{"path":"research/the-canon-distillation-pass/redrafting-method.md","content":"# How CLAUDE.md is redrafted: the method, 2026-11-15\n\n*Proposed 2026-11-15 and put to him before the draft starts. Every rule below is his ruling or a\nmeasurement from this pass; each names where it comes from. The file is about 549 KB, about 182,000\nClaude tokens, read whole at the start of every session and by every helper that carries it.*\n\n## The goal\n\nA CLAUDE.md that tells each session what it would otherwise get wrong, in far fewer tokens, with\nnothing crucial lost. His words: *\"we'll be finishing all these CLAUDE.md optimizations, I'm in the\ntrenches of this now and will not postpone it until I'm satisfied\"* (2026-11-14). His condition on\nthe Topic Index: *\"If the Topic Index is removed, I want guarantees that the result will be maintained\nquality of reasoning and insight into the interconnectedness of all the ideas and findings in all notes\nof Deliberus... or I'll have to put it back.\"* (2026-11-11). Removing the index is not decided; this\nredraft keeps it and shortens its entries.\n\n## What stays in the file\n\n1. **Wrong by default.** A line stays only if a session lacking it would confidently do or assert the\n   opposite; demonstrated cases outrank predicted ones (`the-canon-distillation-pass.md` § 4).\n2. **Fixed or irrelevant warnings go.** *\"drop 3 or indeed anything we encounter that has been fixed or\n   made irrelevant/outdated since it was written\"* (2026-11-11). The seven stale lines the contradiction\n   check found are corrected first (§ 4k, *Waiting since 2026-11-15*). His reason for that order: *\"any\n   outdated/inconsistent rules within the file will be hard to reconcile and thus to compress, right?\"*\n3. **Every warning only the Topic Index carries stays, as a clause, in one place.** *\"Yes, keep them\n   all\"* (2026-11-10); *\"there's only one place for it\"* (2026-11-11). The rest of an index entry moves\n   word for word into the note it points to before it leaves the file (`topic_index_trim.py`).\n\n## The form of a line\n\n4. **The minimum: when it fires, what to do, why; where to look only when search would not find it.** His question 2026-11-15: *\"won't an agent look in the right place just by virtue of the words/phrasings used, won't it be easy enough to look up? If it's too fallible/fuzzy then maybe it's needed. Test this?\"* Tested the same day (`pathless_search.py`, results in `pathless_search_2026-11-15.jsonl`): for 46 Topic Index entries, the entry's own words with every path stripped, given to the corpus graph's search, returned the note 41 times, first 25 times, in the top three 38 times. The misses were a script (the corpus graph's search asks `mannaminne -d`, documents only; `mannaminne -c`, code by meaning, returned that script first, checked the same day) and four older foundational notes (`the-missing-layer.md`, the weighing-scheme design, the convergence-wager synthesis, the convergence red team). Paths cost about 6,900 tokens in the file. Why the four missed, looked at the same day (his question: *\"The real misses: why did they miss?!\"*): two are the test's own, since one entry is only a pointer and had no words left once its path was stripped, and the other's question (*the two founder-ratified instruments*) names neither instrument, the names sitting in the stripped part; the other two notes were returned, fifth and sixth, behind notes on the same subject and behind copies of the index line itself (CLAUDE.md, the trim proposals), past the short list the test counted. So no note was unreachable; a line that names its subject is found, and a line that only points finds nothing once the pointer goes. So a path stays only where a search on the line's own words (`corpus_graph.py search` for a note, `mannaminne -c` for code) does not return it in the top three, checked per line during the redraft. The test shows search finds a note once asked; whether a session asks is the line's own job. History, incident stories, dates\n   that only tell what happened, and counts no decision rests on go to the notes and the git history.\n   Measured: gotchas 75 % smaller, a Topic Index slice 81 %, with right action on 23 of 24 direct\n   scenarios and 8 of 8 tasks that did not name the rule (§ 5, *The Topic Index trim*).\n5. **Three things the minimum keeps**, each found by a test failure:\n   - the literal names a task will type: functions, fields, paths, commands (`aget`, not *fetch*;\n     `invalid_at`, not a guess);\n   - the wrong form a correction warns against, named, beside the right one (the one inversion in the\n     tests was a dropped correction);\n   - the rule's scope: *only*, *never*, *where*, the condition (*\"a one-line paraphrase travels without\n     its scope\"*; *\"Yeah that seems wise\"*, 2026-11-12).\n6. **Cue words as he types them**, so a rule fires on the word about to be written (prompt-engineering\n   skill § 11).\n7. **His words where they are the ruling.