TheLadder/tools/story-to-pack/probe/poe/README.md

Poe reference set: The Works of Edgar Allan Poe, Volume 2 (tales)

A hand-made comparison baseline for the story-to-pack pipeline, the fourth beside O. Henry (../v2/claude/), Wharton (../wharton/) and Jacobs (../jacobs/). Claude wrote every summary, predicament, group and option here by reading the scenes. No local or hosted inference was used. Compare the local-model runs against it. It is not ground truth, just one careful reading.

File	What it is
corpus/tales-vol2.txt	Project Gutenberg #2148 (Raven Edition, vol. 2), plain text, public domain
split_gutenberg.py	→ stories.json (22 tales) and chunks.json (369 scenes, median 258 words), using the same 320-word paragraph rule as ../chunk.py. Titles are matched against the upper-case headings after normalising, because the headings differ from the contents list: trailing periods, a curly apostrophe, "MAELSTROM" without the ring, and "SILENCE—A FABLE" with no spaces. The front matter, the redactor's note, "Notes to the Second Volume" and the licence are dropped.
dump.py	python3 dump.py A B prints scenes A..B for reading (copied from ../jacobs/)
claude/batches/sNN.json	per story: {chunk, story, summary, predicament}, the same schema as the sibling probes
claude/groups_src.py	the groups, options and per-story protagonist standing
claude/build.py	checks everything, then writes claude/records.json and claude/groups.json. It is identical to Jacobs' copy, since its paths are relative to the file.
claude/page.py	→ claude/REFERENCE.html, the readable page

Rebuild: python3 split_gutenberg.py && cd claude && python3 build.py && python3 page.py.

What is in it

• 369 predicaments, one sentence each: median 20 words, range 15 to 24. They follow the redescribe.py rules, so there are no names, titles, places or story-only objects. The topic_words leak check flags 67 of 369. Most of those are role words (visitor 18, rival 13, host 7, guest 5, husband 4), kept on purpose because the design's cast is roles. The real leaks are place words: "house" 17, "room" 5, and "street", "boat" and "church" once each. Poe's situations are often about a house, so "house" usually belongs to the predicament rather than naming a setting.
• 24 groups, each a situation that recurs in two or more stories. Each scene sits in at most one group: 253 are assigned and 116 fit nothing that recurs. The unassigned scenes are concentrated in a few tales, and that is a fact about Poe, not a gap in the reading. Landor's Cottage (17 of 17 unassigned) and most of The Domain of Arnheim (15 of 22) are landscape description with no predicament in them. So are the metaphysics of Mesmeric Revelation (8 of 14), the Sinbad voyage inside Scheherazade, and the essay openings of Von Kempelen, The Premature Burial and The Imp of the Perverse. They were summarised anyway.
• Options: 4 or 5 per group, 113 in all. 97 come from what someone in a scene did (scene:N), 6 are archetypes and 10 fill gaps. 12 scene options cite a scene outside their own group, and groups.json marks each one in_group: false. All 12 are within two scenes of a member in the same story (8 are the scene directly before or after), the same pattern the Wharton reference found.
• Moves: spine paths with a direction only (+, -, risk, set). There are no magnitudes. Skill labels proposed for this corpus: nerve, holding steady under dread, and reason, keeping your head and thinking it through. reason covers the searcher in The Purloined Letter, the fisherman reading the whirlpool, and the prisoner working out the strap under the blade. Poe's heroes survive by that faculty far more often than by courage.
• Stage comes from protagonist standing, judged per story, not per scene, and recorded in STANDING in groups_src.py. The corpus leans high: 9 of 22 tales have a noble or wealthy protagonist, so 14 of 24 groups land in the high stage.

Groups against the existing catalogue

The 36-entry catalogue in ../catalogue/ is built from O. Henry, Wharton and Jacobs. Judged by reading, not by classify.py:

• Overlapping (5): G16 A sound in the night is the catalogue's dread almost exactly. G22 A fortune everyone is watching is after_your_money / strings. G19 Going to the bad is close to habit. G15 At the mercy of someone without pity is a harsher power_over_you. G23 Hidden in plain sight shares ground with cant_prove_it.
• Partly overlapping (4): G17 A place that works on you is a sharper dread: the thing to face is the place itself, and reasoning feeds the fear. G09 A rival who is your mirror is not the courtship rival but an equal who thinks your thoughts. G20 A love you are not free to have touches rival and past_comes_back. G03 Hiding what you did borders on caught_out.
• New (15): G01 The urge you cannot resist (the perverse compulsion), G02 Safe, and about to give yourself away, G04 Shut in the dark, and no one knows (buried or sealed alive), G05 The one you gave up for dead, G06 Death coming slowly, and one cool head, G07 Nobody will believe what you saw, G08 The reminder that will not go away (guilt given a shape), G10 A revenge they cannot escape, G11 Walled off from the danger outside, G12 A companion who may not be sane, G13 Sent for by a friend in a bad way, G14 Experimenting on the dying, G18 Your companion breaks, G21 Someone you love fading into a stranger, G24 A world of your own.

dread is the only existing entry this corpus lands on squarely. The Jacobs dread group means "something may be out there in the dark". Poe's material splits that into several sharper situations. In G01 and G02 the danger is inside you. In G04 and G05 it is someone being buried. In G08 it is your own guilt, and in G12 it is someone else's mind.

How this corpus differs

• The danger is usually internal. The O. Henry, Wharton and Jacobs catalogue is social: rivals, secrets, money, scheming. In Poe the counterpart is often the protagonist's own compulsion (G01, G02), their conscience (G08), or their perception (G12, G16, G17). Actor, wants, counterpart still fits, but "counterpart" is often "the thing they are fixed on" or "the reminder", not a person. A pack built from these needs options that move nerve and reason far more than relationships.*.
• Groups are more concentrated. With 22 tales that mostly run on a single obsession, many groups are one tale plus a few echoes. The top story takes 86% of G23 (The Purloined Letter), 78% of G20, 77% of G09 (William Wilson), 75% of G11 (The Masque of the Red Death) and 71% of G14 (Valdemar). The Wharton rule of at most 50% from one story would cut 16 of the 24 groups. G06 has 19 scenes from only two tales, the whirlpool and the pit, because both are long, slow ordeals of the same shape.
• Some tales hold a group almost alone. G07 Nobody will believe what you saw spans 6 tales, but 9 of its 15 scenes are the queen whose true wonders are dismissed as lies. They are all the same situation, so they are kept together, not trimmed to a quota.
• Several tales are not stories. Mesmeric Revelation and Von Kempelen are mostly essay, and Arnheim and Landor are mostly description. They inflate the scene count without adding predicaments. A tool should expect a "no predicament here" answer for whole tales, not just for stray scenes.

Known defects

• The tail rule drops five endings. The chunking rule discards a story's final paragraph run when it is under 80 words. Five tales lose their endings that way, all of them load-bearing: the rescue in The Pit and the Pendulum (66 words), the dying double's last words in William Wilson (66), the host found dead in The Assignation (61), the ghostly absolution in Eleonora (41) and the closing lines of The Imp of the Perverse (20). The rule is kept as specified so the probes stay comparable, but for this corpus it cuts the resolution of tales that end on a short, sharp paragraph. The summaries describe only what is in each scene.
• Four scenes are epigraphs or subtitles only: 131, 196, 221 and 287, at 6 to 46 words. Each is an epigraph or subtitle followed by a paragraph long enough to start a new scene. They are kept so that chunks.json stays the corpus as split, and none is in a group.
• Poe's footnote markers ((*1) and so on) are left in the text. The notes themselves are in the dropped end matter.