to, and the Special Rules say "While your Office Track is considered Operational Rail, Rolling Stock may not be dropped off here." So a train may stop there and may not leave cars — which is what canDropCarsAt already did. Worth stating plainly because it is easy to read the "no cars here" half as "not Operational Rail", and the A/D track mechanic depends on trains being able to hold at the Office.

The real gap was §9.3's lockout. "While ANY loads are in the MEN | AT | WORK track, the industry's track is locked down... It loses its status as Operational Rail. No cars can be picked up or dropped off, and no trains may occupy or move on it." Losing Operational Rail status only stops a train FINISHING somewhere — a turnout is not Operational Rail either and trains run through one all day. Passage was never blocked, so a crew rolled straight over a locked industry and, because coupling is automatic and mandatory, picked up the cars spotted on it on the way past. Those are precisely the two things the safety lockout exists to prevent.

isLockedByWork is now separate from isOperationalRail for that reason: "cannot stop here" and "cannot pass through here" are different properties and only a locked industry has both.

Covered by a new suite in track.test.ts that walks each card type against the wheel-icon rule, asserts the Office may be stopped at but not unloaded on, asserts a turnout may be passed but not stopped on, and asserts a locked industry blocks both stopping and passage — then reopens when the work clears. It also checks the supply catalogue's isOperationalRail column against the rule, so the data and the diagrams cannot drift apart.

Industry cards: on a stub, one of a kind, and never both ends of a chain

From the sheet, the "Placed" column reading identically for all six industries — "Straight, Stub (not on Running Track)" — and the "Lockouts" column beside it.

An industry may no longer be built on the Running Track. It was allowed and merely warned about: the card text noted that a car left standing there would be hit by the next arrival. That is a hazard, not a rule, and it let a player skip the district entirely and spot cars on the main line — which removes the whole switching puzzle, since the point of a siding is getting a car down off the main and back. check now returns ON_RUNNING_TRACK.

No two of the same industry in one Office Area. The sheet states this in the Freight House row, which lists Freight House among its own lockouts; the catalogue had dropped that self-reference as if it were a typo. It is a general rule, so it is enforced for every kind in isLockedOut rather than repeated in all six entries.

Producer and consumer of the same commodity stay apart. Mine Tipple makes the coal a Power Plant burns; the Refinery makes the oil it also burns; Packing Sheds fill the reefers a Grocer's Warehouse empties. Build one end of a chain or the other, never both — which is what pushes freight to run between districts instead of circling inside one. The pairs were already right; they had no test coverage at all, and now have a suite that checks each pair in both directions, checks that industries sharing no commodity may stand together, and checks the catalogue against the sheet column so a change to it has to be deliberate.

Measured, 60 solitaire Standard games: 0 industries on the Running Track, 0 duplicates. Industry placements fell from 3.84 a game to 2.23 and revenue from 2.87 to 1.35, because the bot builds shallow districts and there are now far fewer legal squares. Not rebalanced — deliberately. The counts, the industries and the track mix are all due a pass together once the rules are right.

One consequence worth naming: flyingSwitch stopped firing in the 60-game reachability sweep. The rule is fine — mainline-cards.test.ts exercises it end to end on a hand-built siding — but the bot no longer gets a crew next to an industry. It is exempted by name in that test, with the reason written next to it, so the other forty-odd event checks stay live and deleting the line is what proves the bot has been fixed.

§6.2's reshuffle, and the Departments and Salvage Yard on screen

The reshuffle existed as a fiction. events.ts declared { type: 'deckReshuffled' } and narrate.ts had a line of prose ready for it — "Home Office deck ran out — Salvage Yard reshuffled back in" — and nothing anywhere emitted or reduced it. check just returned DECK_EMPTY. A declared event with narration written for it reads as an implemented feature to anyone grepping for one, which is worse than an obvious gap.

Now real: when a draw takes the last card, the Salvage Yard and all three Department decks are collected, reshuffled, and §4.6-4.7's opening is re-run — three cards turned face up as the Departments, the rest face down as the deck. Cards played onto the board are not recovered; they are on the table, which is where they belong. A game that has genuinely used everything still ends on DECK_EMPTY rather than reshuffling an empty sweep.

The full shuffled order rides the event rather than being recomputed from rngState. A save is a seed plus the intents, so events are never serialised and the size costs nothing — and an event that states the outcome outright cannot drift from the reducer the way a re-derivation can. Covered by a test that builds the same position twice and asserts the two decks come out identical.

It has not fired in play yet, which is worth knowing. Solitaire Short/Standard/Campaign end with 186.8 / 176.2 / 168.8 cards left of 243 and never ran dry across 60 games each; four-player Campaign ends with 86.6 across 25 games. It is a safety net, not a live mechanic. It is not the case that the Departments only grow — a player takes the top card of one as their draw as readily as discarding onto it, so a Department can be drawn down and refilled from the Home Office deck. What the reshuffle guards is the Home Office deck itself running out, which is possible whichever way the piles happen to be moving.

The Salvage Yard is a pile like the others, and now shown like them. It was already modelled as a list; what it lacked was any presence on screen, which mattered the moment it became the thing that comes back in a reshuffle. Watching it fill is the only warning a player gets that the deck is about to turn over.

All four piles now show their top card and their depth. Each is drawn as a card with the pile's name and a count badge on it, then the face-up card underneath. The depth is a count and not a hint: only the top card may ever be drawn, so everything below it is out of reach, and choosing where to discard is choosing what to put there.

The Departments are decks, and you choose which one to discard onto

From the designer: "when discarding from their hand, the player can select which department card deck they want to place the discard on top of. Department card decks are shared across all players. When pulling from a department card deck players may only pull the top card."

The rules already said so and the code had read them the other way. §6.2: a discard is placed "face up on top of one of the three Department slots", and a draw takes "the top face-up card". Both phrases only mean something over a pile. Gap 4a had concluded the Departments were "three face-up market slots fed from the one deck, not decks with their own contents"; that finding is now marked corrected in open-questions.md.

It was destroying cards. cardDiscarded did departments[toSlot] = cardId — assignment, not a push — so discarding onto an occupied Department annihilated the card already face up there. A closed deck was quietly leaking. It survived because the only card-conservation test ran at setup and never again; there is now one that counts after a full game, and one that asserts no id is ever in two places at once.

And the choice was invisible. All three discards described themselves as discard X, and the action list drops duplicate labels — so three genuinely different decisions collapsed into a single button and the Department could not be picked at all. Discards now read discard Freight House onto Department 2, burying Brakeman, and a draw reads take Depot from Department 1, 3 buried beneath it. The browser shows each pile's top card with a +n under count, because a deep pile is where cards have been put beyond reach and that is what a discarding player is choosing between.

Refill timing changed with it. §6.2 refills an empty Department from the Home Office deck. The old code refilled after every Department draw, which was harmless when a slot held one card and would now drain the deck into the piles. It refills only when taking the last card empties one.

The bot got the strategy this opens up. It used to take the first discard option, always Department 1, burying whatever sat there — including the Depot it was waiting on. It now covers the face-up card least worth keeping reachable, never one it would take, and breaks ties toward the shallowest pile. Measured over 100 games: spreading discards 2.87 revenue, concentrating them on the deepest pile 2.67, indifferent 2.67. Three piles offer three face-up cards, and piling onto one of them leaves the other two showing whatever they started with.

