Honest answer: no. There's no place in the repo where the phenomenon is actually explained, even privately.

What exists:

• notes/canon/world/silent-soil.md — 36KB, but it's all observations (Established sections per arc) plus a long "Open questions (do not resolve casually)" list. It explicitly instructs canon to stay agnostic between the two in-world readings, and green-revolution.md and dungeons.md both carry keep-agnostic notes.
  redacted sentence Nothing anywhere else uses the word.
• There's no author's-eyes-only file — no "truth" note, no sealed solution in maybe-later.md, deferred-beats.md, or theme.md.

So what you have instead of an explanation is a set of hard constraints an eventual explanation must satisfy — the Weave's four distinct years of beginning, the north-east march of centres, rings-not-circles, sites arriving one at a time, differing rates, the redacted's non-level top surface, "nine or nine hundred." That's a decent net; it's tight enough that a lot of lazy answers are already ruled out. But it's a shape of a hole, not a thing in the hole.

That's fine for Book 1 — mystery-first drafting is the right call and the two-column ruling is doing real work. redacted sentence

Worth deciding, at some point, whether you want a sealed private answer or whether you're genuinely composing it forward.

Agreed. Let me start with the CLAUDE.md rules, then build the workflow.

The workflow is running (task w2h8r032p). Shape:

1. Dossier — 5 extractors: chapters 1–10, 11–21, 22–31, all canon/world-rules files, and the thematic/series-plan constraints.
2. Propose — 5 independent designers, each forced to start from a different angle (ecology, System/Weave, history, physical ground, dungeons), each producing a full mechanism with a point-by-point evidence table and explicit discrepancy list.
3. Judge — each proposal scored on evidence fit, rule fit, reveal satisfaction, plot generation for Books 2–5, originality; judges are told to verify the evidence tables rather than trust them.
4. Synthesize — one agent builds from the winner, grafts what the judges flagged as worth keeping, re-derives the evidence table from scratch, and writes notes/sealed/silent-soil.md (answer, mechanism, site-by-site, why magic reads it this way, reveal schedule, discrepancies with minimal rewrites, clue budget going forward, rejected alternatives).
5. Verify — three adversarial attackers (evidence lens against the actual chapter text, world-rules lens, sharp-reader lens), then one revision pass.

Every agent is under a secrecy rule: substantive content goes to files only, and return values are restricted by schema to paths, counts, and scores — so nothing that reaches me (and therefore you) carries the answer.

CLAUDE.md now has a "Sealed answers and load-bearing mysteries" section: answers must be written down when a mystery is introduced, they live in notes/sealed/, and I don't relay them unless you very specifically ask for spoilers. I'll commit that with the sealed file once the workflow finishes and I've done a spoiler-free sanity check on its shape (discrepancy count, verdicts).

You've reached your Fable 5 limit. Run /usage-credits to continue or switch models with /model.

Now adding explicit model overrides to the workflow script and restarting it:

Killed the Fable run, added the rule, restarted both workflows on opus/sonnet. Details:

Fable fix. Sorry — that's twice. The mistake was relying on model inheritance inside the Workflow script; with the main session on Fable, inheritance meant Fable. Every agent() call in both scripts now carries an explicit model. Added to project CLAUDE.md as a "Subagent models" section (never Fable for subagents; explicit model on every spawn; treat an omitted one as a bug), and rewrote the existing memory file to say the same rather than the softer "Fable only for the main session" it had.

Silent soil workflow — restarted as wfm9sbu3a. Sonnet for the five dossier extractors, opus for the five proposal designers, five judges, the synthesizer, three adversarial attackers, and the revision pass.

Character bios — running as wunwu302m. Sonnet scopes the cast and assigns word budgets weighted by on-page presence (mention counts: Edwin 512, Peri 348, redacted 252, Mattie 106, Sorrel 77, redacted 40, Ferrin 20), sonnet gathers 15–30 verbatim quoted lines per character, opus writes each bio — opening characterisation, two paragraphs of grounding, a voice section, then "In their own words" as the largest part — plus a one-sentence-and-one-line closing section for the minors. A final opus pass assembles print-CSS HTML under a hard 10-page A4 budget, page-breaking between majors. I'll render it to PDF with headless Chrome (pandoc's here too, but Chrome gives better control over the quote typography) and send you the file.

Neither workflow can leak the mystery: the bios agents are told not to speculate about it, and the silent-soil agents are under a secrecy rule with schemas that only let them return paths, counts, and scores.

All 88 quotes check out as verbatim with correct chapter cites — sending the PDF now.

Character reference PDF — 10 A4 pages exactly, 4,115 words of body text.

Allocation, weighted by on-page presence: Edwin 950w / 13 quoted lines, Peri 700w / 12, redacted 520w / 10, Mattie 320w / 12, Sorrel 260w / 8, redacted 180w / 7, Ferrin 150w / 5, then "The rest of the cast" — 21 minor characters at a sentence and a line each (Tess, redacted, Burl, redacted, Mrs. redacted among them). Each major starts on a fresh page; each entry runs opening characterisation → two paragraphs of grounding → a voice section on rhythm and register → "In their own words" as the bulk of the entry → a short "wrong about / watch for".

I verified all 88 quotes mechanically against chapters/*.md: every one is verbatim and every chapter cite is correct. Six initially failed my matcher, all for benign reasons — quotes that elide an interior dialogue tag ("I'll carry a mallet," said redacted, "if somebody else carries the auger."` printed as one line) or elide with an ellipsis. Nothing invented.

Source HTML is alongside the PDF in the scratchpad if you want typography changes — it's Chrome-rendered from print CSS, so leading, measure, or quote spacing are one edit and a re-render away. I didn't commit either file, since it's a working document rather than story content; say the word if you'd like it in the repo.

Also committed as 04c6c8b: the two CLAUDE.md sections (sealed answers for load-bearing mysteries; never run subagents on Fable).