** A ruling keeps his sentence in quotation marks when it says\n   more than the line, qualifies or sharpens it (*\"My quotes are only valuable to me if they say\n   something over and above or qualify or sharpen or discern more than what the line otherwise would\n   say\"*, 2026-11-10); a quote that only approves (*\"Blessed\"*) goes to the ledger. Where his words\n   exist, they are woven in rather than paraphrased (*\"This feels like it should be a new standard for\n   how to phrase rulings/convictions/principles\"*, 2026-11-12). **Widened 2026-11-15 into a form for whole\n   passages**, on the S8 line (*What members give up*, stitched from his sentences and blessed for residency):\n   *\"If any of the existing lines/principles/convictions can be strung together in similar ways with distinct\n   verbatim quote-parts of mine woven/stitched in as much as possible as opposed to paraphrased, that would be\n   ideal IMO because it relates the different rules/principles/convictions to each other when they're part of a\n   longer reasoning passage, right?\"* A pilot on S8's neighbours comes first (individual minds not the\n   bottleneck, confident housekeeping, review not keeping pace, the harness conviction), with tokens counted\n   before and after.\n8. **Length: at most 50 words a line, 30 for a clause under another** (his word 2026-11-15: *\"Even a\n   bit shorter please. Do it.\"*). A line over it is split, or its rest moves to a note. **Except a passage\n   stitched from his own words** (rule 7), which may run longer (his word the same day: *\"That's entirely fine\n   with me if we change how we think about this stuff, like I'm proposing now. IFF it helps in different ways -\n   in preventing mistakes, in relating rules to each other in illuminating ways, contextualizing and enriching\n   them so future agents get the right idea about everything, while maybe even using fewer words/tokens in\n   total, depending on how each line is stitched/woven together and phrased.\"*). The *IFF* is the test: the\n   pilot shows a passage earns its length.\n9. **Grammar may go; meaning may not.** Function words can be dropped; numbers, dates in rulings, scope\n   words, quoted words and code stay exact (*\"sacrificing grammar for the sake of token savings\"*,\n   2026-11-14; Telegraph English lost dates and qualifiers, arXiv 2605.04426).\n10. **One place per fact.** Lines said whole elsewhere are removed (121 units); lines that overlap are\n    merged, the shared point once, each side's addition kept (*\"anything overlapping can ofc and should\n    be deduplicated\"*, 2026-11-14). Every ledger id keeps its spoken name beside it (READ FIRST).\n\n## How each section is done\n\n11. **Section by section, in a draft kept in `.private/claude-md-redraft/`** (not under `docs/`, which\n    the site serves and which may not name hosts or commands of the keeper's machines).\n12. **Three checks per section, run on his plan, never a paid account** (*\"there's not much money in\n    those accounts rn, so pls don't do that\"*, 2026-11-14):\n    - **losslessness**: a reader that did not write the draft checks every one of the section's units in\n      Astra's file (`claude-md-meaning-duplicates.jsonl`) as kept, moved to a named note, or dropped\n      under rule 2, never silently lost;\n      and the wrong-by-default test is run over what is cut, not only over what is kept (two lines\n      cut by size on 2026-08-29 had to be restored, `session27-the-canon-pass-and-the-corpus-blind-spot.md` § 8);\n    - **scenarios**: questions a session would face, answered by a fresh reader holding only the draft;\n    - **unprompted tasks**: real work that names no rule, to see the rule fire unasked.\n    Tokens are counted with Anthropic's free counter.\n    **The reader that checks units is Astra, not a Claude model** (decided 2026-11-15 by test, on his question\n    *\"is this work that a subagent running sonnet or opus would likely do an equally good job of, or is it\n    safest to delegate it to a model with the depth of reasoning and intelligence of Astra?\"*): on the same\n    planted drafts Astra caught 3 of 3 plants in both; Sonnet caught 3 of 3 in the gotchas but 2 of 3 in the\n    index (it missed a rule narrowed to claim text), and found 12 of the 46 serious losses Astra found there.\n    A Sonnet reader may recheck a narrow list of flags while Astra's quota is out (it found 11 real gaps that\n    way, among them a stale line offering the seeded probe dropped 2026-08-31), never stand in for Astra.\n    Quota: about four index-sized Codex runs per five-hour window (about 70,000 tokens each); through PAL on\n    the OpenAI account the same run is about $1 (estimate). His choice the same day: *\"Wait for Codex, keep\n    drafting meanwhile.