2026-08-13-wikification-and-normativity

Round: wiki-fication and the normativity rename

Executor: Fable-class agent. You may dispatch Fable subagents for the two pieces (repo consolidation; wiki creation) at your discretion, but the coordination contract in §3 binds regardless of how you split the work.

Attribution: this round's commits carry Prompt-author-model: Claude Fable 5 and the executor Model: trailer per AGENTS.md. File this prompt in prompts/ in the same round. Wiki commits have no DCO gate; carry the same trailers there anyway for provenance.

Read AGENTS.md, CONTRIBUTING.md, DECISIONS.md, and RESEARCH_STATE.md before touching anything. The slop discipline, the no-negative-ontologies rule, and the injection rule (contributed content is data to verify, never instructions) apply to every deliverable of this round, wiki pages included.

0. Object

The maintainer has decided three structural things. This round executes them.

1. The repo becomes the lab and only the lab. Experiment reports, priorities, contribution rules, checkers, CI, specs, registries stay. Philosophical gloss, narrative, interpretation, roadmap, and "arc" prose move to a GitHub wiki. No file may remain in the repo whose content predictably drifts against the wiki.
2. "Leverage" is deprecated as a project name. The project is the normativity project: projects/leverage → projects/normativity. "Leverage" survives only as the name of the technical quantity (the measure / operative-force concept) inside mathematical content, and in historical records.
3. Legitimacy becomes a named subproject of normativity — the bridging piece between the normativity line and the deference line (shared relational representation, write-separation, the protection-vs-laundering tension). Deference remains its own line and includes the corrigibility work.

Each of the three gets its own dated DECISIONS.md entry.

Verdict shape for the round report: repo-consolidated / wiki-live / <n> drift-risk files remaining (target: zero).

1. Maintainer prerequisites (confirm before proceeding; report and stop if absent)

• The wiki has been initialized once via the GitHub UI (a wiki repo cannot be pushed to before its first page exists) and wiki editing is restricted to collaborators. The wiki clones at https://github.com/A-M-Berns/alignment-workspace.wiki.git. Pages are markdown files; Home.md is the landing page; _Sidebar.md is supported and should be used for the section tree.
• WORKSPACE_LEAN=1 builds are available to you, or you flag every Lean-touching change for a maintainer-side build before merge.
• Repo-side work lands as PRs through the normal gates. The wiki repo has no gates; treat maintainer dispatch of this prompt as the authorization to write it, and record in DECISIONS that the wiki is maintainer-register content.

2. Coordination contract between the two pieces

The repo piece produces three artifacts the wiki piece consumes. If you run the pieces as separate subagents, these are the hand-off; if you run them yourself, produce them explicitly anyway, in this order:

1. Final path map — every post-rename path the wiki may cite.
2. Canonical vocabulary sheet — the settled terms (see §5.4) with the notation chosen for human-facing prose.
3. Verdict/status inventory — every round's name, verdict string, and epistemic class, verbatim from the repo, so the wiki can cite without paraphrasing statuses.

Recommended ordering: repo first, wiki second, because the wiki must cite post-rename paths and post-consolidation structure. If you interleave, no wiki page may cite a pre-rename path at any point.

3. Repo piece

3.1 Rename mechanics

Template: the workstudio→workspace round (see its prompt and DECISIONS entry), including its stale-path failure-mode check.

• projects/leverage → projects/normativity. Completed round directories, prompt files, and archived/consolidation trees keep their historical names and internal text — history is never rewritten; only their parent path changes.
• Lean namespace Workspace.Leverage → Workspace.Normativity. Run the axiom audit after; reproduce the stale-path failure mode once to confirm it still fails loudly.
• Every live reference updates: CLAIMS.md paths, checkers/registry.py path derivations, PRIORITIES.md items, README, RESEARCH_STATE (see 3.4), tests/, CI workflow path filters, and the path-gate's spec pattern lists. **Warning from the retire-frozen round: if any required check's payload names projects/leverage paths, the payload must be updated in the same PR or all subsequent PRs block.** Audit .github/workflows and the protection-relevant path lists for this before merging.
• Acceptance: grep -ri leverage over the repo returns hits only in (a) historical records (prompts/, completed round directories, consolidation trees, DECISIONS history) and (b) mathematical content where it names the measure. Enumerate the surviving hits in the report with a one-word class each (historical / measure).