The silent-soil workflow (`wfm9sbu3a) is still running — it's the longer of the two, with five opus proposals and five judges before synthesis.

You're right to push on this. Let me pull their actual lines side by side before I theorise.

Diagnosis first, then let me check it with numbers rather than vibes.

You're right, and the numbers are worse than "too much like" — redacted is out-Edwining Edwin on every marker of his voice:

What's actually shared

Four constructions do the blending:

1. The self-justifying imperative. Short command + because-clause aimed at an adversarial future reader. redacted redacted sentence: unpublished prose redacted can take apart in one reading."* Same sentence, same ethic, same tactical grounding of honesty. Cover the tags and you cannot tell.

2. redacted sentence: unpublished prose That is Edwin's virtue — the record as conscience — handed wholesale to redacted.

3. Numbers as the payload of the line. Both open on a bare figure and unpack: "Eight," said redacted. "Four each." / "Four sites," said Edwin. "Three soils."

4. Closed utterances. This is the abruptness. Nearly every line either character speaks is a complete, pre-defended ruling. Neither ever picks up the other's word, leaves a sentence unfinished, or asks a question they don't already have the answer to. Two people who each finish their own argument produce adjacent monologues, not an exchange — so the scene reads as a series of hard landings regardless of sentence length. The problem isn't clipped prose; it's no uptake.

The fix, and it's already in the book

029:111 is the seed and everything else contradicts it:

redacted sentence: unpublished prose

redacted's inherited ethic is oral, public, witnessed. Edwin's is written, measured, defensible on paper by a stranger. That's a real and generative difference — and outside that one speech she's the book-and-record character, which is his. Four rules follow:

• He records; she says it in front of people. Move redacted's accountability from the ledger to the witness. redacted sentence: unpublished prose Then their disagreement about which one counts becomes live material instead of a shared temperament.
• Authority by transmission vs. redacted sentence: unpublished prose Edwin fundamentally isn't — his whole project is measuring the received wisdom. Currently she gets this register twice (018:137, 019:177) and it's her best material. Make it her default reach where he reaches for a figure he took himself.
• Split the numbers. redacted counts people, days, miles, stitches — things you get by walking and headcount. Edwin measures rates, yards per day, assay figures — things you get from an instrument and a column. Hard rule: she never quotes an assay figure, he never rosters.
• Strip her because clauses. Her characteristic unit should be two flat sentences whose relation is left for the listener to make — "It's a hand. That's all." / "West." / "It could be me." Those are her three best lines in the book and they're all uninflected. Give the subordinate clause to Edwin; give her the full stop. This alone would pull the 26.4 toward 15 and stop her drowning her own flatness.

And one scene-level rule for the abruptness: in any Edwin/redacted exchange, at least one party must fail to complete their case — a repeated word, a real question, a sentence that stops. Right now neither ever does.

I haven't touched any chapters. If you want, I can do a pass over the heaviest scenes — 014, 016, 018, 019 hold most of the collision — reassigning the ledger lines and de-because-ing her. Say which of the four rules you'll actually take first, since they're separable and the second one (transmission vs. measurement) is the one that changes plot texture, not just voice.

The sealed-answer workflow finished. Verifying spoiler-free — I'm checking structure and counts only, not reading content.

Already committed — the synthesis and revision agents committed their own work (3a9e465, then 36b2ea7).

The sealed answer exists: notes/sealed/silent-soil.md. Shape only, no content:

• 33,865 words, 62 sections, all template sections present, spoiler header in place.
• Evidence coverage: 194 items from the canon file explained, 33 neutral, 0 contradicted. The evidence table is 304 rows — every catalogued observation from chapters 001–031 checked individually against the mechanism.
• Discrepancies requiring rewrite: 15. These are places where what's on the page needs changing to fit. That's your big-rewrite list, and it's itemised in the file with the minimal change for each.
• Open questions resolved: 26 — including several the canon file had flagged as unresolved.
• Judging: four angles competed rather than five (the ecology proposer died on a server error mid-response and I let the run continue). Scores were 43 / 42 / 41 / 29 — the top three were close, which is a good sign; the synthesis grafted from the runners-up rather than taking the winner whole.
• Adversarial pass: three independent attackers, all three returned revise — 13 fatal, 30 serious, 22 minor flaws between them. The revision agent then rewrote against all of it, which is where the discrepancy count went 7 → 15 (it stopped papering over conflicts it couldn't actually satisfy) and coverage went 189 → 194.

One honest caveat: the revised version was never re-attacked. The three critiques hit the pre-revision draft. I'd recommend a second adversarial round against the final text before you build Books 2–5 on it — cheap relative to what's already spent, and the first round found real fatal problems, so I don't want to assume the second wouldn't.

Two follow-ups I'd suggest but haven't done: a pointer (pointer only, no content) from notes/canon/world/silent-soil.md to the sealed file, and adding notes/sealed/ to the read-list in the consistency-check and write-arc skills so future drafts are checked against it.

Nothing about redacted has been changed — that's still waiting on which of the four voice rules you want.

Design: each scene gets written twice-over across two rounds, and every scene is blind-tested — the writer also emits a tag-stripped version, and an independent judge attributes each line before scoring. That turns "do the voices differ?" into a measurable number instead of an opinion.

Running as wsu7v768i. No chapters will be touched — it writes only to a scratchpad, notes/canon/characters/, notes/dialogue-voices.md, and samples/dialogue/.

Structure — 23 agents, all opus except a sonnet brief-builder:

1. Brief — a voice-differentiation dossier seeded with the measurements above, the four blending constructions, and the four proposed separation axes, with instructions to test them against the text rather than accept them.
2. Round 1 — five scenes, each written from scratch and then blind-judged, pipelined so each scene flows independently.
3. Distil 1 — updates each character file with a ## Voice — binding (Claude) (Working) section: characteristic sentence shape, what they reach for as warrant, which numbers are theirs, what they do at an emotional peak and what they do instead when avoiding it, three constructions that are theirs alone, and three that belong to someone else and are banned for them.
4. Round 2 — the same five scenes rewritten from a blank page under the new rules. Not patched — re-drafted, so the rules have to actually carry the writing.
5. Distil 2 — revises the rules against what round 2 proved, writes notes/dialogue-voices.md, and copies the best scene per pairing into samples/dialogue/ as exemplars.