In a competitive game the same call reads the other way round — burying a card a rival wants is an attack rather than housekeeping — which is exactly why the choice belongs to the discarding player.

Measured, 100 solitaire Standard games: Departments hold 8.3 cards between them at game end, deepest single pile 17. The Home Office deck ended with 176.6 of 243 and never ran dry, which matters because §6.2's reshuffle — collect the Salvage Yard and all three Departments, reshuffle, re-establish the deck — is not implemented. It has never been reachable in a 5-Day solitaire game; it will matter for Campaign length and for four players. (Implemented in a later entry.)

Track is a deck card, not a private supply

Reported by the designer: "all track cards are included in the home office deck and are played from there like any other card."

The error is visible in the spreadsheet. docs/Deck cards2.xlsx has a column B headed "Number in Deck" — 32 straights, 16+16 curves, 4+4 sharp curves, 16+16 turnouts, 104 cards — and a LAST column headed "Track Per Player" reading 8/4/4/1/1/4/4 = 26. The code took the last column as a separate physical stack and wrote "Track is NOT in the Home Office deck — this is the single biggest structural change from the placeholder." It is 104 shared among four players, not a second pile. The sheet's own totals settle it: "Sum other 115", "Total track 104", grand total 231 — and 115 + 104 + 12 start cards is exactly 231.

What went. TRACK_SUPPLY/TRACK_PER_PLAYER, OfficeArea.trackSupply, the track.lay intent, the trackLaid event, protoTrackCard, TurnState.laidThisTurn (the one-piece-a-turn cap, which only existed because a private supply had nothing else bounding it), and the "Your Track Supply" panel. CardKind for track gained a hand, because handedness is the diagonal and a track card without one cannot say what it may be joined to.

Most of the plumbing was already there and dead: checkPlay had a track branch, protoCard had a track branch, and the cardPlayed reducer already placed a track card with a rotation. Track had been a card once, and moving it back was largely deleting the parallel path.

The deck is 243 cards, of which 104 are track — the largest category by some way, and the point of the change: building a district is now paid for in the industry or train you did not draw, and a three-card hand is the real constraint on how fast a railroad grows.

Measured over 100 solitaire Standard games, against the same run with track as a private supply:

	private supply	in the deck
deck size	139	243
district size	28.5 cards	28.0
mean max depth off the main	1.45 rows	1.94
turnouts / curves / straights per game	7.6 / 7.2 / 3.4	5.6 / 7.4 / 5.6
facilities placed	4.56	3.84
districts with a run-around	70/100	29/100
facilities on a run-around	0.51/game	0.18
revenue	5.80 mean	3.05

Revenue nearly halved, and that is the headline for the next decision, not a defect to paper over. A run-around needs a turnout, a matching curve, straights, a second curve and a second turnout — all of the right hand, arriving in a three-card hand in a usable order. It used to be a shopping list; it is now a draw. Two test floors were re-baselined against the measurement with the old figure recorded beside them, deliberately set BELOW what was measured so they detect the loop machinery breaking rather than endorsing 29%.

Two bot fixes fell out of it, both real. Track became an ordinary card.play, so the generic "play anything placeable" fallback started dumping track on whatever square was legal — bypassing bestTrackLay, which had already looked at the same piece and declined it. Measured: 26 of 60 districts ran the siding past the last column with a way up. A track card the scorer will not use is a card to discard. And arcsLeft, which asked a supply that no longer exists, became arcInHand: "have I got a curve of this hand?" is now a question about the hand, not a certainty.

Still open, and now urgent. The office counts are doubled (Depot 4→8, Station 2→4, Terminal 1→2, Q12) and the industries tripled (Gap 12), both tuned by measuring a deck with no track in it — 25 of 100 games never drew a Depot and never escaped Whistle Post. Adding 104 cards dilutes every draw by 43%, which is exactly what those multipliers were compensating for, so they are now either badly needed or badly wrong and only a measurement will say which. test/setup.test.ts records both departures and flags them for re-measurement.

The placement labels described the mirror of the card they would lay

Reported as "I can't play a turnout north or south of an existing turnout". It was always legal — a turnout under a turnout is a crossover, and it is how a siding gets a track running parallel to the Running Track. The engine accepts left-over-left and right-over-right today, and the square was offered and clickable. What was wrong was the words on it.

rotationNote in src/web/game.ts called variantsFor(geometry) without the hand. Hand is the diagonal, so without it variantsFor answers for the left-hand card whatever you are holding. Every right-hand turnout was offered as "stem west, through east, diverges south" — the exact mirror of the card it would lay — and every right-hand curve named the wrong edge. The placement was always correct and only the description lied, which is the kind of bug that survives a green test suite and makes a working feature feel broken.

Fixed by threading hand through rotationNote → variantLabel, and covered by a test that walks every geometry × hand × rotation and checks the label against variantsFor's own answer.

And a placement now says what it would connect to. Two cards meeting at an edge is not a rail — on a north or south edge their 45° legs must also share a diagonal — so "is this square legal" and "does this piece meet the one I am aiming at" are different questions and only the first was on screen. Spots now read (-1, 1) — stem east, through west, diverges north at 45° · joins the track above, which is the crossover named outright.

A New game button, instead of finishing the one you have

start() restores from localStorage on every load and the only "new game" button lived on the game-over screen, so a game you no longer wanted followed you across reloads with no way out.

New game sits beside Save replay in the header. It confirms once past the opening Stage — the save is the game, there is no undo, and the replay download is right there — then clears the save and reloads. It drops any ?seed= from the URL as well: leaving it would deal the same game again and look like the button had done nothing.

The track is 45° geometry, drawn from the printed cards

The prototype designer's feedback: there are no north–south tracks. Straight runs are always east–west, the east–west line sits dead centre on the card rather than biased to the top, and a turnout is an east–west through track plus a curved leg meeting the north or south edge at 45°, which must line up with the corresponding leg on the card it abuts.

Measured off docs/tracks.png rather than inferred. Card grid lines at y = 15/164/314/464/614/764/ 914/1064/1213 on a 2136×1397 sheet; rail centres at 89.5/236/–/–/689.5/839/989/1135.5/1288.5 — the card's exact vertical middle, within a pixel, on every one of the nine rows. Cards are 259×150 (aspect ≈ 1.73). Every leg crosses the middle of its edge. Four shapes exist: turnout with the leg off the west end, turnout with the leg off the east end, curve west↔south, curve east↔south, plus the plain straight. Depot, Station, Terminal, Refinery, Freighthouse, Coal Tipple, Manufacturing and Town are all plain east–west straights with no diverging leg at all.

Slope is now part of adjacency. Name the two diagonals after the pair of arcs that mate across a horizontal card edge: ne_sw (port pairs {n,e} and {s,w}) and nw_se ({n,w} and {s,e}). An sw card above an ne card is one unbroken rail; sw above nw is a V — both cards have the port, both legs meet the same point on the edge, and they still do not connect. joins(a, p, b) in src/engine/track.ts is the single adjacency test, and every hasPort(nb, opposite(p)) in the engine, the bot and the tests now goes through it. Leaving one behind reintroduces the V.

Handedness is that diagonal. A printed card has a back: it turns 180° but never flips. So a straight has ONE orientation, and a curve or turnout has two — 0° sends the 45° leg south, 180° sends it north, and neither changes diagonal. Left-hand reaches sw/ne, right-hand se/nw. That is what makes the 4-left/4-right split of the curve and turnout supply mean something: a run-around needs one card of each hand — a left turnout down, its matching left curve, straights along, a right curve back up into a right turnout. variantsFor takes a hand argument for this reason.