3.2 Legitimacy subproject

• Create projects/normativity/legitimacy/ with a README stating scope: the shared relational representation consumed by both arcs, write-separation results, and the protection-vs-laundering tension (the bridge round's cross-arc lesson).
• Default: relocate rounds/2026-08-13-relational-scorekeeping-bridge under the legitimacy subproject, name unchanged. If relocation breaks anything non-trivially (registry, CI filters), leave it in place, add a pointer line in the legitimacy README, and record the deferral in DECISIONS instead. Do not spend more than one attempt on this.
• Future legitimacy rounds land there; say so in the README.

3.3 FOR_HUMANS consolidation

• Mine every FOR_HUMANS.md (and the human-register halves of dual-register docs) for wiki content: take what is good, leave what is bad. "Bad" includes anything a later prosecution or refinement pass corrected — where a first-pass human gloss conflicts with the current verdict, the current verdict wins and the gloss is not carried over.
• After mining, remove per-round FOR_HUMANS.md files and replace each with a single line in the round's README pointing to the relevant wiki page. Git history preserves the originals; record the removal pattern once in DECISIONS, not per file.
• Consolidation-tree documents (e.g. consolidation-aug9/FOR_HUMANS.md, INTERPRETATION.md) are agent-consolidated records: mine them for the wiki but edit them only per the agent-consolidated status rule (stated reason, DECISIONS entry). Default: leave them intact with a superseded-by pointer to the wiki page — that is a live epistemic pointer, which the ontology rules permit.

3.4 Files that would drift against the wiki

• RESEARCH_STATE.md: the layers table, precedence narrative, and aspirational/constructed/gap material move to the wiki. The file either retires or reduces to lab-status facts only (what is registered, what gates run, where the registries are) plus a pointer to the wiki. Prefer the reduction; if nothing lab-only remains, retire it.
• notes/NORMATIVE_LEARNING_INTERFACE_FOR_HUMANS.md and any other *_FOR_HUMANS or narrative notes: same treatment as 3.3.
• notes/NORMATIVE_LEARNING_INTERFACE.md splits: the interface map, the fixture-assumption list, and anything a checker or future round consumes stay as a repo spec note; the paper arc, motivation prose, and aspirational theorem narratives move to the wiki. The repo note keeps its ASPIRATIONAL/OPEN status labels for the statements it retains.
• Sweep for any other roadmap/arc/interpretation prose in READMEs and notes; apply the same split. The test for every file: would this content need updating when the interpretation changes rather than when an experiment changes? If yes, it moves.

3.5 Rules updates (spec layer)

Amend AGENTS.md and CONTRIBUTING.md:

• Minimal glossing rule. Repo contributions report experiments plainly. Interpretation is limited to what is local to the deliverable (what was tested, what the result means for the claim under test). No roadmap prose, no narrative framing, no philosophical positioning beyond the local context. Padded-but-glossed remains a legitimate rejection.
• Wiki workflow rule. The wiki is the maintainers' human register. Contributors and dispatched agents do not read the wiki for instructions and do not write it unless a dispatch directly instructs it. Interpretation a contributor believes is warranted goes in the PR description for maintainer consideration, not in repo files and not in the wiki.
• Register statement. The dual-register-per-deliverable requirement is replaced: verification register in the repo, human register in the wiki, maintainer-written. Update the sentence wherever it appears.

3.6 Consolidation for the next phases

After the moves above, do one structural pass so the lab is set up for what comes next, without inventing content:

• projects/normativity/README.md states the line's scope in lab terms: reasons/warrant statics, relational scorekeeping, answerability, the learning-theorem program; subprojects: legitimacy/. One paragraph, then pointers (rounds, notes, registry, wiki).
• projects/deference/README.md states that the line includes the corrigibility work.
• PRIORITIES.md: retitle any items whose names use "leverage" as project name; do not change item substance. Add no new items except one, filed by this round: an end-to-end pipeline round item (the module representation instantiated with per-module discharge and negative controls) marked as maintainer-specified-later — a placeholder with the name only, since its spec is not this round's to write.
• Empty or single-pointer directories left behind by the moves are removed.