The five scenes, chosen to stress the failure and force feeling:

• Edwin & redacted, after redacted night — redacted alone with the mistake she's already confessed in public, Edwin past being clever. Fear and fondness, no tidy resolution.
• Edwin & redacted on whether writing it down counts — his conscience is the checkable document, hers is the thing said aloud in front of the household. Neither wins. This is the axis, and the brief demands it be audible in syntax, not opinion.
• redacted & Peri, the nudges — Peri asks for something redacted won't give and gets a smaller true thing instead.
• Edwin & Peri, redacted — she's angry in a way she can't justify; the feeling arrives through their hands.
• redacted & Mattie in a kitchen — the control case. If redacted only sounds distinct when Edwin is in the room, her voice isn't real yet.

Every writer must also emit a tag-stripped, numbered version with the speaker key sealed at the bottom; the judge attributes every line before reading the key, scores itself, and is told plainly that 50% means no differentiation. Scores tracked across rounds: blind attribution, distinctness, emotional truth, caring, uptake, LLM-tells, Voice D fidelity.

Two things I'll flag rather than assume: the judges are graded on self-honesty but nothing structurally prevents an inflated attribution score, so I'll spot-check one blind file against its key myself when it lands. And if round 2's attribution doesn't move meaningfully, the right response is a third round with different pairings, not accepting the notes as written — I'll tell you which it is.

The final agent reports fixed: false and argues its own scores are near-worthless. That's the claim to check first, since everything else depends on it.

Verified independently, and the agent's self-criticism is correct: the round-2 blind files run 81–96% strict speaker alternation. One anchored line recovers the whole scene by counting. So 47/48 and 34/34 measure alternation, not voice. The honest number is the agent's own out-of-order figure — roughly a fifth to a third of lines attributable on grammar alone.

Since the voices aren't right yet, I've launched round 3 with the test repaired and pointed at the holes rather than repeating the same exercise.

Rounds 1–2 are committed (d8729c5, adc9808): binding voice sections in all seven character files, a new notes/dialogue-voices.md, and four verbatim exemplars in samples/dialogue/ — each carrying a "Known defects — do not imitate these" list, on the reasoning that cleaning up an exemplar destroys the evidence and teaches its own faults. No chapters touched.

What genuinely got fixed:

• Uptake — solved. Six named mechanisms, each with a working line: grant-the-fact-reverse-its-sign; a single word returned as a bare interrogative and answered by narrowing (*"Behind." / "Behind the post. On redacted redacted sentence: unpublished prose This was your root cause of the abruptness and it's gone.
• Caring. *"I'm angry. redacted sentence: unpublished prose
• redacted's sentence.** Her long speeches now stop where the list stops. She went from the character the exercise broke on to the most attributable in the cast.

What isn't fixed — and the headline is that the measurement was bad. Round 2's blind files ran 81–96% strict alternation; I verified this myself. A judge anchors one line and counts. So 47/48 meant nothing. The honest number, out of order, is a fifth to a third.

Beyond that: short mode is the biggest hole and nobody planned for it — every round-2 scene had a stretch where both speakers went short and the voices vanished (17 of 35 lines in one). The character files are written almost entirely about long speeches, and most dialogue isn't long speeches. Peri is separated by who knows agronomy, not by grammar. Mattie is measured at 64 number-tokens/1000 against redacted's 58 — the round-1 gap closed only because redacted's density doubled. Sorrel, redacted, Ferrin untested. And the narration bought differentiation by going **inert — two of five scenes had no Voice D in them at all.

Round 3 is running as wyom9snbb (11 agents): shuffled-line attribution with adjacency destroyed and short lines scored separately, a whole scene constrained to short mode, an alternation ceiling, an explicit narration-liveness score, and the four untested characters put in scenes without Edwin or redacted to lean on. The distiller is instructed to delete disproved rules rather than accumulate them, and to tell you plainly if iteration has stopped paying.

Two things for you, not for the workflow: the Mattie/Ferrin three-fragment-report collision is a real unresolved clash that I've told it to evidence and recommend on but not rule. And it kept edwin-redacted-mistake(24/25) as the exemplar overrecord (34/34), because record has a canon breach at its peak — Edwin given Peri's dash false start. Overrule that if you'd rather have the higher-scoring one; it's still in the scratchpad, which is ephemeral.

Round 3 committed (09df726). And I checked its most important claim against the published book, not just the exercise:

Peri drifts in the actual chapters too. Her dialogue near Edwin averages 15.1 words with an and-rate of 27.7/1000; near redacted redacted sentence: unpublished prose She becomes whoever she's standing next to. That's independent corroboration that her problem isn't the exercise — it's on the page.

The honest number moved from ~a fifth to ~a third across three rounds (that's lines placeable with props, proper nouns and subject expertise covered — the raw shuffled scores are still flattered). One decisive result: the fifteen-word cap. The all-short Edwin/redacted scene is the only one that beat the band, and it beat it decisively — because removing the long speech removes where every discriminator was written and where drafters hide.

Two findings worth your attention:

A rule can be obeyed and defeated. Peri's countable false-start rule was met exactly — two em-dashes, both hers, none elsewhere — by a scene that replaced redacted with an announcement of redacted ("I'll say it plain"). redacted sentence: unpublished prose Both loopholes are closed, but any rule stated as a count invites satisfaction without effect.

Descriptive rules fail; countable prop-free grammatical markers work. Every one found has held — redacted's anchor, Edwin's figure-opener, redacted's place/time/page, Peri's subordination, Mattie's class-dismissal, Sorrel's animate subject. Every rule phrased as function or intent has failed under test: "address", the fragment rule, "he goes plainer at a peak" until it was made mechanical.