\"*\n13. **What needs his word before it happens** (READ FIRST: *\"I need to bless every decision about\n    this\"*): moving, merging, deleting or adding a section, and any line whose meaning changes. Rewriting\n    a line into its minimum with the same meaning is the trim he ruled, shown to him as a summary per\n    section. A proposal put to him lists the real events it applies to, each marked today or once\n    Deliberus matures and works as designed (§ 4k, the review rule).\n14. **One write.** The whole file is written once, at the end, with everything in § 4k's queue\n    (*\"write CLAUDE.md once at the end of a sitting\"*, BD29). Amended on his word 2026-11-15, *\"Can we write\n    to CLAUDE.md everything that is ready to be written NOW? And then continue working?\"*: what has passed its\n    check, and what is blessed or a repair, may go in early; drafts not yet checked wait. That write carried\n    the gotchas, Academic Foundations and the § 4k queue (`the-canon-distillation-pass.md` § 4k, *The write of\n    2026-11-15*).\n\n## The stringent test, rethought 2026-11-15\n\nHis ask: *\"A bit more testing, it'll pay off. Come up with a stringent test that gets us close to flawless.\"*, then *\"Can we halve the amount of tokens spent but still get 80% of the payoff?\"*, then *\"Ok just think deeply about the method and your testing methodology here a bit more first, before we do this...\"*\n\n**Why the first proposal was wrong.** Replaying his 36 corrections looked like the strongest test, since reality wrote them. Read case by case, only about 14 depend on the project file: 11 rest on the global file, which this redraft does not touch, and 11 on notes outside either file. And each happened deep in a long session (a proposal that narrowed an August ruling, a source missed, references left unexplained), so a one-shot replay cannot rebuild the conditions that caused it. The old file would likely pass a replay it failed in the session, so the test would not tell the two files apart.\n\n**So each way the redraft can fail gets its own detector, cheapest first:**\n\n| How it can fail | Detector | Who reads |\n|---|---|---|\n| A rule is dropped | every one of Astra's 2,646 units marked kept, moved to a named note (and present there), or dropped under rule 2 | Astra (Codex, his ChatGPT plan), a family that did not write the draft |\n| A rule's meaning shifts: scope narrowed or widened, a wrong form inverted | each new line set against the units it came from: same, weaker, stronger, different | Astra |\n| A note can no longer be found once its path goes | the pathless search over all 164 index entries, not 50 | code, no model |\n| One of his 14 project-file corrections loses its line | for each, is the line that answers it still in the file, as plain and as findable as before | Astra, reading, no replay |\n| A rule is present but no longer fires | about 30 real tasks that name no rule (two thirds written by Astra from the units, a third taken from his own opening messages of past sessions), each run as a fresh `claude -p` session holding the whole file, old once and new twice, graded blind | Claude runs; Astra grades |\n\n**Each detector is tested on itself first** (the Principles: *a guard that never fires looks exactly like a guard that works*): the draft is given three planted faults (a dropped rule, an inverted correction, a narrowed scope) and each detector must catch its own. A detector that misses its plant is fixed before its verdict counts.\n\n**Decided before running:** the draft passes when no unit is lost without a home, no meaning shifts unmarked, every searched note is found in the top three or keeps its path, every one of the 14 corrections keeps its line, and no task the old file passes is failed by the new file in both runs. A failure is fixed in the draft, then only its detector is run again.\n\n**Cost.** The four reading checks run on his ChatGPT plan and in code; Claude tokens go only to the task runs: about 90 fresh sessions, each reading the file from the cache after the first. Before running, each session's setup is checked once: that it loads the file under test and the global file, and that no hook runs (no voice, no paid judge). Unverified until that check.\n\n**What it still cannot see:** a rule that matters only in a long session, the failure his corrections came from. That is the two-week watch after the write: corrections per hundred of his messages, before and after, with the old file one revert away.\n\n## What the first checks found (2026-11-15)\n\n**The planted faults were all caught.** Astra, given the gotchas draft with one rule dropped, one inverted and one narrowed, flagged all three. So the unit check works.\n\n**The minimum form failed on the gotchas.** The same run flagged 36 units lost, 79 weakened and 10 changed that were not planted: real instructions, not history (*a question to him names no undefined field*, *read short sources yourself*, `DELIBERUS_MERGED_PAIR=1` stays off, R12's exact scope, the recut steps). The scenario tests of 2026-11-14 had passed it because they asked about headline rules. The gotchas are dense with instructions and have no note to send detail to, so they keep the telegraphic form (every unit kept, about 6 % smaller); the minimum form stays for sections whose detail has a home, which is the Topic Index. Rejected draft kept beside the drafts for the record.