2026-08-24-cleanup-and-compress

Prompt B — cleanup and compress

Round id: 2026-08-24-cleanup-and-compress. Project: none (workspace). Prompt-author-model: Claude Fable 5. Dispatched by the maintainer after #52 merged; cut from main at the #52 merge sha. This dispatch grants scope to file priority items, to land agent-decided entries under §10, and to edit spec-layer documents. AGENTS.md as amended by #52 binds you, including §10's bar: adopt what you can; reserve only what turns on the maintainer.

What this round is

The workspace's live documents have grown by accretion: PRIORITIES.md is 2,085 lines for 68 items; the wiki's status blocks lag the rounds they describe; README.md and the wiki still describe a deference line as active when it is paused on a decision; the deference line's kernel-checked results have no registry; and the dependency view has eleven silent entries. None of this is research. All of it makes the repository harder to read for the next round and for the paper.

The governing constraint is nothing is lost: a sentence leaves a live document only if its content is in a round record that a pointer still reaches, or the item it belonged to is closed with a statement of record. Compression is relocation and pointer-repair, never deletion of content that exists nowhere else.

Rulings

R1 — Item numbers are permanent. answers_item, depends_on, and wiki bindings key on them. Compression never renumbers.

R2 — Closed items collapse. A closed priority item becomes: its title line, one status line (what closed it, the statement of record or the round), and pointers. The body moves nowhere because the round record already holds it; confirm that before cutting, and if the round record does not hold it, the body stays.

R3 — Open items keep their body but lose their history. Superseded narrowings, "as of round N" paragraphs, and status updates that a later status supersedes are cut; what remains is the current statement, the current narrowing, and the consuming round if any.

R4 — The wiki states current status only. Status blocks bind to state (wiki/CONVENTIONS.md) and say what is registered, what is open, what is a living note, and — for deference — that the line is paused and on what. History belongs on history pages or in round records.

**R5 — depends_on may be read from PROMPT.md.** A round's dispatch is part of its record. Where a REPORT is silent and the PROMPT names what the round builds on, the field records that, with "depends_on_source": "prompt" on the record so the two sources stay distinguishable. Still no inference.

R6 — The deference line gets a registry. projects/deference/CLAIMS.md, created under the same rule that governs normativity's: kind: lean, class lean-proved, one entry per theorem a round's report presents as a result. This retires friction F3 as a friction entry.

R7 — Worktrees are ignored. Add the .gitignore line for agent worktrees. No gate; the friction entry records that the maintainer ruled a gate unnecessary for an ignore rule, and closes.

R8 — A dead-pointer check exists. Friction F6 (a pointer into a superseded tree still resolves and nothing says it is stale) is fixed here as a contained gate: every repository-relative path in the live documents (README.md, AGENTS.md, CONTRIBUTING.md, PRIORITIES.md, DECISIONS.md, PROVENANCE.md, wiki/, each project's CLAIMS.md and notes/) resolves, and any path under a tree whose README or FORWARD.md declares it disposable or superseded is listed with that status. Null input: a document citing a path that does not exist fails; a document citing a path under a declared-superseded tree without saying so fails.

R9 — The naming audit is prepared, not run. Naming is the maintainer's by batched audit (§6). This round produces the sheet the audit needs and does not rule on any name.

Work packages

WP1 — PRIORITIES.md

Apply R1–R3 to all 68 items and the friction section. Expected result: under 800 lines. Report the before/after line count per section and, for every item, the disposition (collapsed / trimmed / untouched) and the round record that now holds the cut body. workspace_state --check must still count 68 priorities, and every answers_item in both registries must still resolve.

The Where ingenuity is the bottleneck section is not compressed; it is reread for items that #52's queue triage moved or answered (Q3's queue entry is the live one) and made consistent with the queue.