What went, and what opened. TrackAxis and every axis field are deleted outright — there is one axis now, and a one-valued field invites the old branching back; deleting it turned tsc into the migration checklist. The Office's e-s/w-s stubs are gone, as are north–south facility placements. Q7's "a district hangs below the Running Track" is lifted: a turnout turned 180° reaches north, so placementCandidates and adjacentFacilityCoord no longer reject the rows above, and §9's "nine nearby spots" really is nine. legalIntents offered six variants per placement per card; the widest set is now two.

The renderer draws the card rather than approximating it. RAIL moves from 30 to H/2, and the card widens from 132×96 to 166×96 to match the sheet's 1.73 — the aspect is not decoration, since the frog where a 45° leg meets the through track sits H/2 from the centre and a squarer card pushes it almost to the edge. The fixed elbow at (W/2, RAIL + (H-RAIL)/2) is replaced by the real frog at (W/2 ± H/2, H/2) and a segment at exactly 45° to the edge midpoint. That elbow sat below the rail on the assumption everything diverged downward, so a leg reaching north was drawn as a hook that dropped past the rail and came back up. Four tests in test/web.test.ts now measure the emitted SVG: the through rail level and centred and spanning the full card, every leg at 45°, ne/sw on one diagonal and nw/se on the other, and the leg crossing the edge at its midpoint. The enhancement label moved above the rail, where it used to print along it.

Measured, 100 solitaire Standard games, against the same run before the change:

	before	after
mismatched 45° joints on finished boards	—	0
district size	25.4 cards	28.5
mean max depth off the main	2.38 rows	1.45
built above the Running Track	0/100 games	93/100
straights laid	1.01/game	3.44
districts with a run-around	99/100	70/100
facilities on a run-around	0.82/game	0.51
revenue	7.34 mean	5.80

The shallower districts are intrinsic: no card joins north to south, so descending a row costs at least two cards — the leg out, then a curve turning the run back east–west. Districts are now wide and shallow, which is what the printed sheet and a real yard both look like, and the run along the siding is straights, which is why the bot lays three times as many (13 of the 18 enhancement cards need one).

The revenue and run-around drops are honest, not a regression to chase. A run-around costs more pieces than it did, and the Office no longer offers a free way back up onto the main — it used to carry e-s/w-s stubs, so every district had one guaranteed climbing point. Two bot heuristics were tried against it and both rejected on measurement: preferring one side of the main changed nothing at −4 and made things worse at −100 (62/100 run-arounds), and paying more for the second turnout that closes a siding bought 86/100 run-arounds and 1.08 facilities on a loop but cost 9% of revenue (5.80 → 5.25). Revenue is the game's own measure, and both structural floors in test/sim.test.ts still pass, so neither shipped.

The bot is otherwise rewritten around the new geometry: descendFrom/waysOff/reachableOffMain/ runAroundCells take a side rather than assuming "below", the arc score asks joins instead of reading the arc name, and arcsLeft is asked per hand — a left-hand turnout's leg can only be continued by a left-hand curve, and a supply full of right-hand ones is no help to it.

Q13 — a train that catches the one ahead runs into it

Answered, and implemented as option B: collide on catching up, which is the version that rewards judging the gap.

§10 makes a Mainline collision the Superintendent's fault and removes both trains, and ABS Signals exists to stop trains rear-ending each other — but §8.3's trigger list never named one and nothing was implemented, so granting clearance was FREE: both trains survived, no penalty, and ABS Signals protected against nothing. Now a following train that closes on the one ahead runs into it, and one that never closes is fine, so clearance is a bet on relative speed rather than a formality. §2.1 divides the card into two regions, and sharing one is what "caught up" means. ABS Signals does what it prints instead: the follower stops short and holds.

Not on a card that prints "trains may pass". The first version fired 0.41 times a game while the bot never once granted clearance, which is the tell — those were all Double Track and Uncontrolled Siding, cards that hold two trains because they HAVE two roads. Catching up there means going past, which is what the card is for.

With that corrected the mechanic is invisible to the current bot, because it always denies clearance. That is the right shape, and the teeth are real:

Superintendent	revenue	rear-enders	ABS holds
always denies (the bot)	7.34	0.00/game	0
always allows	−5.13	2.20/game	19

So the bot's always-deny policy — a deliberate choice made when clearance was free and the arithmetic only guessed at — turns out to be correct, and is now correct for a measured reason.

Tested deterministically rather than through the bot: two trains built onto one single-track card with the follower closing, which collides whichever order the phase processes them in, and the same pair again with ABS Signals to prove it holds instead.

The engine has a place in the train, and the yards are visible

The engine had a position the game recorded and never used. engineFront was a boolean, written in three places and read in none — so every consist was drawn as an anonymous row of cars. It is now engineAt, an index into the consist, because a Crew Tray is an engine plus its Rolling Stock and the engine may be PULLING (ahead of everything), PUSHING (behind everything), or in the middle doing both at once. A boolean cannot say the third thing. Consists are drawn with ENG where it sits.

The engine is deliberately NOT one of the consist entries: §8.2 counts the consist as Rolling Stock, and the four-car limit (§A.4) is a limit on cars, not on the locomotive hauling them.

Both yards are now on the page, by car type and split loaded / empty, with the Division Yard outlined the moment it goes bare. This matters more than it did: the Classification Yard returns to service only when the Division Yard is empty, so the supply genuinely runs down, and a game that never showed either yard gave no warning at all.

It shows the pressure immediately. In a finished game the Division Yard held 30 cars — hoppers, tanks and cabooses — and no boxcars, coaches or reefers at all, while 19 of them sat in Classification unable to come back, because the Division Yard was not bare.

Crew Tray scarcity was already implemented and is left alone: a train with no free tray is held (trainHeld), which is §7.

The Classification Yard rule, from the source — and my guess was worth 2.4 Revenue it should not have been

Answered: used Rolling Stock is set out in the Classification Yard, used engines and cabooses go straight back to the Division Yard, and the Classification Yard empties only when the Division Yard is bare — then all of it returns at once.

That is a much harder rule than the one I invented. Returning cars at every DAY boundary keeps the yard topped up continuously; this lets it run down to nothing and refill in one go, which is the whole of the supply pressure the game is meant to have. Measured paired over 400 seeds:

my Day-boundary guess 9.67   the real rule 7.25
paired change  -2.42 ± 0.49   (t -9.67)   235 of 400 seeds affected

So the +2.32 celebrated when the Classification Yard was first made readable was very largely an artefact of getting the trigger wrong. The refill is now checked wherever a car leaves the Division Yard, so it fires the moment the yard empties rather than at the next convenient tick.

Poling is dealt zero copies rather than deleted. Its effect is "TBD in the source", so there is nothing to implement and a card that cannot be played is worse in a hand than absent from the deck. The catalogue entry stays so the gap remains visible. Deck 140 → 139, solitaire 118 → 117.

Heavy Grade orientation stays rolled from the seed, and is now documented as temporary in the code rather than only in TODO: the card says the player sets it, but it is dealt during setup and setup has no decision point — createGame is a pure function of the seed, which is also what makes a save portable.