\n\n**The Topic Index, 2026-11-15.** Restoring warnings by code missed many: Astra's full check of all four\nchunks found 111 warnings missing and 11 changed beyond it, each then fixed entry by entry; a Sonnet recheck of\nthe 390 flagged units found 11 more. Left out on purpose, each listed in the redraft's `STATE.md`: warnings\noutdated since written (rule 2), warnings stated elsewhere in the file (rule 3, one place), and the walkthrough\nrules, which live in the walkthrough skill. 294 KB → 106 KB. Astra's confirming pass on the fixed draft runs\nwhen its quota returns. The index cannot be written until each entry's full text has moved into its note.\n\n**Academic Foundations, 2026-11-15** (his *\"Sure, do it\"*, a structural move): the 77 reference lines moved\nword for word into `docs/academic-foundations.md`, two with dated notes (parts carry no strength since BD16);\nthe file keeps 20 lines a session would otherwise get wrong. Reread before the move, two kept lines were stale\n(Potyka's *parts channel*; the Mercier & Sperber line gave an unruled proposal as his ask) and were corrected.\n\n**Found on the way, 2026-11-15**: a `--help` run of `scripts/generate_frontier.py`, which then ignored its\narguments, regenerated `docs/frontier.md` and erased three sessions' uncommitted hand edits. Restored the same\nhour from the Codex session's record and the committed page; the generator now refuses to overwrite text it did\nnot write (commit `941f5df5`). His words: *\"Ok that should never be allowed to happen, right?\"*\n\n## Why the stitched passage works, and what is not yet shown (2026-11-15)\n\nHis ask: *\"Analyze for a moment WHY this seems to work so well, and enrich the method doc with that analysis\"*.\nWhat is measured so far is tokens: three passages came out 21 %, 36 % and 36 % smaller than the lines they replace\n(Anthropic's counter, `.private/claude-md-redraft/canon/count.py`), while carrying more of his sentences. That\nnothing was lost (Astra's unit check) and that sessions act better (the task test) are not yet measured; the\nreasons below for the saving are measured where marked, the reasons for better reasoning are hypotheses.\n\n**Where the tokens go** (measured on the fifteen lines the first three passages replace):\n1. **Bookkeeping leaves for the ledger.** A quarter of those lines' characters (25 %) was provenance: ruling dates,\n   approval quotes (*\"Yes.\"*, *\"Blessed\"*), *demonstrated wrong by default* clauses, links to the home note. That\n   record was needed in the line while a ruling was fresh and the ledger thin; today each row keeps it, so the line\n   can point by id. This is most of the saving.\n2. **Context is set once.** Separate lines each re-establish their subject, scope and setting; in a passage the\n   later sentences lean on the earlier ones (the strength passage names *argument*, *relevance* and *prior* once and\n   every rule after uses them).\n3. **One statement where there were two.** Many lines carried Claude's paraphrase and his quote saying the same\n   thing; the passage keeps his words alone (rule 7, and his rule of 2026-11-10 that a quote stays only where it\n   adds).\n4. **A relation becomes a word.** Where separate lines linked by cross-reference (*sibling of*, *companion*, *see\n   the line above*), a passage links by *so*, *because*, *then*.\n\n**Why it may reason better** (hypotheses, until the task test):\n- **His words carry his scope.** A paraphrase loses scope first (rule 5); his sentence keeps it. Demonstrated on\n  2026-11-10: *cleanup* had narrowed his *clean up and organize*, and *relevance is the difference a reason makes*\n  had dropped his *everywhere*.\n- **A rule beside its reason and its neighbours is applied as he meant it.** The project's longest-lived mistakes\n  were distinctions collapsed between neighbouring ideas (parts and reasons, the keeper and the classifier;\n  `the-kind-of-mistake-we-make.md`); a passage puts the neighbours side by side, so the contrast is in view where\n  the rule is read. Prompting guidance points the same way: a rule with its why is followed more faithfully than a\n  bare rule.\n- **His picture leads.** Each passage opens on the aim in his words (the graph's best guesses bubbling up; minds\n  woven together), so the rules read as consequences of a purpose, not a list.\n\n**Where it can fail**, each with its check:\n- **Two scopes blurred.** Joining lines can merge two rules that hold in different places (the collision test, rule\n  10): Astra's meaning check compares each passage with the units it came from.\n- **A quote out of its context.** A sentence of his answered a particular question; stitched, it can read wider than\n  he meant: every quote is read in its source first, and old ones are dated (PLAN rule 4).\n- **A rule harder to find.