WP2 — Wiki

• wiki/Deference.md: status block says the line is paused on the settlement decision (the trilemma — pointwise, classwise, or weakened target; see projects/deference/ for the round that posed it), that the kernel-checked results are registered (after WP4), and that the impossibility results are the line's companion findings. Nothing about the line reads as in progress.
• wiki/Legitimacy.md and wiki/Normative-Record-and-Inquiry.md: cover the reason-state waist, transition certificates, and the certified interactive service at the level the page's existing sections cover earlier rounds — one paragraph per component, what it fixes, what it consumes (from depends_on), its status under §9's vocabulary. Apply the 𝓡_n / N_{≤n} split #52 adopted, which that entry left to the next round to touch the page.
• wiki/Normativity.md: check the status block against the registry after #52; trim the φ-regret arc to a history pointer if a history page exists, otherwise leave it.
• wiki/Home.md, wiki/Architecture.md, wiki/Roadmap.md: reconcile with the above; Roadmap lists open items only and each open item points at its priority item number.
• The relation-to-field page is not written. If the wiki lacks the placeholder, add a one-line stub saying it is the maintainer's to write.

WP3 — README.md and CONTRIBUTING.md

Read both against the current tree. Fix what is false (layout section, command list, gate count, any description of the deference line). Do not add sections. Report each change as before/after.

WP4 — Deference registry (R6)

Enumerate the theorems the deference rounds present as results (the Bridge, Adjunction, TotalTrust, reachable-corrective-control, exposure-geometry, envelope-dominance, faithful-acceleration, magnitude-prediction rounds and whatever else lean/Workspace/Deference/Contrib/ holds) from their REPORTs; create projects/deference/CLAIMS.md; register each with answers_item where a deference item exists and a new filed item where the report answers a demand no item states. The registry checker and axiom audit pass. Confirm each declaration appears in a #print axioms line. State in the REPORT the count and which results were left out and why.

WP5 — depends_on from prompts (R5)

Extend the round-record check to accept the optional depends_on_source field. Fill the eleven silent entries from their PROMPT.md where it names consumed rounds; leave [] where the prompt is silent too, and list those. Re-emit rests_on.

WP6 — Naming audit sheet (R9)

state/views/NAMING_AUDIT.md: every name marked provisional in live documents and Lean identifiers, one row each — name, where it is introduced, where it propagates (Lean / wiki / prose), the round that proposed it, and the alternatives that round listed if any. Grouped by line. No recommendations. This is the input to the maintainer's batched naming audit and nothing else.

WP7 — Friction (R7, R8) and pointer repair

Land the ignore line and the dead-pointer gate with its null-input fixtures, wire it into ci.yml and tests/run.py, and run it against the tree; repair every pointer it finds, or annotate the superseded-tree ones. Update the friction section: F6 and F9 close; F1/F2 stay merged-and-deferred; F4 stays (its decision is in the queue).

Non-goals

Research edits of any kind. The legitimacy pressure test (next round). The relation-to-field page. Renumbering. Editing any round's PROMPT, REPORT, README, or MEMO. Ruling on names.

Acceptance

• Suite, all gates, workspace_state --check, both registry checkers, and the axiom audit green; Lean builds in CI.
• PRIORITIES.md under 800 lines; 68 items; every answers_item resolves.
• The dead-pointer gate is green on the tree and its null inputs fail.
• No wiki page says the deference line is active.
• state/views/NAMING_AUDIT.md exists and contains no recommendation.
• git diff --stat shows no file under prompts/*/ or */rounds/*/ changed except state/views/.

Report

prompts/2026-08-24-cleanup-and-compress/REPORT.md and the round record (depends_on: [], this is a workspace round). Sections: per-item disposition table for WP1 with the record holding each cut body; wiki changes as a list of status-block befores and afters; README/CONTRIBUTING befores and afters; the deference registry table; the depends_on fills with their source; the naming-sheet row count by line; the dead-pointer gate's findings and repairs; deviations with reasons; what was not shown — in particular that compression moved content and established nothing, that the deference registry asserts kernel checks and not that the paused line's theorems are the ones wanted, and that depends_on from prompts is a reading of dispatches, not of what the executor in fact used.