Fixed and measured: Mattie's number leak is gone (61.5 → 17.8 against Peri). redacted passed outright — the only clean short-mode result. Edwin's peak held for the first time. Narration is no longer inert, though it now front-loads Voice D and coasts through the middle.

Peri has not moved in three rounds — 11 of 62 in the scene built around her. The mechanism is now visible and it isn't syntax: every rule she has is about how she records things, so she's either the one with the pen or she's nobody. Her voice is a job description.

I've launched round 4 as the distiller recommended, but only two of its three scenes:

Running as wtbg7nmd6 (5 agents). The two scenes:

1. redacted sentence: unpublished prose Opposite a farmer who knows less than she does, so for the first time she holds the expertise rather than the instrument. Hard fifteen-word cap on every turn. This is the only untried configuration and it's the whole of her problem.
2. redacted sentence: unpublished prose redacted thread. Tests the real candidate discriminator (her clauses take animate subjects; Edwin's take the matter) after round 3 disproved the stated one.

The instrument is rebuilt: shuffled and prop-covered, so props, proper nouns and subject vocabulary are replaced with placeholders before the judge sees a line — only grammar can place it. Plus the swap test round 1 specified and no round has run: three lines from each speaker put in the other's mouth, scored on how many a reader rejects as impossible. A voice that survives no swaps doesn't exist. The raw shuffled score is gone from the report so nobody can quote it.

redacted sentence: unpublished prose redacted sentence: unpublished prose Worth knowing before you rule: Ferrin currently places on vocabulary, not grammar — cut the word "fact" out of his scene and four of his seven passports fall over. Resolving this casually would take a construction off one of them permanently, so I've left it. No scene with both of them should be drafted until you've called it.

I'd also cap the exercise here regardless of round 4's result. Three rounds moved the honest figure from a fifth to a third, and the returns are now concentrated in one narrow move — find a countable prop-free grammatical marker, write it mechanically. That's a method you now have; the remaining work on Peri especially may be an authorial decision about who she is rather than something more iteration can find.

Round 4 committed as 647cc49. I've stopped the exercise there — the distiller's own recommendation, and I agree: round 4 produced two rules and two retractions, and the retractions were worth more, which is what a paying round looks like near its end. Four authorial decisions are now queued, and none of them can be resolved by shuffling more lines.

I spot-checked its central measurement against the published chapters. Sorrel opens on a first-person pronoun in 29% of her utterances against Edwin's 15% — the direction holds, though on 17 utterances her base is thin enough that I'd treat the rate as indicative, not settled.

Solved: Edwin, redacted, redacted. Each has at least one countable, prop-free marker that survives a swap. redacted sentence: unpublished prose redacted's is the anchor under six words, with one named hole — she has no short question form, so any short question in her mouth is voice-neutral by construction. redacted passed outright.

The method is the biggest thing four rounds bought. Prop-cover plus the swap test, and where they disagree the swap test is right: round 4's scenes scored 86% and 100% on attribution and 3/6 and 2/6 on swaps. redacted sentence: unpublished prose Use these on real chapters in the editing pass.

redacted sentence: unpublished prose That plus the dash false start is the whole of it; nineteen of her twenty turns could have been Edwin's. Both declared separators died: her "professional modal" is Edwin's own canon line from ch. 019 with the nouns changed, and subordination is licensed to him outright.

Four decisions for you:

1. The Mattie/Ferrin collision — evidence complete, recommendation on the page, ruling is a judgement. Still untouched.
2. redacted sentence: unpublished prose Either that's the character — a girl who became a clerk because it was the one job in the room she could do and has nothing under it yet, which is a real thing to be at eighteen and would need to cost her something on the page at least once — or something has to be written underneath her.
3. What Sorrel wants. Load-bearing for succession, still unwritten. redacted sentence: unpublished prose Round 4 put the ask correctly and then granted it two turns later, so nothing cost anything.
4. Whether Peri's twelve-word cap binds retroactively — the chapters have her at 15.1 words beside Edwin, which is most of her page-time. Three options are written into the rule; don't let it get resolved in passing in a draft.

Two smaller things: samples/dialogue/edwin-sorrel-trench.md was added because it beat round 3 on the stated test, but its Sorrel is off-canon and the risk is a drafter imitating her — its findings are recorded independently, so pull it if you'd rather. And one test is still worth running before anything in peri.md gets promoted: Edwin and Peri, fifteen-word cap, prop-covered. I've left that for you to call rather than launching a fifth thing unasked.

No chapters were touched at any point.

That observation is testable, so let me check it before answering.

You're right, and the shape of it is sharper than "stronger earlier." Her mean line length across the three arcs:

She doesn't lose page-time — 41% → 32% → 40%, she's as present at the end as the start. She loses sentence. Her turn shrinks by two-thirds, and her question rate drops with it. redacted sentence: unpublished prose redacted sentence: unpublished prose redacted's polysyndeton on the way down.

That reframes the decisions, including one the workflow got backwards.

1. The twelve-word cap is the wrong instinct and you should probably kill it. The exercise proposed capping Peri at twelve words because the fifteen-word cap improved distinctness under test. redacted sentence: unpublished prose redacted sentence: unpublished prose Her strongest, most distinctive period is the 21-word one. The right move is almost certainly the opposite: give her turn back its length and fix the distinctness inside it.

2. The real decision: what is under the clerk. Four rounds established that every rule she owns describes what she does with information. That's a job description, not a person, and it's why she vanishes when you take the pen away. Two honest options:

• It's the character. A girl who became a clerk because it was the one job in the room she could do, and who has nothing underneath yet. That's a true thing to be at eighteen, and it's compatible with redacted arc — but it only works if it costs her something on the page at least once, and if the series is going to put something under it later. Right now it reads as absence rather than as a fact about her.
• Something goes underneath now. A want, an appetite, a grudge, a form of pleasure that has nothing to do with the work. She's the only major with no non-professional appetite: redacted has her grandmother's people and her dead, Edwin has three hundred and fifty years of accumulated opinion, Mattie has her land. Peri has a notebook.