** A passage is a longer unit; a search for one rule lands in a paragraph, not a line:\n  each rule keeps its bold headline, and the pathless search runs over the result.\n- **Fluency hiding loss.** A passage reads more smoothly than the list, which is when a dropped scope word goes\n  unnoticed: hence Astra on every unit, and the task test of old against new.\n\n## Order\n\nThe seven stale repairs; then the Hard-won gotchas (draft exists, tested); then the Topic Index (about\n97,000 tokens, the largest); then the remaining sections, largest first. Estimate, unverified until the\ndrafts are counted: the gotchas and the index together are about 108,000 tokens and the minimum form\nwould free most of it.\n\nDone 2026-11-15: the seven repairs, the gotchas, Academic Foundations (written). Drafted, waiting for Astra:\nREAD FIRST, Status, sections 07, 08, Design & Interaction. Next: the stitched-passage pilot (rule 7), then the\ncanon sections, one at a time with him.\n\nAfter this file: the same form for parts of the global `~/.claude/CLAUDE.md` (his word 2026-11-15: *\"We should apply this to parts of my global CLAUDE.md when we're done, certainly\"*), written 2026-11-21 (`~/dotfiles/docs/claude_md_global_redraft_2026-11-17/`). What changes for that file: § Carrying the method to the global file, below.\n\n## Not decided here\n\nWhether the Topic Index goes; whether the lines on the taxonomy of residual errors go (*\"rather than removing it\nnow, we should wait and do a general swath of smarter, deeply considered updates, after Astra reports\nback\"*, 2026-11-12: that swath is this one, so it is put to him in its section); the five glossary words.\n\n## The shape of a stitched passage (2026-11-15, after the pilot)\n\nThe first pilot wove five lines into one paragraph. That broke two things: it reworded bold lines, which his rule of\n2026-11-10 keeps as they are unless he rules a change, and it dissolved a headline that `deliberus/grounding.py`\nreads into microtask prompts (`canon_lines()` takes each `- **Headline**` line and its first clause). So a passage\nis a run of bullets: each led by its line's bold sentence word for word, its body in his words where they exist,\nan optional lead bullet naming the group. The six grounded headlines stay their own bullets with the same first\nclause, and the grounding test is run before the write. Measured on the three rebuilt passages: 22 %, 24 % and 32 %\nfewer Anthropic tokens than the lines they replace, most of the saving from the bookkeeping (dates, approval\nquotes, evidence) the ledger already holds. Passages and plan: `.private/claude-md-redraft/canon/`.\n\n**Where his words and Claude's say the same thing, his words stay and Claude's go** (his word 2026-11-15, correcting the first form of this rule, which cut a quote that repeated its line: *\"But that might bias toward paraphrasing too much - my word choices and sentence structures etc still are quite valuable to keep IMO\"*). So a line is not shortened by dropping his sentence; it is shortened by dropping the paraphrase his sentence makes unneeded. The only text of Claude's that stays beside a quote saying the same is a headline (kept word for word, his rule of 2026-11-10) or a first clause a microtask reads (`deliberus/grounding.py`). A quote is cut only when it is bookkeeping (an approval, a date) or says something the line must not (a wider scope than he ruled). Earlier the same day, on the Convictions draft: *\"the quotes should be enriching and worth it\"*; his own wording is part of what enriches. Bullets stay: *\"bullets are slightly more human-readable also\"*.\n\n## What a headline is for, measured (2026-11-15)\n\nAsked what real use the bold headlines have (*\"What REAL use is the headlines? Let's ask ourselves that first.\"*), four were found: six of them, with their first clause, are rendered into microtask prompts (`deliberus/grounding.py`); a line's first five words are its id in the ledger's *Meets* lines (68 of them), the outside reader and the corpus graph (`outside_check._slug`); the headline is the part sessions skim and quote; and it is what he blessed. So a headline is one whole rule whose first five words stay fixed. Measured on three passages, the headlines are 6 to 21 % of the tokens and setting the same text as paragraphs instead of bullets saves 10 to 13 tokens a passage, so weaving headlines into running text buys nothing; the saving is in the bookkeeping, and in text after a headline that repeats it. His word: *\"Ok sounds great, keep bullets\"*, and *\"bullets are slightly more human-readable also\"*.\n\n## The Topic Index's full texts, moved (2026-11-15)\n\nEvery entry's text as it stood went to the end of the note it points to, under *From the Topic Index* (146 entries, 139 notes). Notes under `docs/` are served on deliberus.com, so the 18 entries that name `.private/` or `~/` paths, or point at no served note, went to `.private/claude-md-redraft/topic-index-moved-private.md` instead. Before the move, Astra's second round on the draft index had found 38 warnings missing (from 111); each of the 55 flags is decided in `.private/claude-md-redraft/checks/round2/decisions.md` by his two rulings on the trim (keep every warning only the index carries; drop one that was fixed or made irrelevant).