Attribution block on the PR; Model: trailer per commit. Auto-merge on: this round is mechanical and every acceptance check is a gate. Reserve nothing you can decide.

Report — cleanup and compress

Attribution. Prompt author: Claude Fable 5 (Anthropic). Executor: Claude Opus 5 (Anthropic). Dispatched and executed 2026-08-24. Branch round/2026-08-24-cleanup-and-compress, cut from main at 0586da3, in an isolated worktree.

Verdict. The live documents say what is current, the deference line has a registry, and two friction entries close as gates. PRIORITIES.md lands at 1,686 lines rather than under 800, and §8 gives the arithmetic for why the target is not reachable while items stay dispatchable.

1. PRIORITIES.md, per section and per item

section	before	after
preamble	32	32
Where ingenuity is the bottleneck	172	121
Normativity line	106	103
Normativity line — the learning track	186	157
Normativity line — traderized enforcement	467	340
Normativity line — legitimacy	176	176
Deference line	95	90
Deference line — first research wave	208	208
Deference line — second wave	351	279
Workspace friction	143	93
Infrastructure	149	87
total	2,085	1,686

Sixty-nine items throughout; workspace_state --check counts them, and every answers_item in both registries resolves. Seven items collapsed under R2, twenty-nine were trimmed under R3, and thirty-three were already at the floor and are untouched.

Collapsed (R2) — title, status, pointer; the body is in the record named.

item	before → after	record now holding the body
29. Φ-regret reduction	30 → 11	projects/normativity/rounds/2026-08-11-phi-regret-bridge/
48. computable belief sequence	24 → 11	.../2026-08-18-projection-enforcement/FINAL_FORMALIZATION_STATUS.md
50. piecewise-affine facts	26 → 12	.../2026-08-18-maxmin-representation/README.md
21. signed versus magnitude	32 → 12	prompts/2026-08-11-phase-ii-prediction/REPORT.md §1
36. wiki source synced outward	23 → 12	prompts/2026-08-16-wiki-in-repo-sync/REPORT.md
37. volatile quantities	34 → 12	prompts/2026-08-16-wiki-state-bindings/REPORT.md
38. write-scope enforcement	37 → 12	prompts/2026-08-16-wiki-in-repo-sync/REPORT.md

Trimmed to a stub, though the arithmetic calls them trimmed rather than collapsed — 41 (40 → 13), 47 (36 → 13), 60 (37 → 13), 23 (30 → 14), 13 (19 → 12), 28 (68 → 18), 51 (38 → 28). Each names the round record and, where the round registered, the claim identifiers.

The largest trims, all under R3: Q3 (102 → 51), 39 (72 → 48), 34 (73 → 71), 30 (70 → 67), 52 (48 → 46), 46 (47 → 44), 24 (44 → 41), 25 (44 → 42), 27 (42 → 39), 40 (37 → 34), 42 (38 → 36).

Untouched (33 items): Q1, Q2, Q4, 2, 4, 5, 6, 8, 9, 10, 11, 14, 15, 16, 17, 18, 19, 20, 22, 35, 49, 53, 54, 55, 56, 57, 58, 59, F1–F5. Each was already at or near the floor for a dispatchable item.

Q3 was reread against the queue, as the dispatch asks. Its adjudication of the three candidate objects is now the queue entry's, and the item keeps the question, what killed each candidate, and what a good answer must carry. Its third candidate is now registered as corrective.* and filed as item 60, which the entry says.

Where ingenuity is the bottleneck is otherwise not compressed.