3. Does the drift get repaired retroactively. She currently averages 15.1 words beside Edwin and 6.8 beside redacted — she becomes whoever she's standing next to, which is the same fault as the arc-by-arc decline seen sideways. redacted sentence redacted sentence

redacted sentence: unpublished prose redacted sentence: unpublished prose redacted rather than earning it. redacted sentence: unpublished prose

I haven't changed anything. Say which way you want to go on what's under her and I'll work the other two out from it.

Agreed on both. Let me pull what her early personality actually consists of, so restoring it means restoring something specific rather than "more personality."

Recorded as f723040 — the cap is struck and the clerk direction is rejected in canon, marked as your ruling rather than a Claude proposal.

Reading her early chapters back, the personality that drains out is seven specific things, all citable:

1. redacted sentence: unpublished prose I'm not sorry about the rest, because the world needs saving, and it's you, and I'd rather have you angry than have the world end. redacted sentence: unpublished prose redacted sentence: unpublished prose
2. redacted sentence: unpublished prose redacted sentence: unpublished prose redacted sentence This is the largest single loss, and it's the one that produces the passivity you're seeing — everything else follows from her wanting nothing on her own account.
3. redacted sentence: unpublished prose
4. She lands things without knowing it. *"The scones were for you. redacted sentence: unpublished prose
5. redacted sentence: unpublished prose redacted sentence: unpublished prose
6. redacted sentence: unpublished prose redacted sentence
7. redacted sentence: unpublished prose

I also demoted every rule the workflow wrote that separates her by what she does with information — filing, columns, recording. redacted sentence: unpublished prose The two prop-free markers survive, since they're hers whatever she's holding.

redacted sentence redacted sentence

Still queued from earlier and untouched: the Mattie/Ferrin collision, and what Sorrel wants.

That decision simplifies both questions, so let me record it first — it changes what's at stake in them.

Recorded as 4a607c1. That also closes the last open item on Peri — "does the repair bind retroactively" is moot if nothing's being repaired.

Who they are

redacted sentence: unpublished prose I measured with the chain.* She can't pay and said so. redacted sentence: unpublished prose

Ferrin — the priest of Fenholt, spare, sixty, a lantern. redacted sentence: unpublished prose redacted sentence: unpublished prose He stood at the gatepost for the entire six-hour field walk, carrying a lantern long after it had gone out.

The Mattie/Ferrin question

They share a construction — the three-fragment report. His: "Firm. Dry for Seedmonth. Grass good." Hers: "redacted and his brother. Stood at F for a while. Neither of them knocked." Below eight words you cannot tell them apart at all; above it, only by what they're talking about. redacted sentence: unpublished prose

redacted sentence: unpublished prose redacted sentence: unpublished prose

The alternative is a clean split: one of them gives the construction up.

There's a caveat worth more than the ruling. redacted sentence: unpublished prose redacted sentence: unpublished prose If you adopt the recommendation, the cap on that lexicon matters as much as the rules.

Given the rewrite decision, this is now cheap either way — there's no accumulated prose paying for a change, so pick the one you find truer rather than the one that costs less.

The Sorrel question

Different in kind: it's a plot hole, not a voice hole. Nobody has written what Sorrel wants for herself. She's been in three test scenes and walked past it every time — the last one had her ask correctly and then Edwin granted it two turns later, so nothing cost anything.

It matters because her file flags it as load-bearing for redacted thread. redacted sentence redacted sentence: unpublished prose

Related and worth knowing: round 4 disproved her stated voice rule too. redacted sentence: unpublished prose redacted sentence: unpublished prose

Let me record the Mattie/Ferrin ruling, then read Sorrel properly before suggesting.

Ruled and committed (bd07e06).

Sorrel

redacted sentence: unpublished prose She broke a hail front herself. She worked the equinoctial turn for eleven hours as a one-mage day, decided to lose a third of the late grazing in the salt-burn parishes, and then wrote: *"I did not send the rider. redacted sentence: unpublished prose redacted sentence: unpublished prose

That is someone defending a decision in advance to a man who has never once overruled her. She doesn't need his approval and she asks for protection from his manner of giving it. redacted sentence: unpublished prose

Three candidate asks, with the constraint that it must be his to refuse and stay refused:

1. To be the one who goes. She asks to come east — to be on the road, at redacted, in the work that matters — and he has to refuse, because she is the only person who can hold the vale. This is my recommendation. The refusal is genuine, it is permanent, and it costs him something to say: she is caged by being good at it. Nobody is at fault and there is no version where she wins. redacted sentence: unpublished prose And it's true of real institutions in a way this book is otherwise good about.

2. A name for what she is, written down. Not a warm acknowledgement — a title, in the record, that makes her decisions hers rather than his by delegation. Weaker as drama: he'd want to grant it, and any refusal has to be manufactured out of guild machinery, which makes it a paperwork problem rather than a human one. Better as the thing she asks for instead, when the real ask has been refused and she needs to leave with something.

3. To be told plainly where the vale goes when he dies. The most load-bearing, and the most dangerous. redacted sentence: unpublished prose His refusal is honest (he doesn't know, or he does and can't say it yet), which makes it worse, not better. redacted sentence

I'd use 1 as the scene and let 3 be the thing underneath it that she doesn't ask — she opens her mouth for redacted question and asks for the road instead, which is the smaller sayable version. That gives you the ask that stays refused, the emotional truth, and the groundwork, without spending redacted confrontation early.

Two things to be careful of. Her ask must not be granted two turns later, which is how every test scene failed. redacted sentence: unpublished prose

Say which and I'll write it into sorrel.md as your ruling, along with the correction that her stated voice rule is false about her.