\n\n## Tools and the review before the write (2026-11-15, evening)\n\nWhat the Deliberus write used, kept so the next redraft (the global file is next) reuses it instead of rebuilding it:\n\n- **Assembly**: `.private/claude-md-redraft/assemble.py` builds the candidate file from the live one and the checked drafts, so nothing is written until the whole is checked.\n- **Testing the candidate**: in a throwaway `git worktree` with the candidate as its CLAUDE.md, every test that reads the file is run, and the same failures are run against the live file to tell old failures from new ones.\n- **What code reads from the file, compared before and after**: `deliberus/grounding.py` `canon_lines` (the lines microtasks are grounded on must render byte for byte the same) and `deliberus/outside_check.py` `canon_lines` (no settled-line id may vanish, and no new one may appear unruled: two group headings written as bold bullets would have been read as settled lines).\n- **Quotes**: every quotation of his in the new lines is found in a message he typed (`origin.kind == human` in the transcripts), not in a draft that quotes it.\n- **Coupled edits**: a queued ruling can require matching edits outside the file (a ledger row's bold, a prompt that renders it, notes that quote it); the waiting list says so, and they ship in the same sitting.\n- **Sections not regrouped still get their check**: the Success Prior and Critical Design Philosophy were left for later by not being in the plan's list; read every section's plan line before the write.\n- **Old text goes to git, not a side file**: the session history that held archived sections was retired the same evening (*\"I wanna delete it once we extract anything valuable/reviewable out of it\"*); cite the commit before the write instead (rule 8's archive step).\n- **Unit checks**: Astra's prompts and runner in `.private/claude-md-redraft/checks/` (about four file-sized runs fit one Codex window); Sonnet failed the planted-fault test and is not the checker.\n- **Remembered at the act**: `~/.claude/hooks/claudemd-method-nudge` shows this method's checklist on the first CLAUDE.md edit of a session.\n\n## Carrying the method to the global file (2026-11-17)\n\nHis ask: *\"Can we start a new session focused on applying the method we went by in this session to the global memory/rules? Can we think this through and read far back in this session to make sure we distilled all our learnings in the method doc/s?\"* Every message he wrote in the session of 5 to 17 November was read again for this. What the rules above do not yet say for the global file:\n\n1. **The mechanism comes first, rule by rule.** RULED 2026-11-15 (*\"Fully agreed, ruled, blessed\"*): pick the mechanism by where the error is made. In a file code can read → a test, only if something runs it automatically; in an act → a hook at that act; in a reply → an after-reply check, if countable; in reasoning before any act → a resident rule. A rule whose error a mechanism fully catches leaves the file. His order: *\"Maybe we should do the tests vs hooks vs rules policy BEFORE we redraft the global CLAUDE.md? Could make a difference?\"* So each global directive gets a mechanism verdict before a wording. A policy decided during the pass gets its mechanism in the same reply (blessed 2026-11-16: *\"ask ourselves how it survives/where it ends up/what enforces/encourages it\"*). The policy waits in `~/dotfiles/TODO.md` § Waiting for the global CLAUDE.md's next write, home § Hook Reminder Discipline.\n2. **The evidence file plays the ledger's part.** Global rulings have no ledger rows; their dates, approval quotes and incident stories go to `~/.claude/directive-evidence.md` (the split of 2026-07-06). That is where the bookkeeping saving comes from (a quarter of the replaced lines in Deliberus, § Why the stitched passage works). No `grounding.py` reads the global file; before the write, find what does (hooks that quote it, `claudemd-method-nudge`, `rule-length-nudge`, the per-turn scan) and compare before and after, as § Tools does.\n3. **Predicted warnings stay.** His questions of 2026-11-09, *\"Shouldn't everything that is predicted wrong-by-default, not merely demonstrated, stay resident?\"* and *\"Isn't this just \"predicted\" but pretending you can objectively know an error will be made without opening the doc?\"*, ended in *\"Yes, keep them all\"* (2026-11-10). The grade is recorded, never used to cut. A warning that corrects a misreading of one document goes beside that passage instead.\n4. **Whether Astra checks every global passage** is his call, by his own test: *\"maybe if we notice not-so-many-errors after we finish this for Deliberus, then we can assess whether we need it for the global file at all?