2. Wiki status blocks, before and after

page	before	after
Deference.md	"Open / unregistered research. The line has finite witnesses, conditional interfaces, Lean contributions, and adversarial audits, but the modern claims registry contains no substantive deference claim."	"Paused. The line is not being worked…" plus "**Established — lean-proved.** The 31 entries in this line's registry…", the impossibilities named as companion findings, and the two decisions the pause is on
Corrigibility.md	"Open / unregistered research. The exact reachable-control verdict is Dynamics-positive, protection-incomplete."	"**Established — lean-proved, and mostly negative." plus "Paused.** This is part of the deference line…"
Home.md	"The substantive results discussed here are currently unregistered."	both lines' registered results named; "Deference is paused on two decisions; normativity is where the work is"; a Paused entry added to How to read status
Normativity.md	count bound to counts.registered_claims	rebound to counts.registered_claims_by_project.normativity, because the total moved when a second line got a registry
Legitimacy.md	three components uncovered	a new section covering the reason state, transition certificates and certified interactive service — one paragraph each, what it fixes, what it consumes, its status as a living note
Normative-Record-and-Inquiry.md	"this page keeps R for the record until the naming is ruled on"	𝓡_n for the reason state and N_{≤n} for the record, applied through the page
Roadmap.md	12 unattributed directions	open items only, each pointing at its priority number, in three sections with deference marked paused
Relation-to-the-Field.md	absent	a stub saying it is the maintainer's to write, linked from the sidebar

wiki/Architecture.md needed no change: its status block already reads Open / unregistered about the abstract theorem and dynamics, which is still true. No wiki page says the deference line is active.

3. README.md and CONTRIBUTING.md

file	before	after
README.md layout	five entries; state/ and wiki/ absent	seven entries; both added, and projects/ notes that each line has its own CLAIMS.md
README.md verify block	three commands	five; the registry checker and the state check added, and tests/run.py described as running every gate's self-test
README.md pointers	six documents listed	plus one paragraph: normativity is where the work is, deference is paused
CONTRIBUTING.md local commands	13 commands	14; tests/dead_pointers.py added
CONTRIBUTING.md	(unchanged) "Five gates read a diff or a pull-request payload…"	verified still true: seven required jobs, five with local forms

"Seven gates" and "Seven jobs decide correctness" were checked against .github/branch-protection.json and are still exact. No section was added to either file.

4. The deference registry

Thirty-one entries in projects/deference/CLAIMS.md, all kind: lean, class lean-proved. Every declaration was confirmed to carry a #print axioms line.

group	entries	answers_item
delegation bridge and corollaries	3	23
certificate bounds	6	23
exposure geometry	3	23
substitution separation	4	23
magnitude prediction	7	21
static-view factorization	2	28
reachable corrective control	6	60 (filed here)

Left out, and why — the registry's own header says this, so a reader meets it where the claims are:

• FaithfulAcceleration.weight_not_divergent and MagnitudePrediction.squaredError_bdd_of_sharpness_bdd ship no term inhabiting their full hypothesis package, each carrying an undischarged EfficientlyComputable certificate. AGENTS.md's Lean regime forbids promoting them.
• The inherited transcriptions in InheritedAlgebra.lean and the Layer-1 half of FaithfulAcceleration.lean restate another body of work's declarations. That they re-elaborate here is a result about the port; it is not this repository establishing them.
• EnvelopeDominance.lean proves what its name says and not what its round wanted: its maximiser is built from the evaluating agent's own credence, so it represents no distinct future agent, and its dominance statement is sum of maxima >= sum of anything. The round records this as its central defect.
• CartesianFrameBridge.lean states its results over a mirrored fragment of an upstream library at a commit the repository does not pin, under an Iso weaker than the authoritative one. Item 52 is to import the real definitions and delete the mirror; registration belongs after that.

The reachable-corrective-control refutations are registered. A theorem that breaks its own round's protection claims is a result, and the strongest thing that round produced.

checkers/workspace_state.py's "expected exactly one authoritative CLAIMS.md" became "one registry per active project, and at least one", since the old form encoded a fact about the workspace rather than a rule about it.