A counting question this raised, and could not answer. A census of every holder of rolling stock comes to 92 against the 80 dealt. That is not proof of duplication: outboundBox, inboundBox and menAtWork all hold RollingStock, and stocking a green box takes a LOADED CAR out of the Division Yard — so some of those objects are cargo in transit rather than cars, and nothing distinguishes them. An accounting test was written and then withdrawn, because it could not tell the two apart. Logged: until a load is its own type, "is any stock being created or destroyed?" is unanswerable.

The freight figures were counting one half of freight

A measurement fix, not a game fix — but it is the instrument every balance decision is read from.

rev.freightUnload was assigned eventCounts['unloadBegan']: unloads STARTED, not Revenue EARNED, and the two differ by every unload that never finished. grossFreight then used freightLoad alone, so freightShare omitted the unload half outright. A completed load and a completed unload each earn a point, on two distinct revenueChanged reasons.

The comment that stood there claimed "an unload scores through the same event as a load completion in the reducer". It does not — apply.ts emits freightUnload separately. A comment asserting a fact about code a few lines away, and wrong.

freight share of gross:  39%  ->  49%

The harness now prints both halves. strategyBuckets counted "a scoring game" the same wrong way, and so did the test guarding it — so a game that scored only by unloading was bucketed but not counted, which is how the fix first showed up as a failure.

This matters backwards as well as forwards: the "freight is only 13-18% of gross" finding was read from this number, and it is what drove the Gap 12 industry-density change. That decision was taken against an instrument reading roughly 60% low.

Turnouts that go nowhere — reported, confirmed, and mostly not the problem

Reported from two replays: of 8 turnouts off the Running Track, 2 formed a run-around, 2 served industries and 4 went nowhere. Measured across 120 games, that holds exactly:

what a turnout leads to
part of a run-around	30%
a stub, but serves an industry	12%
a stub ending in bare track	44%
nothing below it at all	15%

4.61 wasted turnouts a game. Then three attempts to stop it, each measured paired over 400 seeds, and each WORSE than leaving it alone:

attempt	paired change
no new way down while one leads nowhere	run-arounds → 0	deadlocked: a run-around needs TWO ways down, and the second cannot be justified by what hangs off the first
first two free, gate the rest	−0.84 ± 0.53 (t −3.08)	track spend collapsed 15.4 → 4.8
forbid rail that butts an incompatible card	−0.62 ± 0.57 (t −2.11)

The reason is that a turnout is not only a way DOWN. It is also a way UP, and both the east-west extension and the closing arc are gated on one existing beyond them — so cutting the turnouts cuts the places a siding can rejoin, and the sidings stop forming too. The apparent waste is optionality. This also re-confirms, with proper statistics, a note left in the code by an earlier attempt.

Rail that can never go anywhere

The one that did work, and only as a tie-breaker.

A port facing an EMPTY square is a promise: something may be built there later. A port butting an OCCUPIED square whose card has no matching port is not — that square is taken, so the rail stops dead and always will. Reported from seed 618682, where an arc came off a turnout with its far end jammed into a curve that could not accept it. Measured: 28% of all pieces laid, 4.26 a game.

Forbidding it cost 0.62 revenue a game. Applying it as a tie-breaker on the distance score — never able to veto a piece, only to choose between two the heuristics rate equally — measured +0.43 ± 0.49 (t 1.74), with 108 seeds better against 67, and cut these from 28% of pieces to 7%. Not significant on its own, but it is the only one of four attempts pointing the right way, and the mechanism is sound.

And it fixed the reported problem after all — sideways. Re-running the turnout taxonomy:

what a turnout leads to	before	after
part of a run-around	30%	66%
a stub, but serves an industry	12%	13%
a stub ending in bare track	44%	2%
nothing below it at all	15%	18%
wasted per game	4.61	1.65

Bare stubs all but gone and run-arounds more than doubled, without ever refusing a turnout. Refusing them directly had destroyed the run-arounds; declining to lay rail INTO a dead end leaves the bot free to cut every turnout it likes and quietly stops it building the stubs. The remaining waste is the last turnouts of a game, cut with no turns left to build beneath them.

Also added, and honest about it: a turnout is not cut when no arc remains to hang beneath it. Measured at 0% today — the bot lays the arc immediately after the turnout and never runs the supply dry — so it is a guard against the supply changing rather than a fix for anything happening now. Its test constructs the situation by draining the arcs.

A source file git was hiding

Found while checking the above: src/web/replays.html had never been committed. .gitignore carried replay*.html to catch the throwaway files generated at the repo root, and unanchored it also matched a source page — the website's replay viewer. git archive HEAD confirms it: a fresh clone does not contain that file, and build-web.ts copies it unconditionally, so the build would have failed for anyone but this working copy.

The pattern is anchored to the root now (/replay*.html), which still ignores the generated files and no longer ignores the source. A test asks GIT — not .gitignore — whether each source page would survive a clone, because that is the actual question.

Three places draw a game, and they had drifted

Reported from playtesting: the replay had lost its "extra slow" speed, and neither the sound nor the auto-hide could be found. Both true, and the same cause — a game is drawn in THREE places and only some of them had kept up:

	speeds	sound	auto-hide
the playable page	—	yes	yes
the standalone replay file	5	yes	yes
the website's replay viewer	3	no	no

The website viewer is its own implementation — it replays a save through the engine in the browser rather than reading a rendered file — and it never got what the other two grew. It now has all five speeds (extra slow through very fast), the sound, and the auto-hide, using the same shared cuesFor and playCue as everywhere else.

Three tests hold them together from now on: the two viewers must offer the SAME set of speeds, all three pages must carry the sound and auto-hide controls, and replays.ts may not ask for an element its page does not have — the same total check the playable page already had, and the one that would have caught this. Verified by removing a speed and a control and watching them fail.

The published replays had stopped replaying

Both saves in public/replays/ were dead. A save is a seed plus the intents, replayed through the real engine — so it cannot describe a position the rules could not produce, and an intent that no longer applies stops the replay rather than being forced. That is the safe direction, but it is silent: seed-202 got 42 intents into 360 before halting, and seed-430 managed 4 of 338.

They were recorded before this run's rules work — Modifier hosts, the industry-to-car mapping, the Classification Yard — so most of what they described is no longer legal. Replaced with three games generated against the rules as they stand, each verified to replay every intent to the final Day:

A winning run	36 Revenue, seed 1038389
A strong run	32 Revenue, seed 618682
Collisions	10 Revenue and 27 smashes, seed 919604

The third is deliberately a bad game: a full Office is a collision, and it costs more than the freight was worth.

This is the "save/restore is not version-aware" item in TODO doing exactly what it warns about. The saves are cheap to regenerate, so the fix is not to freeze them — it is for a stale save to say so instead of quietly ending early.

The replay behaves like the game it is replaying

Sound and the district auto-hide were built for the playable page and the replay had neither, which made watching a game back a poorer experience than playing it. Both are there now, and both are the SAME implementation rather than a second copy:

• cuesFor decides what happened — a Stage ended, a Day turned, a train was built — and is imported by the live game and called by the replay recorder, so the two cannot disagree about when a Stage ended. Frames carry their cues.
• playCue decides what that sounds like. It was three module-level functions with a shared AudioContext; it is now one self-contained function, embedded into the replay by toString() exactly as the board renderers are. The context is parked on window because the embedded copy has no module scope to keep it in.

Sounds fire only when stepping FORWARD one frame. Scrubbing across a hundred frames would otherwise fire a hundred whistles at once, and stepping backwards would sound a Stage ending that is being un-done.

Both default to the quiet, tidy setting: muted, and auto-hide on.