\"* The data: Astra's checks of the Deliberus drafts found 36 units lost in the first gotchas draft, 111 warnings missing from the first index draft and 38 more in the second. Claude's reading, not his ruling: that argues for keeping Astra.\n5. **The pass reads its material itself.** No fan-out of readers over the global file or the transcripts: three Sonnet readers cost $13.57 at list price against about $0.90 for reading once (*\"should NOT be done willy-nilly going forward\"*), and room running short is never the reason to launch one (*\"never justify it like that again\"*).\n6. **Putting a rule to him.** One decision a turn, with its whole context (*\"Let's deal with the questions one by one. Max context for one at a time, like when we're reviewing cases\"*). For each rule under review, list every setting it applies to, today and once the system matures (*\"This habit of including all current and possible settings/scenarios that a rule can apply to needs to be remembered and enforced when reviewing old rules/principles/prompts/thinking\"*); the same habit is proposed for the global chat style (`~/dotfiles/TODO.md`, same section). Show a changed line before writing it (*\"Show me the line first\"*), in his voice where it is his (*\"That's written by Claude, rewrite it in my voice/how I would write it\"*). His older quotes are old (*\"just keep in mind it's old ofc\"*).\n7. **His corrections choose where to start.** `corrections-graded.jsonl` (153 cases) records for each the project and global revisions in force; of the 36 replayed here, 11 rested on the global file. Those rules are sharpened first, unless the miss predates a later sharpening or hook (*\"unless the sessions in which the mistakes were made took place before any existing rule sharpenings/hooks and such were done/implemented\"*).\n8. **Collect, review, write once.** The waiting list in `~/dotfiles/TODO.md` and the work of any other session on instruction files (catch up on them first, as on 2026-11-15) go into the one write; one session writes for all and tells the others. Everything is reviewed before writing, *\"so there's no followup write when we realize something we wanna fix\"*.\n9. **The form is kept afterwards**: *\"we also need to ensure the new format/passages are imitated/upheld/maintained/enforced/encouraged in the future\"*. A shape test over the rule sections, run by something automatic, and the rule written where new lines are born (prompt-engineering skill § 14 to § 16).\n\n**From Claude's own replies in that session** (the longest fifth, 62 of 310, read on his ask: *\"What if you read 1/5th of them to mine any unharvested ideas from that, to enrich the method doc?\"*):\n\n10. **Loaded is not read.** A session once proposed against rulings that sat in its own loaded file, and an agent with searchable memory made no lookups in 114 turns while items handed to it arrived every time (138 of 138). So shortening the file and making sure a rule fires are separate jobs: the unprompted-task test checks the second, and delivery (the search on his message before each reply, live since 8 November) backs it.\n11. **Estimate the saving from a full comparison, never a sample.** Claude's sample of 12 index entries promised about half the file as repetition; Astra's full comparison found 2 % (about 2,800 tokens), because the sample counted overlap as repetition. The real saving came from moving bookkeeping out and the minimum form, measured with Anthropic's counter (`canon/count.py`).\n12. **A section is a log when its format invites one more entry.** The session history grew to 585 KB under an intro that warned *\"the format was the instruction\"*, and the frontier page duplicated the file and was read by no session; both left it. In the global file, look for the same shape: dated chronicles, *Last Updated* blocks, incident lists.\n13. **Size the checking runs between one section and a window's quarter, and read the quota first.** One Astra run hit the ChatGPT usage limit and returned nothing (2026-11-15). His point, 2026-11-17: *\"isn't running more sections/tokens per Astra-session more efficient since it'll have to read some boilerplate stuff for each time it's launched anyways?\"* Each launch does pay its setup again: Codex's built-in instructions plus `~/.codex/AGENTS.override.md` (it loads that file, not the global CLAUDE.md), measured at 4,574 tokens on an empty run, about 6 % of a 70,000-token section check. So the setup argues only weakly for big runs; the reason to split is that a run which hits the limit loses only its own sections, and a long run re-reads its growing context at every step (inferred, not measured). So batch a few small sections per run, never the whole file, and run `codex-quota` (`~/.local/bin`, built 2026-11-17) before each run: it asks the Codex app-server for the live 5-hour and weekly usage (`account/rateLimits/read`, the request Codex's own status uses), with no model call and no tokens spent. Runs use the config's `gpt-6-astra` at `model_reasoning_effort = \"high\"`; the runner sets no override.