Delve logs

Sorted worst-reading first, because the interesting log is the one that failed. The bar is the share of the poetics a delve met — a sorting key, not a verdict.

50% of delves miss this — Rooms should be in the order they are for a reason.

delve 1: 1 of 9 rooms answer an earlier one (11%) · delve 2: 1 of 13 rooms answer an earlier one (8%) · delve 7: 1 of 13 rooms answer an earlier one (8%) · delve 9: nothing any room did reaches any later room

Mechanics that could serve it: encounters/RoomEncounters.js (forcedFormation, starBlessed, forewarned), agents/Adventurer.js wounds

28% of delves miss this — No sentence should appear twice in one delve, verbatim.

delve 2: 2 effect lines printed verbatim more than once (worst ×3) · delve 6: 1 effect line printed verbatim more than once (worst ×2) · delve 8: 1 effect line printed verbatim more than once (worst ×2) · delve 9: 2 effect lines printed verbatim more than once (worst ×2)

Mechanics that could serve it: sim/Simulator.js pickLine, narrative/Narrator.js pick

24% of delves miss this — The back half should cost more than the front half.

delve 2: front half 3, back half 1 · delve 3: front half 5, back half 3 · delve 10: front half 5, back half 3 · delve 13: front half 5, back half 3

Mechanics that could serve it: world/DungeonGen.js floorScale, game/Progression.js

5181 lines of prose, 2854 of them distinct (45% reused).

A reader reads one delve closely and a saga loosely, so the sentence they are sure they have seen before is the tell. Most reused:

1. Who was this delve about, and does the transcript know?
2. Where did it turn, and was the turn a decision or a die?
3. What did the party pay, and for what?
4. What is known at the boss door that was not known at the entrance?
5. Is there one sentence a reader would repeat to somebody else?

Different numbers are not a different story.

Two delves can differ in every value the ledger records and read as the same delve, because shape is what a reader remembers. Rolling wider does not fix this; it makes the sameness better disguised.

Would change its mind: Readings across many logs showing high variance in shape (which rooms cost, where the turn falls) even where outcomes are similar.

A log that needs its ledger to be understood has failed as prose.

The ledger exists so that nothing can move silently, which is an auditing goal. If the reader has to open it to know what happened, the curation above it is not doing its job.

Would change its mind: Readers reporting they prefer the ledger, or prose-only readings that consistently answer the INTERESTS below without it.

A mechanic the reader never meets is not in the game.

Already house doctrine for coverage (CLAUDE.md rule 12); it is a narrative rule as much as an auditing one. The stairhead camp was built, tested, and met by 2% of delves.

Would change its mind: Nothing. This one is settled; it is here because it is easy to forget.

Repetition between delves costs more than repetition inside one.

A reader reads one delve closely and a saga loosely. The second delve is where a generated game either becomes a campaign or becomes obviously a machine, and the tell is always a sentence they are sure they have read before.

Would change its mind: Cross-delve line-repetition measures staying low while readers still report the saga feeling samey — which would mean the tell is structural rather than lexical.

A loss should be traceable to a decision, not to a die.

The party deliberates, chooses, and pays. When the payment cannot be traced back to the choice, the transcript reads as weather. This is the argument for consequences that reach forward: they make the earlier decision visible at the moment it is paid for.

Would change its mind: Logs where random loss reads as tragedy rather than noise — plausible if the writing frames it as fate rather than as mechanism.

The protagonist should be discovered, never assigned.

If the systems are too even, no one emerges and the delve is about nobody; if a character is designated in advance, the others are luggage. What is wanted is uneven outcomes that a reader can narrate afterwards.

Would change its mind: Readings where an assigned viewpoint character measurably improves the other values without flattening the rest of the party.

The middle of a delve is where generated narrative dies.

Openings get novelty free and endings get the boss free. Rooms four through eight have to earn attention with nothing but what the systems put there, so they are the honest test of the mechanics.

Would change its mind: Per-room engagement measures showing the drop-off elsewhere.