That's better than all three of mine, and it fixes the craft problem rather than answering it: nothing has to stay refused, because nothing gets asked. Let me check the goats before I build on it.

It lands, and the book has already seeded the mechanism — I think more precisely than you may have had in mind when you reached for it.

Peri asked the Weave where am I most needed and got a goat shed. Her own reading of that (ch. 003):

"Either the goats matter in a way that'll be clear later, or the Weave answers the question you actually ask. I asked about needed. redacted sentence: unpublished prose That's the hard part."

redacted sentence: unpublished prose redacted sentence: unpublished prose redacted sentence: unpublished prose

That also retro-justifies the Peri ruling from earlier. redacted sentence: unpublished prose That's a personality, it's prop-free, it works with no pen in the room, and it's exactly what dragged Edwin out. The clerk arc buried the one thing that made her the engine of the story.

Your structural point is the sharp one, though: the refusal isn't Edwin's. He'd say yes in a heartbeat and both of them know it. redacted sentence: unpublished prose redacted sentence: unpublished prose redacted sentence: unpublished prose

Two things I'd add, and flag as mine rather than yours:

redacted sentence: unpublished prose redacted sentence: unpublished prose That's the version where he has something to actually be sorry about, which the theme notes already want (maintenance as hiding).

redacted sentence That's much more interesting than a competent woman being unfairly overlooked, and it's non-obvious enough that a reader won't get there ahead of her.

redacted sentence: unpublished prose I'd also add the correction that her stated voice rule is false about her while I'm in the file.

Now the two corrections that follow from it — Sorrel's false voice rule, and Peri's actual faculty.

Rule 1a was already corrected by the round-4 distiller. So just rule 8's superseded framing, and Peri's faculty.

Written and committed as 11fd401.

notes/theme.md unpublished — "The trap", binding, as an extension of anchor 2. redacted sentence: unpublished prose redacted sentence: unpublished prose Also flags redacted redacted sentence: unpublished prose

notes/canon/characters/sorrel.md — "What Sorrel wants", the long one. redacted sentence: unpublished prose It keeps your caveat that maybe it is what she wants, and forbids prose from deciding that early. Plus how it shows on the page: she's not aggrieved, not thwarted, not waiting to be recognised, and the danger in her scenes is Edwin being kind at her. redacted sentence: unpublished prose

Rule 8 is now marked answered rather than open, with the note that the reason it stayed unwritten through three test rounds is the character, not a gap in the notes.

redacted sentence: unpublished prose redacted sentence: unpublished prose

redacted sentence: unpublished prose redacted sentence: unpublished prose

redacted sentence: unpublished prose

That clears every decision that was queued. The state of things: the sealed answer exists with fifteen known discrepancies, the voice work is closed with Edwin/redacted/redacted solved and the method (prop-cover plus swap test) available for the rewrite, and chapters 001–031 are now formally groundwork. The obvious next piece of groundwork is the rewrite's plan — but that's a bigger conversation than a follow-on, so I'll leave it with you.

Now the targeted canon/notes edits.

Now the character files, farming rule, and the book plan.

Read all nineteen. Committed as 1bfba56 — no chapter prose touched.

notes/reader-notes-book-1.md is the new home for them, sorted by what each one binds. The useful surprise: most aren't open questions. Four are prose contradicting canon that already existed, so the rewrite has no licence to repeat them:

redacted sentence: unpublished prose redacted sentence: unpublished prose

• "First one since the Choosing" at Level 26. Twenty-five levels unaccounted for. The scene wants both "it's her first" and "it's a big number."
  redacted sentence: unpublished prose redacted sentence: unpublished prose In this world that's just someone old.
  redacted sentence: unpublished prose Wystow existed; its agricultural everything is his monument, and his biography can't contain it.

Two comments turned out to be the same error twice — "nobody writes it down" about emptying a mana pool, and "nobody sings about that." Both reach for oral-culture ignorance in a society with near-universal literacy and circulars. The better line in the first case is that it's recorded and simply not done: the world's knowledge is fine, its patience is the scarce thing.

Peri. The 029 note is the one with the most downstream weight, so I gave it a section. Ch. 001 isn't the defect — "I'm sorry about the sheep, I'm not sorry about the rest" is one of her best lines and exactly the unapologetic version the earlier ruling wants. redacted sentence: unpublished prose redacted sentence: unpublished prose redacted sentence: unpublished prose

Your two ch. redacted sentence: unpublished prose redacted sentence: unpublished prose

The six "reads clever but doesn't parse" lines I made into a named entry in llm-tells.md rather than six fixes: the sentence has a good line's shape without a proposition that survives being restated plainly. redacted sentence: unpublished prose Dangerous here specifically because Voice D's register is dry compression, so merely-compressed is hard to tell from compressed-around-something.

The ending. Diagnosis in book-revision.md: redacted rule is doing exactly what it's told, and that's the problem. It governs chapters, where the reader has the next one. Book endings still can't be cadences, but they have to land a change in standing — and the strongest shape is redacted redacted sentence: unpublished prose The mallet-and-auger exchange is nearly that and misses; it's a party negotiating logistics.

redacted sentence: unpublished prose

The two I'd flag as still needing you: "She was quoting a rulebook — what rulebook?" either becomes a real written instrument binding fate magic, which fits the oath-held-mystery material in healing.md and is probably a good fact, or the line goes. redacted sentence: unpublished prose

Answered all three as Working canon, committed as e0adf59.