The replay's DOM stub had no classList, so the page threw while rendering rather than folding a panel. It has one now — the same gap the playable page's stub had, and the same fix.

Playtest fixes

The end-of-line labels were cut to three letters. "trains enter" and "trains leave" hang off the ends of the route, and the canvas padding was sized for the buffer stops alone — so the west end read "TER" and the east "tra". Padding now allows for the words.

"End my turn" became "End Local Operations". It ends the phase, and the old wording invited the reading that it ended something smaller.

The hand limit is a limit, not a toll on drawing. Drawing a card set "you must now play one", so a player who had already played two cards and drew back to three was still forced to spend one. §6.2 is a hand LIMIT — "reduce his hand to no more than three cards", four with a Red Flag — and that is now the only thing that blocks the end of a turn. It is derived from the hand at each render, and a test asserts the page and the engine never disagree about whether the turn may end. The engine removes draw.end outright when the hand is over the limit, so the page draws a disabled button naming the reason rather than silently offering no way out.

A train now says what its card calls for. Reported on seed 22222: Extra X22 offered a caboose and nothing else with no reason given. The rules were right — "Pee-Dee" is a per-diem train whose consist is one caboose and no cars, and §8.2 forbids the wrong cars however few are carried — but the screen never said so, which reads as a broken game. The New Train header now names the train and its consist: Making up Extra X22 "Pee-Dee": its card calls for 1 caboose — Per-diem train. May only pick up MTs.

The auto-hide control said what the panel was doing, not what pressing it would do. "auto · folded" reads as a status line and was missed; it now reads "auto-hide: on — click to keep open", and looks like a control rather than a caption.

The whistle is twice as long, and sound now defaults to OFF.

A legal move you could not click

Reported from playtesting: laying a straight offered three places in the list and highlighted one on the board.

Empty squares were drawn by a separate ghostSvg and spliced into the SVG afterwards. It computed its origin over cells PLUS the offered spots, while officeSvg sized itself over cells alone — so the two disagreed the moment a legal square lay outside the played cards, which is every square that would EXTEND the district. Reproduced on seed 555: three placements offered, the empty one drawn at y=99 on a canvas 103 tall, i.e. off it entirely.

officeSvg now takes the ghosts and sizes itself over both, so there is one origin and one canvas, and ghostSvg is deleted rather than fixed — a second coordinate system was the bug, not a detail of it. It was also imported by the replay and never called there.

A test walks six seeds and asserts every square the menu offers has a target drawn inside the canvas: 129 squares across 47 placeable subjects. It fails on the old code with the exact coordinates above.

The railroad, heard

A whistle at the end of each Stage, the grade-crossing bell at the end of each Day, and the conductor when a train is built. Off by default — everything here is synthesised rather than recorded, so it is a placeholder for real audio and a playtester who did not ask for noise should not get any. One click in the title bar turns it on, and that click doubles as the gesture browsers require before audio may start.

Synthesised, not sampled, and that is a constraint rather than a preference: the site is a static folder that fetches nothing — there is a test asserting no page reaches an external host — so audio would have to be committed to the repo, and a plausible whistle is not something to invent. A real steam whistle is a CHORD of several chambers slightly out of tune with each other plus the breath of the steam, which three detuned partials and a band of filtered noise get most of the way to. The bell is inharmonic partials struck twice, which is what separates a bell from a beep.

"All aboard" is speech, and speech cannot be faked with oscillators. speechSynthesis is built into the browser, needs no asset and works offline, so it says the words; where the platform has no voice installed a two-note conductor's call takes its place rather than nothing happening.

The model names WHAT happened — a Stage ended, a Day turned, a train was made up — and the page decides what that sounds like. A Day boundary rings the bell only: both would collide, and the bell is the bigger event. The first Stage of the game announces nothing, because a Stage beginning is the previous one ending and there is no previous one.

Counted against the clock rather than trusted: over a full game, 60 Stage boundaries produced 55 whistles and 5 bells, and 5 Days produced 5 bells.

Regions, drawn from what the crossing already cost

§2.1 divides a Mainline card into two regions and §8.2 moves a train one region per Stage. The engine had replaced that with crossingStages() — implications.md records it plainly: "The Region model is gone" — because ten card types with real speeds cannot be expressed by one region a Stage. Both REGIONS_PER_MAINLINE_CARD = 2 and entryPoints survived as dead constants.

The map draws regions again without reinstating the mechanic, because the printed cards say how: they carry Start positions, so a train with Brakemen enters further along the card. That is the same fact as "takes a Stage off the crossing", in different coordinates. So position falls out of what the engine already knows:

entry = REGIONS - stagesTotal        position = clamp(entry + elapsed)

crossing	drawn as
1 Stage (a 60 card)	enters at region 2, gone next Stage	the printed Start position
2 Stages (a 30 card)	region 1 → region 2	§8.2 exactly
3 Stages (slow train)	region 1 → region 1 → region 2	fixed distance, slow train

entry is deliberately allowed to go negative and only the final position is clamped: that keeps a slow train's extra Stage at the START, where being slow shows. Clamping the entry instead parked it at the exit, reading as a train that raced across and then waited — which is what the first version did, and what the test now forbids.

Transit gains stagesTotal, set on entry. State, not rules: nothing reads it to decide anything, so the bot and every balance figure are untouched, and a save is a seed plus intents so there is nothing to migrate. It cannot be recomputed later — a modifier played onto the card mid-crossing would change the answer and make the train jump backwards.

Division Points draw one region, the queue. Running Track cards inside a district draw none: a crew moves there by Moves, not Stages, so it occupies a card outright.

The Division map shows the whole route

The map drew one box per node, which collapsed each player's entire district into a single "Office" tile — so the track a train actually runs along was invisible on the only view that shows where trains are.

An Office now expands into its Running Track, Limits to Limits. That is the right cut rather than a compromise: the through route IS the Running Track, and everything hanging beneath it is secondary track a crossing train never touches. Sidings, industries and the load pipeline stay in the district view, which is where they can be read.

West DP · Mainline · [Limits … Office … Limits] · Mainline · [ … ] · Mainline · East DP

Trains are shown wherever they are. On a Running Track card, on a Mainline card, queued at a Division Point. A crew working BELOW the Running Track has no position on the through route, so it is reported against the district — "2 switching below" — rather than drawn somewhere it is not. Projecting a siding onto the through line would be a lie the map cannot support.

Seating. Players sit around a table, so the route is laid out the way they do: one row alone, two rows facing, a horseshoe of three, a square of four. The Division is a LINE and not a loop — trains enter at one Division Point and leave at the other — so the shape is deliberately left open, with buffer stops at both ends and the gap labelled "trains enter" and "trains leave". Closing it into a ring would promise a connection the rules do not have.

Layout is geometry with no visual feedback loop, so it is tested rather than eyeballed: for 1, 2, 3 and 4 players no two cells may overlap and none may fall outside the canvas.

The district folds itself away

Auto-focus. The district is worth its vertical space during the phases that change it — Local Operations and Cargo — and not during New Train, Mainline or Supervisor Shift, where the Division map is what matters. Folded, it leaves a summary line rather than vanishing, because a panel that disappears entirely reads as broken.

A manual toggle overrides it and stays put. This is DISPLAY state and never game state: two players at the same table may reasonably want it set differently.

The replay was storing the same things over and over

Halved, near enough: 3415 KB → 1877 KB on a 735-frame game.