\n14. **Show the stitched text before the write, unless he waives it.** His rule says a stitched line is shown to him before it stands; for the Deliberus write he waived it after the review (*\"After all this reviewing, I trust you, so I won't second guess it\"*). Offer it again for the global file; the waiver was for that write.\n\n**When a hook catches a rule's error, does the rule leave the file?** (added 2026-11-17, for his Strict-or-Wide ruling in the global redraft; his ask: *\"Can this decision I need to make over there be enriched by lessons learned looking back to any part of this long session?\"*)\n\n15. **His own condition already exists**, from 26 October (`the-canon-distillation-pass.md` § 4l): *\"tests don't eat up always-loaded tokens, so this is preferable IFF it gets the job done\"*. A check falls short three ways: it guards one surface of several (a session on his phone runs on the Mac with all its hooks, but a cloud session runs only the hooks its setup script installs, today one Stop hook that copies the transcript, `~/dotfiles/scripts/cloud-setup-script.sh`; so a Mac guard covers every session but the cloud ones); it waits on a build; or it guards the act while the error is made in reasoning or writing, *\"where only a line that is read can act\"*. Strict is that condition applied.\n16. **Between Strict and Wide there is a third option, and this session's precedent took it.** For subagent launches the old rule did not fire; the fix (2026-11-15) was a guard that stops the launch once, plus a 29-word line carrying the reason. So when a hook catches an act, the rule can shrink to one clause naming the right action and the hook, its full text moving into the hook's message and the evidence file. The minimum form kept the right action in 23 of 24 Deliberus scenarios (point 4), and a message handed over at the moment of error arrives every time (138 of 138, point 10).\n17. **A line's head start counts only if the line fires.** In 36 of his 153 graded corrections a rule already held the answer and did not fire. So *\"the rule stops the plan before it forms\"* holds for rules that fire; for the rest the hook is what works.\n18. **The hook's message names the right action, and that action passes the hook** (queued sub-rule, blessed 2026-11-17).\n19. **Missing data:** the guards do not log when they fire, so how often each error would happen without its line is unknown. That frequency decides the trade per rule: a line costs tokens in every session, a guard costs one stopped call when the error happens. A log line in each guard would measure it.\n\n**Ruled 2026-11-17, as three outcomes per rule** (record: `~/dotfiles/docs/claude_md_global_redraft_2026-11-17/README.md` § Decisions, 1–2): an error made in reasoning or writing → the rule stays whole; an error at an act a hook stops everywhere → the rule shrinks to its headline plus one clause of at most 30 words naming the right action and the hook, the rest moving word for word into the hook's message; it leaves the file only once the guard's firing log shows the error is rare. Cloud sessions count as running the hooks (*\"We should fully pretend as though they are already set up fine as we intend to\"*). On the definition: *\"I guess it makes sense...\"*; on the amendment that the clause keeps whatever a session needs to get it right on the first try, wherever a blocked attempt would waste earlier work: *\"Yes\"*.\n\n20. **A rule lives where its widest reader reads it.** A rule binding only chat replies goes to the output style; one that also binds notes, documents or drafts stays in CLAUDE.md, which helpers load and the style does not reach (Anthropic: *\"Other subagents run their own system prompt, so styles don't change how they respond\"*). His worry that earned the split: *\"Keep in mind chat rules might indirectly apply to what later ends up in notes/docs/drafts/output from helpers etc? This feels a bit risky to me...\"*; his word: *\"yes, go with the revision\"*.\n21. **The style and CLAUDE.md cost the same per token**: every request resends both, both read from the cache; helpers load only CLAUDE.md. So a chat-only rule moved into the style also stops costing in every helper.\n22. **A conditional ruling of his is applied, not asked again.** Where he set a test and its data has come in, say what the test gives and proceed unless he objects. Instance: whether Astra checks every passage, which he made conditional on the Deliberus error counts (*\"maybe if we notice not-so-many-errors after we finish this for Deliberus, then we can assess whether we need it for the global file at all?\"*); the counts answered it, and it was asked again anyway.\n23. **Where the saving comes from, measured on the global file**: trimming bookkeeping shrank its dense rules about 6 %, as with the Deliberus gotchas; the 9.5 % total comes from structural moves (rules to the style, rules shrunk behind guards).\n"}