Why are there fens, and what made them worth draining? (continent.md) — Ordinary geology: a shallow basin with no fall, slow rivers that can't get away, a silted sea end. redacted sentence: unpublished prose redacted sentence: unpublished prose

The part I think earns its keep is that the fen was never worthless. redacted sentence: unpublished prose That's the same fact as the crown chartering a guild at Fenholt because lordship never paid.

redacted sentence: unpublished prose

redacted sentence: unpublished prose A drainage scheme is never finished. redacted sentence: unpublished prose It also gives redacted redacted sentence: unpublished prose

Why coastal towns? The premise is that the sea isn't uniformly lethal and the towns aren't looking at it. redacted sentence: unpublished prose redacted sentence: unpublished prose redacted sentence: unpublished prose Add fish as protein at scale, salt in bulk to keep a surplus through winter, and salt-marsh grazing with seaweed carted inland as manure. redacted sentence: unpublished prose

Are monsters edible? Mostly yes, and it doesn't matter — that's the joke. redacted sentence: unpublished prose redacted sentence: unpublished prose redacted sentence: unpublished prose redacted sentence: unpublished prose redacted sentence: unpublished prose Not a magic-item economy.

redacted sentence: unpublished prose

Both committed and verified in place. Now the two new rulings.

Yes — all three of yesterday's answers are in continent.md and dungeons.md, committed as e0adf59. Nothing was left in the reply only.

Both new rulings are now in canon (56d03ea).

Enchanting (magic.md). redacted sentence: unpublished prose redacted sentence: unpublished prose The skill unlocks at 1 from the class with a boost, and the hard road exists but almost nobody walks it, which is what makes it almost always second-tier.

Two corollaries I drew that I think are worth having:

redacted sentence: unpublished prose redacted sentence: unpublished prose
redacted sentence: unpublished prose Also means items are institutional and heritable rather than personal; nobody is shopping.

And the bit that keeps it from eating the premise: the green revolution didn't run on items, and that's the point. redacted sentence: unpublished prose redacted sentence: unpublished prose

Skill levels (system.md). This is the more valuable of the two, because it supplies a cause for something canon has always asserted without one. The stall was a rule; now it's a consequence. redacted sentence: unpublished prose redacted sentence: unpublished prose redacted sentence: unpublished prose That's why Edwin's warning is about forty.

redacted sentence: unpublished prose redacted sentence: unpublished prose redacted sentence: unpublished prose

One fork I didn't resolve, flagged in the file for you. "A class also does it" reads two ways: (a) the class moves the number, so her sheet says nine and she didn't earn nine; or (b) the class supplies a boost on top of an unmoved number. (b) preserves the older canon line that classes never grant levels past 1 and avoids soundness and level being two separate hidden quantities. I've set (a) as the working assumption and amended that older line accordingly, because (a) makes the level a lie the character can read on their own sheet — which is the more useful lie. But it's a real change and it's yours to make.

Also flagged: whether enchantments need maintaining. My working answer is yes, at rates from "recast at the quarter day" to "twice a century" — it's the theme's own answer and it turns an heirloom into a standing obligation rather than a free asset — but that's a genuine design commitment rather than a derivation, so I marked it for you.

Both ruled and recorded (211d38e), with the two older lines they contradict amended in place so consistency-check won't have a live contradiction to trip on.

The class moves the level. Your framing turned out to be the sharper version of what I'd written, so I made it the operative text: free information, a wider net for nudges, and no procedural skill at all. What I did with it is turn the pair into a test to write from —

redacted sentence: unpublished prose redacted sentence: unpublished prose

That's the bit I think earns its place, because it explains why the failure is invisible from inside, not just from outside. The work passes. redacted sentence: unpublished prose Which is exactly why the warning is about forty.

redacted sentence: unpublished prose redacted sentence: unpublished prose Otherwise this becomes a stat check and stops being a character observation.

Enchantment durability keyed to the object is the tidier of your two rulings, and it does more work than it looks like. redacted sentence: unpublished prose redacted sentence: unpublished prose

Three things fall out that I recorded:

redacted sentence: unpublished prose redacted sentence: unpublished prose redacted sentence: unpublished prose

• It sets what gets enchanted: durable things. Stone, iron, a millwheel, a boundary. redacted sentence: unpublished prose
• Failure has an undramatic shape. redacted sentence: unpublished prose

No open flags left on either of these.

Written to notes/canon/world/demographics.md, cross-linked from system.md and magic.md (44b87f9). Your numbers check out against what was already there: 1 in 100 of the living against magic.md's 1-in-10,000 field-scale rung makes the top rung one mage in a hundred, which is a sane pyramid.

redacted sentence: unpublished prose The fix is the ordinary one: professionals concentrate. redacted sentence: unpublished prose So a parish has one or none, often old and part-time, and produces a mage-child once a generation or two. The 1-in-100 is real, it just lives in the towns.

redacted sentence: unpublished prose redacted sentence: unpublished prose redacted sentence: unpublished prose Working split, adopted unless you overrule: one propertied child in ten, one labouring child in four hundred, averaging to your 1 in 200.

Two payoffs fell out that I don't think were in your prompt.

redacted sentence: unpublished prose That's the concrete social content of redacted redacted sentence: unpublished prose Noted that Edwin must never say this; he's standing in it.

redacted sentence: unpublished prose

I also flagged one inference as Working and genuinely optional. Longevity is bought by continuing to climb, so a mage who plateaus ages near-normally from then on — which means a magic class isn't a life sentence of centuries, and, at one remove, the stall is eventually lethal. redacted sentence: unpublished prose Say if that's more than you wanted and I'll pull it.

Now the two demographics sections that this changes:

Both written (c905619), and the old "possibly stat-linked; details open" line in system.md is now closed rather than left dangling.

The aging mechanism. The thing your two attributes buy is a clean visual distinction, which is what I leaned on: they produce two kinds of old person who don't look alike.

• Vitality refuses to break. redacted sentence: unpublished prose The parish reads them as tough.
• Magic slows the clock. Such a person isn't hard to kill; they're just not old yet. Edwin at three hundred being physically a healthy sixty is this and not toughness.

redacted sentence: unpublished prose redacted sentence: unpublished prose

redacted sentence: unpublished prose That's what makes 500–600 a record rather than an open-ended number, so the ceiling now has a reason.

redacted sentence: unpublished prose That gives it a good restraint — it isn't cosmetic and can't be aimed, which is why very old mages should read as settled rather than youthful. redacted sentence: unpublished prose Noted as open with a do-not-spend-casually.