The TODO said to carry links forward. Measured first, that would have bought 5% of the cells payload — cells is 62% of the file, and inside it the cost is elsewhere:

what       32% of cells   long prose, identical on every turnout in the district
facility   24% of cells   the SAME object already serialised in the frame's `facilities`
identity   26% of cells   row/col/kind/label/running/links, fixed once the card is laid

So all three are interned. A card's identity and its description are written once for the whole recording and referenced by integer, and a cell points at its facility instead of carrying a second copy of it. What stays per-frame is what genuinely changes: the cars standing there, the crew, and any Enhancement laid on the card.

rehydrateCells is exported and emitted into the page by toString(), the same trick the two board renderers use — a second copy inside the page's inline script could drift from the packing and the symptom would be a board drawing the wrong cards rather than an error. The round-trip test runs that exact function over every frame, resolving nulls the way the page does.

Rolling stock was leaving the game

Four things were tried against bot revenue. One of them was worth more than everything else in this changelog combined, and it is the one that had been ranked third and predicted not to matter.

The Classification Yard was write-only. Seven places pushed cars into it — retired trains, collisions, unjams, set-outs — and nothing in the engine ever read it. Rolling stock drained one way out of the game: 30 cars dead by the end of a game, 37% of the 80 dealt at setup. Gap 2c says engines and cabooses return to the Division Yard and everything else to Classification; the recovered rules never say how Classification empties.

ASSUMPTION, flagged rather than derived: sorting cars for redistribution is what a classification yard is for, and a Day is its natural cycle, so they now return at the Day boundary. This is a rules decision that wants confirming against the source.

Paired over 400 seeds: +2.32 ± 0.52, t = 8.79, 194 seeds better against 44. Revenue 6.70 → 9.02.

Why it was mis-ranked is worth recording. The Division Yard does not run dry in five Days — 16.6 loaded freight cars remain, empty in 2 games of 100 — so it was reasoned that supply could not be binding. The aggregate was never the point: what starves freight is not having the RIGHT commodity at the moment a green box needs stocking, and returning classified cars keeps the mix alive.

Three things that did not work, kept because the measurement is the result

Measured paired over the same 400 seeds, which is the only way to see effects this size.

Capping the draw option: worth nothing. 62% of all Local Operations actions went to drawing and 12.6 of 29 cards drawn were discarded, so a draw into a full hand converts straight into a discard. Refusing it: -0.10 ± 0.13, and 379 of 400 seeds byte-identical. The branch almost never fires.

Pairing the two halves of a load: worth nothing. Sampling every outbound industry at every loadUnload phase, 76% of the time it had NEITHER a stocked green box NOR a spotted car, and only 5% of Stages had a single workable facility. Letting the bot switch for a stocked box without waiting for a train at the Office changed nothing measurable — folded into the -0.10 above. The diagnosis was right and the prescription did not address it: there is usually nothing to switch.

Waking a dead branch made things worse. chooseLocalOption tested options.some(i => i.type === 'mainline.modify'), but that intent requires s.turn.option === 'draw' and the test runs while the option is still null — legalActions had already filtered it out, so the branch could never fire and never had. Rewriting it to check the hand cost -0.64 ± 0.32 (t = -3.90), 104 seeds worse against 35. Its comment claimed the value compounds like a train card's; it does not. The branch is now deleted, with the measurement in its place.

The Enhancements other than Interlocking are worth exactly nothing — and cost nothing either: -0.01 ± 0.41 (t = -0.04) when the bot is forbidden to place any of them. Left as they are; there is nothing to gain by restricting them. Note the first attempt at this experiment showed 0/400 seeds changed, which was the experiment failing rather than the answer: a card.play fallback with no placement filter was still placing them.

	before	after
revenue (200 games)	6.7	8.7
wins	6.5%	12%
freight loads	2.8	3.2
cars dead in the Classification Yard	28.4	~0

A Running Track with nowhere to put an Enhancement

Every penalty in the game has one cause. Across 100 games, all 48 were collision: no free A/D track — 2.70 revenue a game, 27% of gross, concentrated in about a fifth of games and responsible for the −47 tail.

The rules already answer it. Interlocking prints "may stop an inbound train on the Limit Track", and advance.ts:621 holds the train at the Limits instead of colliding. It had never once been placed. Nor had any train ever been held at the Limits.

The reason was structural and nothing to do with Interlocking. The bot builds minimal two-arc run-arounds — an ne arc meets an nw arc directly — so it never needed a straight and laid none: 0.00 straights on the Running Track across 100 games. Interlocking, Water Column and Telegraph all require one; Yard Office and Small Yard want a Secondary Track Straight; Telephone and Radio chain off Telegraph. Thirteen of the eighteen Enhancement cards that go on the board were unplayable. They were drawn 3.78 a game and placed 0.64.

bestTrackLay now scores one straight onto the Running Track. Enhancements placed went 0.64 → 3.01 across nine types where only Overpass had ever appeared, and Interlocking now reaches the board in about a quarter of games.

On whether it pays — measured properly, and the answer is qualified. Revenue per game has a standard deviation of ~9, so a 100-game run carries roughly ±1.0 of noise. Run against the same 400 seeds, paired:

	without	with
revenue	5.99	6.70
collisions cost	2.70	1.91
worst single game	−47	−24
most collisions in a game	10	6
wins	5.3%	6.5%

Paired per-seed the change is +0.70 ± 0.74 (95% CI), t = 1.87 — not significant on its own. And 154 seeds improved against 165 that got worse: the mean gain comes from removing catastrophes, not from making a typical game better. What justifies keeping it is that the mechanism is measured directly and accounts for the whole effect — collision cost falls 0.79, and revenue rises 0.70.

A correction to the numbers already in this file. The per-100-game revenue figures reported for the earlier changes carry the same ±1.0 noise, so the individual steps (5.0 → 6.0 → 6.5) were stated more precisely than the sample supports. The cumulative move from 3.2 to ~6.7 is far larger than the noise and stands; the individual increments should be read as indicative only.

The bot was throwing away its own freight

Routing turned out not to be the problem, which is why it was worth measuring first. Of 1650 drops across 100 games, 940 (57%) landed on a facility that wanted the car and zero landed on one that did not. The crew makes 9.4 correctly-targeted drops a game; the one-move-only destination test costs nothing measurable. "70% of waiting loads have nothing spotted" was a misleading signal — the cars arrive.

Following the freight from the other end found the leak. Loads were being destroyed:

stockToOutbound   9.45/game
loadStarted       2.71/game
facilityUnjammed  3.10/game  from outbound  <- healthy waiting loads, discarded

facilityUnjammed from outbound splices the load out of the green box and pushes it to the classification yard. That load cost a Local Operations action to stock, so discarding it is strictly negative — and the bot did it more often than it started a load.

Two fallbacks meeting, and the root cause is a familiar one: two rules for one act. canStockProductively decided the Freight Agent option was worth taking if a matching empty car was spotted. The engine's stockOutbound additionally requires a loaded car of that commodity in the Division Yard — which the predicate never checked. So the option was chosen believing a box could be stocked when none could; the follow-through then found nothing stuck, nothing to clear and nothing stockable, and fell through to "clear whatever is stuck" with nothing stuck.