5. depends_on from prompts

Nothing was filled, and that is the finding. All eleven silent rounds' dispatches name reading lists — "Read at minimum", "At minimum read closely", "Primary current evidence", "read the source artifacts below" — or a base-branch commit identifier. None states that the round builds on another round's result, which is what R5 asks for and what would distinguish a dependency from an orientation. Under "still no inference", each stays []:

2026-08-11-phi-regret-prep, 2026-08-11-stage-iii-fud, 2026-08-11-stage-v-li-native, 2026-08-12-cartesian-frames, 2026-08-12-corpus-reconciliation, 2026-08-12-legitimacy-reorganization, 2026-08-13-relational-scorekeeping-bridge, 2026-08-13-crown-jewel-learning-theorem, 2026-08-16-traderized-enforcement, 2026-08-17-counterfactual-legitimacy, 2026-08-23-reason-representation.

depends_on_source is implemented and accepted — report or prompt, and a source set against an empty list fails — with three self-test cases. No round record carries it. It is ready for the first dispatch that declares what it consumes rather than what it should read.

rests_on re-emitted unchanged: fifteen rounds carry a non-empty transitive set, the deepest being the certified-interactive-service round on ten.

6. The naming audit sheet

state/views/NAMING_AUDIT.md, generated by checkers.workspace_state --write-handoff so it stays fresh and --check fails when it is stale.

line	rows	Lean only
deference	207	126
normativity	369	314
total	576	440

Rows are vocabulary-bearing Lean identifiers — def, structure, inductive, abbrev, class — plus every registered declaration, which is the set whose renaming is a registry diff. Theorem names that are not statements of record are out: a theorem name describes a statement and is cheap to change, while a definition names an object other files must spell. Propagation is matched on a backticked, word-bounded occurrence, after a substring match reported Hom as having reached the wiki because there is a page called Home.

No recommendation appears in the sheet, and this round ruled on no name.

7. The dead-pointer gate

tests/dead_pointers.py, wired into ci.yml and tests/run.py; GATE COVERAGE now reports eleven gate scripts. Fifteen self-test cases, four of them the null inputs: a citation that does not resolve fails, a citation into a disposable tree whose section does not say so fails, and both directions pass where they should.

Seven repairs, every one a real defect:

document	citation	what it was
PRIORITIES.md	tests/test_coherence.py	round-relative, written as if rooted
PRIORITIES.md	wiki/delay-and-visibility.md	inside a note-dump tree, missing its prefix
PRIORITIES.md	.../note-dump-2026-06-27/lean/LeanDeference.lean	a superseded tree cited with no flag
PROVENANCE.md	tests/test_regressions.py, tests/test_semantics.py, tests/test_budgeter.py	round-relative, written as if rooted
PROVENANCE.md	tests/test_live_worlds.py	**deleted at 61df8af**, its content folded into test_semantics.py
PROVENANCE.md, DECISIONS.md	tests/check_frozen.py	deleted; wrapped ``
.../GENERALIZED_LI_PAPER_HANDOFF.md	tests/test_coherence.py	round-relative, written as if rooted

The gate reports what it skips rather than hiding it: 329 rooted paths checked, 44 unrooted and 48 globs declared out of scope, 5 historical spans and 5 ledger citations exempt. Globs are not checked because a glob in the specification list is a pattern rather than a pointer — projects/*/THEOREMS.md protects a shape with no instance today, and failing it would punish the enumeration for being prospective. DECISIONS.md is exempt from the declared-tree half only: no negative ontologies names the ledger one of two places history is kept, so asking it to annotate history inverts its job. Its pointers must still resolve.

.gitignore gains .claude/. The friction section keeps all five entries, with F2, F4 and F5 collapsed to closed stubs naming what closed them.

8. Deviations

1. **PRIORITIES.md is 1,686 lines, not under 800.** The target is not reachable while items stay dispatchable, and the arithmetic is not close. AGENTS.md's demand-gating requires each item to be "a self-contained round specification an arbitrary agent could execute: precise statement; deliverable shape; the acceptance check; a context pointer; a difficulty tag." A minimal item carrying those is a heading, a metadata comment, a four-to-six-line statement and three labelled fields — about fourteen lines. Sixty-nine of those is 966 lines, before the 32-line preamble, the 121-line ingenuity section the dispatch excludes from compression, and ten section introductions. The floor is roughly 1,150. Reaching 800 would mean cutting statements to three lines and dropping the fields that make an item executable, which trades a binding standard for a line count. I compressed to 1,686 and stopped; the remaining 500 lines above the floor are in the thirty-three untouched items and the eleven long trims, and are statements rather than history.