On the country distribution — your correction is better than what I had and I've replaced it rather than hedged. redacted sentence: unpublished prose redacted sentence: unpublished prose

redacted sentence: unpublished prose

Now a pointer from the farming policy file, since that's on the read-before-writing list:

Ruled and enforced, and you were right that we were nearly clean already.

The audit. I swept chapters, canon, plans and summaries for New World foods, animals and materials. Exactly one violation, in a plan, not on the page: the frost table in plans/farm-curriculum/a-foundational-five.md unpublished listed "potatoes, beans and garden crops" as killed outright by frost. Dropped the potatoes; the fact survives intact on beans.

Everything else checked out, including two things that looked like hits and aren't:

redacted sentence: unpublished prose The corn factor, a corn stack, "when the corn comes off", the archaeology note about Roman towns coming up in the wheat. Consistent, and correct as it stands.

• Bean straw (ch. 022, the redacted stack) is field beans, Old World and exactly right for the rotation.

Tea, cane sugar and pepper are all Old World imports and stay — and development.md already has your ch. 001 "three sugars" ruling on refined sugar as a noticed small luxury, which is untouched by this.

Written into continent.md as "Flora, fauna and the larder": the rule, the banned list (foods plus turkey, rubber, quinine, guinea pig, llama and the American ornamentals), and the Old World list that remains — which is worth having spelled out, because it's long enough that nobody drafting should feel constrained.

redacted sentence: unpublished prose redacted sentence: unpublished prose

I also put a short rule 1b into notes/farming.md, since that file is on the read-before-writing list and this binds every author's note as much as it binds the prose. The one standing risk is corn drifting toward maize in a drafter's ear, so that's called out explicitly with "do not clarify it".

Fixed — the rule now binds the world, not the page.

redacted sentence: unpublished prose What's banned is any of it turning up in the story — narration, dialogue, or a character's larder, redacted or field.

That also means my earlier "fix" was wrong and I've reverted it. The frost row in the curriculum plan is the real-world mechanism column — the fact the spell is built on, headed for an author's note — so potatoes belong there exactly as they were. Restored.

I dropped the instruction to substitute beans for potatoes when illustrating a real fact, since that was the same mistake wearing a different hat: it was constraining the notes, which are the one place the constraint doesn't apply.

Now I have the full picture. Writing it into canon.

Now timeline and system.

Written into canon (4c5c117) across development.md, edwin.md, timeline.md and system.md.

The reorder fixes a seam that was already there. Under the old version he had two unconnected failures — a shed at 18, then quitting four terms short for no stated reason. Now they're one event: he was three years into Wystow, cast the one battle spell his training had given him, hit a shed, and never went back. "One spell, one shed" is now the complete academic record as well as the military one.

What I added around your ruling:

• The Choosing. Canon already has classes arriving at a choosing around 18. redacted sentence: unpublished prose redacted sentence: unpublished prose
  redacted sentence: unpublished prose That's why a scholar of magic and letters had a battle spell to cast at all.
  redacted sentence: unpublished prose The redacted charms are half-working in front of a man equipped to see what they're reaching for.

Two things that got better rather than needing repair. Ch. redacted sentence: unpublished prose And redacted redacted sentence: unpublished prose

Three prose lines in accepted chapters now contradict canon. I flagged them rather than editing accepted chapters unasked:

redacted sentence: unpublished prose redacted sentence: unpublished prose
redacted sentence: unpublished prose redacted sentence: unpublished prose

• chapters/020-*.md:147 — age only, one word in his own mouth.

redacted sentence: unpublished prose Say if you'd rather it went another way — and say if you want me to make the three prose fixes.

Already recorded in notes/process.md — the drafted chapters are input to the rewrite, not text to be maintained. I'll retitle the flags so they don't read as owed edits.

Done. The three lines are now listed as beats for the rewrite to get right rather than repairs owed to the drafts, so nobody chases them and nobody cites the superseded version.

The origin canon stands as groundwork: Choosing at 18 → three years at Wystow → the shed at 21 → the fall back onto farming that becomes the whole career.

Written into notes/canon/characters/peri.md (commits 9301a8a, plus a cross-reference), replacing the old "needs rebuilding" section — which is retained below the ruling.

The ruling, as canon: the title stays and is Peri's verdict, not the book's. Her two questions — who is the best person to teach me magic and who can show me how to save the world — both truthfully answer Edwin, and neither on a technicality:

• Teacher: most students taught by a mile, and he gets magic into people with no aptitude at kitchen tables. Nobody alive is better at it.
  redacted sentence: unpublished prose redacted sentence: unpublished prose That there's no duel in the method is the education of the book.

redacted sentence: unpublished prose It's the interval: days spent asking human beings about a name. redacted sentence: unpublished prose redacted sentence: unpublished prose The Weave only gave her a name.

redacted sentence: unpublished prose

The boots (marked Claude/Working, overrule if you want it otherwise): she didn't walk most of it. redacted sentence: unpublished prose Appleford is off the coach roads, so the walking and the misery are the last day or two.

All three corrections are in (335b171 and a follow-up).

The Weave gives no names — you're right, and my phrasing contradicted the established format. That's now stated explicitly and made binding on the rebuild: short present-tense fragments, true and under-specified, a hill above a village being the kind of thing it can say. redacted sentence: unpublished prose 001 may let it.

The interval is a loop, not two endpoints. A question a day to the Weave, everything else asked of people, each fed back into the other across coaches and taprooms. The names and the whole genre frame come from humans and the reading; the rumour and fact arrive indiscriminately mixed. redacted sentence: unpublished prose redacted sentence: unpublished prose

redacted sentence: unpublished prose His grounds are his own and honest — farmer first as a plain answer rather than a modest deflection, no knowledge of fate magic whatsoever, and no recognition of "saving the world" as a description of anything he does. He's helping people, a parish at a time. redacted sentence: unpublished prose