Fixed by making the predicate ask the same question the engine does, by no longer using Freight Agent as the idle default (switching at worst moves the crew toward the Office, which a train must reach to depart at all, §8.1), and by ordering the last-resort unjam by what it costs to lose — MEN|AT|WORK first, then the red box whose Revenue is already banked, and the green box last.

	cars fixed	freight kept
loads discarded from a green box	3.10	0.00
revenue	6.0	6.5
wins	5/100	8/100
trains scheduled	3.0	3.4
cards played	16.6	19.3

The gain is development, not freight. Loads started held at 2.70 and freight revenue at 2.6 — the recovered Local Operations actions went into drawing and switching rather than into the freight chain. Green-box stocking fell 9.45 → 6.34 because the bot no longer stocks boxes it cannot serve. The regression test asserts outbound unjams stay at zero and that genuine MEN|AT|WORK jams are still cleared, so gutting the fallback would not pass it.

Half the industries never asked for a car

Freight had not moved through two rounds of fixing the district, and this is why: three of the six industries were invisible when the bot chose what to put on a train.

wantedCars consulted a hand-written industry -> car type switch that had drifted from the sheet. It named produceShed and oilRefinery — neither is an industry — and omitted freightHouse, refinery and packingSheds, which are. An industry it could not name returned null and was skipped entirely, so it never requested a car. The Refinery is the only source of tank traffic, so tank cars boarded a train 0.07 times a game and were dropped by a crew zero times in 100 games, while 23 of 79 waiting loads sat at an industry that wanted one.

It now derives the commodities from INDUSTRY_PROFILES, which is the sheet. The switch is deleted rather than corrected — a second copy of the mapping is the bug, not the values in it.

A second, narrower collapse. facilityCarType returned carTypes[0], so the second commodity of a two-commodity industry was unreachable: a Power Plant burns coal OR oil, a Grocer's Warehouse receives dry goods OR perishables. facilityCarTypes (plural) now returns the full set, and the bot's spotting and switching checks accept any of them.

Worth stating precisely: the engine's WRONG_CAR_TYPE gate was corrected to use the full set too, but that changes nothing today — both two-commodity industries are inbound-only, so freightAgent.stockOutbound rejects them before the commodity is examined. That fix is latent. The measured gain is entirely the bot side.

	facilities fixed	cars fixed
revenue	5.0	6.0
wins	1/100	5/100
freight revenue	1.3	2.6
freight share of gross	25%	37%
loads completed	1.27	2.60
tank cars dropped	0.00	0.62
reefers dropped	0.04	0.67

Both regression tests were checked against the bug they guard: restoring the stale map fails the tank assertion, and collapsing carTypes to its first entry fails the profile assertion.

Still 70% of waiting loads have nothing spotted at all. The commodity mix is right now; the volume reaching the industry tracks is not. That is a routing question — which facility the crew takes a car to — rather than a car-choice one.

Putting the industries on the siding

The run-arounds were being built and the industries were somewhere else — facilities sitting on one stayed at 0.00 a game even at 91 districts in 100 with a closed loop. The cause was blunt: facility placement was options.find(i => i.placement !== undefined), the first legal square the generator happened to list, unscored, while track laying had sixty lines of scoring beside it.

Facilities are now scored, on the thing that decides whether a crew can serve them at all: a placement in line with a siding still being built becomes part of the loop itself (its own through track is a segment), so the crew reaches it from either end and can pass its standing cars (§A.5). Merely touching reachable track is worth less; the Running Track is a penalty, because a car left standing there is hit by the next arrival (§11.2).

A second bug surfaced immediately, and it is the interesting one. Scoring facilities onto the siding row sent run-arounds down, 91 games in 100 to 36 — the industry took the square and the loop stopped closing around it. runsAcross tested the card's KIND ("a track straight running east-west"), so an industry standing in the line read as a dead end and the run refused to extend through it. It now asks the card's PORTS instead. A Facility carries its own rails (§11.2), and so do the Office and a Limits sign; what matters is whether a port faces this way.

The reachability walk is now shared between track laying and facility placement rather than written twice — two copies would eventually disagree about whether a district connects, which is the one thing both decisions rest on.

	before sidings	sidings fixed	facilities fixed
facilities on a run-around	0.00	0.00	1.08
games with a run-around	0/100	91/100	71/100
revenue	3.2	4.1	5.0
collisions	—	—	0.4 (was 0.6)
track pieces spent	19.5	16.1	13.6

Run-arounds fall from 91 to 71 because facilities now compete for the siding squares — which is the trade being made deliberately: a loop with an industry on it is worth more than an empty one.

Freight did not follow. Loads completed 1.42 → 1.27 and freight's share of gross 31% → 25%; the revenue gain is passengers and fewer collisions. Of facilities holding a load, 77% still have nothing spotted at all. The industries are now reachable and the right cars still are not arriving — which is the car-selection problem in TODO, untouched by any of this.

The bot was building stubs, not sidings

Measured first: 0 run-arounds in 100 games. A run-around is the engine's own definition of a useful siding (track.ts) — double-ended, both ends reaching the main, and §A.5's facing-point move is impossible without one. Every district the bot built was dead-end stubs, 3.86 of them a game, plus 2.89 cards below the main that reached nothing at all. Before trusting a zero the detector was handed a run-around built on purpose and found it from both ends.

Three bugs, all the same shape: scoring on local form without checking it reaches anything.

• The +12 rule was commented "close the loop back up to the main: the run-around is complete" and only tested that a neighbour ran east-west — never that the card above had a south port to join. The siding terminated in an arc pointing north into empty space. It now requires a way up above, asked of the engine's hasPort so the Office counts too; divergesSouth had looked for a turnout and missed the one way down that is on every board. An arc's facing also decides what it meets — nw joins west, ne joins east — so closing from the wrong side connected nothing.
• The east-west extension had no stopping condition, so the run went on past the last column it could rejoin at, in 96 of 100 games. The loop then missed by one card.
• bestTrackLay never declined. This was the one that actually mattered, and the first two fixes barely moved the overshoot without it: the function returned its best-scoring option unconditionally, so once the useful squares were taken it kept laying track because track was legal. Bonuses are now tracked apart from the distance score, and a piece that earns none is not laid — the Stage falls through to playing a card instead.

Anchors also have to be reachable from the main now. A stranded east-west straight made both its neighbours look like legal extensions, so a fragment joined to nothing grew in both directions.

	before	after
games with a run-around	0/100	91/100
track laid east of the last way up	96/100	0/100
track pieces spent	19.5	16.1
revenue	3.2	4.1
freight revenue	2.6	3.8
cars dropped	11.6	20.9
loads completed	0.99	1.42

Two regression tests, both walking the district with the engine's own exitsFrom so they cannot credit a connection §A.1 forbids: one asserts run-arounds get closed, the other that no track is spent east of the last column with a way up.

Still open. Facilities sitting on a run-around: 0.00. The loops get built and the industries are not on them, so the run-around is not yet paying for itself in freight — which is the next thread, not a finished one.

Freight, drawn where the work happens

The load pipeline moved onto the card. A load crosses green → MEN | AT | WORK → a spotted car, and that journey is freight. It was drawn only in the side panel, so a Laborer action — the whole of the freight game — changed nothing on the card the player was looking at. The squares now sit under the track, which is where the printed cards put them and why they are printed at all. The tooltip names the same thing in words: which square the load is on, and what the next Laborer action does with it.

A collision found while placing them. overpass and facingPointLocks are placed onCard, so they can land on a facility, and the enhancement label's baseline ran straight through the new squares. The label moved into the gap between the crew tray and the pipeline; a test now asserts the two do not overlap rather than trusting the two constants to stay apart.