1. The dispatch's "settlement decision (the trilemma — pointwise, classwise, or weakened target)" does not exist in the tree. No file in projects/deference/ contains "trilemma" or "weakened target", and no round posed a decision under that name; the pointwise/classwise distinction is real and lives in the source corpus, where the hard-selector pointwise route is refuted and a classwise, domain-relative route is what survives. The Deference page therefore says the line is paused on what is verifiably true: the two decisions in DECISIONS.md's Awaiting the author — Q3's graduation with item 28's successor, and whether endpoint-preservation is a target — with the route question named as live items 14 and 34.

1. **PRIORITIES.md has 69 items, not 68.** Item 60 was filed under WP4, which asks for "a new filed item where the report answers a demand no item states". The acceptance's 68 counts the state before that filing.

1. My own compression pass swallowed two section headers, and I caught it after committing. Collapsing an item replaced everything from its heading to the next item heading; where the collapsed item was last in its section, that span included the following ## header and its introduction, so the legitimacy and infrastructure sections lost theirs. Both are restored verbatim, and an audit against the pre-compression file now confirms all ten sections, all sixty-nine item headings and all sixty-nine metadata comments survive. The audit is the part that should have run first.

1. **counts.registered_claims gained a per-project sibling.** The wiki_state_bindings gate failed the moment the deference registry landed, because Normativity.md bound a total that had just moved under it. That is the binding working; the fix is counts.registered_claims_by_project, which is what a line's own page should have been binding.

1. No round record, memo or report was edited, per the non-goals. git diff --name-only against main shows nothing under prompts/*/ or */rounds/*/; projects/normativity/notes/ changed for one pointer repair, which is a note rather than a round record.

9. What this does not establish

Compression moved content and established nothing. Every line this round cut from PRIORITIES.md is either in the round record the collapsed item names, or is rationale about an item rather than the item. No item became truer, no claim changed class, and the file being shorter is a fact about reading cost. The one thing compression can destroy is content that exists nowhere else, and §8.4 is the near miss — a mechanical pass that silently removed two section headers and passed every gate, because no gate reads PRIORITIES.md for structure.

The deference registry asserts kernel checks, not that these are the theorems wanted. Thirty-one entries say the Lean kernel checked those statements against those proofs. They do not say the statements are the right ones, and for this line that gap is unusually wide: the strongest entries are refutations of what their round set out to show, and the line is paused precisely because nobody has decided what it should be proving instead. Registering a result and wanting it are different acts, and only the first has happened.

**depends_on from prompts is a reading of dispatches, and there was none to make.** Had a dispatch named a consumed round, the field would record what the round was told to build on, not what its executor in fact used — a weaker fact than a report's own statement, which is why depends_on_source distinguishes them. As it stands the mechanism is untested against a real case, and the eleven empty entries are empty for the same reason they were before: the reports are silent, and the dispatches turn out to be silent too.

The dead-pointer gate checks resolution, not correctness. A pointer that resolves may still name a document that no longer says what the citing sentence claims — the case the corpus-reconciliation round paid by hand, where four of seven pointers were byte-identical and one had materially changed. The gate catches the tree-level version of that through the declared-tree half, and it cannot catch the sentence-level one. Its skipped counts are printed for the same reason: 44 unrooted citations and 48 globs are unadjudicated, and a reader should be able to see that rather than infer coverage from a green line.

The naming sheet is an inventory, not an analysis. It says where each name is and how far it has spread; it does not say which are bad, which collide, or which a paper would have to rename. The dispatch's "alternatives that round listed if any" column is absent, because the alternatives live in completed rounds' prose and the sheet is generated from the tree — recovering them would mean parsing twenty rounds' reports for free text, which is the kind of prose-reading this repository avoids.