Teaching the bot to use a siding

Nose coupling, which the rules had and the engine did not. §A.3: "engines also have couplers on the front end, so a train can pick cars up onto its nose". Coupling always appended to the back, so which end cars landed on did not exist — and that is precisely what a run-around is for. Cars met running FORWARD now couple in front, cars met BACKING couple behind, so the approach decides which car is next off the tail:

running forward -> reefer, boxcar, hopper   next off: hopper
backing up      -> boxcar, hopper, reefer   next off: reefer

The bot prefers a run-around to setting a car down, since it keeps the car — but only when the drop can actually follow. Without that guard it ran the loop for its own sake (8 a game, 63 Moves), which the shuttling regression correctly failed.

A serious bug found on the way. The carsCoupled reducer cleared every card in the Office Area, not the ones the crew ran over — its own comment said "every card along the path" while the code iterated the whole grid. One coupling anywhere deleted every standing car and every industry track in the district, so loads worked over several Stages vanished when a crew picked up an unrelated boxcar somewhere else. The event now names the cards it lifted from.

A test replaced rather than relaxed. "Under 60 Moves a game" started failing. That threshold was calibrated when the crew coupled ~0.4 cars a game; it now couples ~8 and drops ~11, and a run-around is supposed to cost several Moves. Counting Moves was always a proxy for aimlessness, so the test now measures the thing itself — Moves per drop-or-coupling — which still fails when the bot is made to wander.

	before sidings	now
revenue	3.9	3.2
freight	0.9	1.0
drops per game	3.3	11.1
couplings per game	~0.4	7.9

Switching is transformed; revenue is not. The crew now does real work — roughly 4 Moves per productive act, which is what running a loop costs — but that work is not yet converting into Revenue. Where it goes next is an open question rather than a known fix.

---

Board rendering, three-page site, curve geometry

Drawing the board as track, splitting the site, tooltips, and the curve fix.

Board rendering

Two renderers in src/sim/board-svg.ts, chosen because they fail in opposite places:

• Office Area — the district stays a map. Cards on a grid with the rails drawn edge to edge, so a join is rail meeting rail rather than two descriptions that happen to agree. The through rail sits at a constant height on every card, which is the alignment the printed cards use.
• The Division — not a map but a queue of sections with hard capacities, so it is a dispatcher's diagram: one line per track, 1 of 2 free under each section, no limit — trains queue at the Division Points.

Both are self-contained — no imports, no module-level helpers — because the playable app imports them normally while the replay is a single HTML file with an inline script that cannot import anything, and embeds them via Function.toString(). One implementation either way; a second copy would eventually draw a different board from the same state.

Rails come from the engine's own connectionsFor, now exported. A drawn rail cannot claim a connection the rules do not have — visible in that Modifier cards draw no rails at all.

Capacity, confirmed from the rules and now shown: Division Points unlimited, most Mainline cards 1, Double Track and Uncontrolled Siding 2, Offices 1/2/3/4 by tier.

Three pages

index.html is now a splash with two doors. The game moved to play.html; replays.html is a directory.

A replay is a save, and a save is {seed, history}. The engine is deterministic and runs in the browser, so re-submitting the same moves rebuilds the position exactly — 18 KB instead of 3.4 MB, about 190× smaller, small enough to email. A save that would describe an impossible position cannot be replayed at all, because every step goes back through applyIntent; the viewer stops and says so rather than showing a board the rules could not produce.

Static hosting cannot list a directory, so replays/manifest.json is generated at build time from whatever is in public/replays/, with each file validated first.

Tooltips

Reference detail is read once and printing it costs the space the board and action list need. A single delegated tooltip (src/web/tooltip.ts) now carries card effects, action reasoning, and facility state. Not the native title: that waits a second, cannot be styled, and never appears for keyboard users.

Curve geometry — the significant fix

A curve was modelled as a through track plus a diverging leg, which is a turnout. Consequences:

• Curves were topologically identical duplicates of turnouts — same connections, no distinction beyond the sharp curve's 2-Move cost.
• Every diverging leg went south, so no piece anywhere reached north except an n-s straight, which connects n↔s and nothing else.
• Therefore a district could only ever be a vertical column: no siding, no parallel track, no run-around, no second turnout back to the main.

The printed cards (docs/tracks.png, rows 3–4) show a curve as a single arc from edge to edge with no through track. A curve is now a two-port arc, rotatable to ne/nw/se/sw. A sharp curve is geometrically identical and costs 2 Moves. Turnouts keep §A.1 unchanged — a two-port arc has no third port, so the absent-edge rule cannot apply to it.

Verified through the engine, not the types: a crew leaves the main at a turnout, runs a siding parallel, and rejoins at the far end.

Measurements

	revenue	freight	sidings built
before	4.1	0.5	impossible
geometry only	3.9	0.9	possible, not sought
bot builds sidings	3.3	1.1	59/60 games

Freight more than doubled and the harness recorded its first win, but overall revenue is down from geometry-alone and the bad tail worsened (−29 → −59).

Capping turnouts at one or two measured worse (revenue 2.5, freight 0.6) — more ways off the main means more industries the crew can reach, and that beats a tidy Running Track. The uncapped version stands, with that measurement recorded at the code.

The bot builds sidings but does not exploit them. It pays about 20 of its 26 track pieces for them while its switching logic still sets out dead weight on a spur rather than planning a run-around. That is the next piece of work and where the revenue should appear.

---

Earlier commits

Recorded from memory of the work rather than written at the time; detail thins going back.

f00255c — more fixes and tweaks

Card descriptions on every card in hand and every face-up slot, the three Local Operations options explained, the objective and pace in the header, and rotations named in words rather than "rotation 2". Build stamp added (version, git SHA, -dirty for an uncommitted tree) because nothing tracked what was deployed. Extra X22 was not a bug — a per-diem train whose card calls for a caboose and nothing else — but "loaded caboose" was.

d1f689d — name the caboose, the train and the Running Track hazard

maneuver.redFlags was labelled "Red Flags on tray3": the earlier tray-id fix checked the history log, and the action list was a surface it missed. An industry on the Running Track does not block traffic (§11.2 gives it rails) but a car left standing there is hit by the next arrival (§10).

dd300ac — fix caching, Limits placement, and explain the board better

The first deploy served a fresh index.html against a cached main.js, which threw missing element: target and never started. Every module import now carries the build tag. The Limits fix was half-done: the sign could move, but nothing forbade building past it, so one straight at (0,2) produced limits · Whistle Post · limits · straight · limits.

2eca9de — playable browser build

The solitaire game as a static site, proven to need no server: full games run with every Node global replaced by a throwing stub. Train cards stopped accepting a board placement — seed 555 had offered Extra X15 at six squares with six rotations, all identical.

160190d — the bot's reasoning in the replay

Decision panel showing what the bot chose, why, and every option it passed over, with the reasons reported by the bot itself rather than re-derived by the viewer.

d261ad7 — parking trains, freight deadlock, Whistle Post lock-in

Three faults found by measurement rather than failing tests. Trains parked because destinationsFor computed reverse as facing === 'e' ? 'w' : 'e', so a crew facing south reversed to east — a port a north-south card does not have. Freight ran at a 3% load completion rate because a load started with no spotted car parks on WORK and locks the industry track, blocking the very car that would clear it. Office density doubled after 25 of 100 games never drew a Depot and never escaped a one-track Whistle Post.