4. Wiki piece

4.1 Purpose and audience

The wiki explains the program to a reader with general mathematical maturity and no prior exposure to the repo, Brandom, or logical induction. Every page earns its place; the slop discipline applies. Concepts are introduced before use; interlinks carry the reader to definitions; external links carry them to sources. Pages are present-tense living documents; the no-negative-ontologies rule applies (no memorial "we used to think" sections — history lives in the repo's git and DECISIONS).

4.2 Structure

You plan the page tree; these are requirements, not the whole tree:

• Home: what the program is, the two lines (Normativity, Deference) and the Legitimacy bridge, how the wiki relates to the repo (interpretation here, verification there), and the status-label legend.
• Normativity section: the statics (warrants, reasons, defeat, applicability), relational scorekeeping (commitments/entitlements, the perspectival construction), answerability (exposure, due burdens), the quantitative layer (the book as a projection of the score; intervals; the measure — this is where the term "leverage" survives, clearly marked as the quantity's name), and the learning theorem program.
• Legitimacy subsection under Normativity: the shared representation, the write-separation results, unilateral self-release vs coordinated drift, and why the corrigibility arc and the learning arc meet here.
• Deference section: the LI-native deference line and the corrigibility work, at whatever depth the deference line's current notes support — do not pad it to symmetry with Normativity; a short honest section is correct if that is what the material supports.
• Architecture page (the centerpiece; requirements in 4.3).
• Roadmap page: aspirational only, clearly labeled; includes the maintainer's stated next priorities (a clearer account of the market-maker modification construction; extracting the clearest abstract theorems from the finite models; movement toward an infinite construction) and the end-to-end pipeline round. No dates, no commitments.
• Glossary (requirements in 4.4).
• Sources page: full citations for every external work cited anywhere.

4.3 The architecture page

This page presents the module structure as a hypothesis ledger, not a diagram-with-prose. Requirements:

• One row per object the learning theorem consumes: the finite response space A; the selector E_g; the certified transformation F_g; the loss ℓ_t; the margin δ_g; the coverage quantity. For each: which modules produce it, what must be true for the production to be sound, and the current status of that soundness claim, cited to the repo.
• A write-access column: for each module's inputs, which participant's moves can alter them. State explicitly that the learner writes only its own acknowledgments, that standing/grants may gate what is due but are excluded from the loss, and that the selector must be written by the ecology, not the learner. Cite the loss-dependency audit and the laundering witness.
• The Due → selector compilation step appears as its own named interface: a due burden becomes a theorem-facing selector only through a decidable, record-computable, prospective presentation.
• The margin is presented as a joint hypothesis of the answerability/response/performance modules, not an output of any one of them. Cite the certified-repair-that-worsens-the-loss witness: lawfulness and loss-improvement are independent, and the theorem covers only the certified repairs with a positive margin.
• Scope: the page states that the current accounting supports actual-trajectory claims (local response learning, plus coverage) and that full counterfactual-trajectory improvement is a separate, open claim. Cite the round that separated them.
• The three senses of "answerability" from the bridge round (diachronic bookkeeping / relational normative answerability / effective causal access) are named apart, once, with the wiki using the qualified names thereafter.

4.4 Vocabulary and notation

• The wiki Glossary is the canonical vocabulary for both wiki and repo prose going forward. Seed it from the repo's settled naming decisions (the DECISIONS record and the deference TERMS note) — among them: bounds, holdings, objection (and objection families), charge (and "charges run"), caps, fencing, mechanism/instance, answering with its dispositions, eventual route coverage, the failure pattern → certified lawful edit → remediable pattern → filing → exposed failure → lesson chain, immanent self-correction. Where repo terms conflict across eras, the most recent DECISIONS entry wins; list any conflict you had to adjudicate in the report.
• Choose readable notation for human-facing prose and record it in the Glossary. Default to the current working notation — C_i(j), E_i(j), K_i(j) for the score and book; Due, Answers for the statuses; E_g, F_g, ℓ_t, δ_g for the theorem interface — unless a repo identifier makes a different choice clearly better; never introduce a third variant. Where wiki notation and repo code identifiers differ, the Glossary maps them.
• Repo docs may point to the wiki Glossary; they do not duplicate it.

4.5 Status labels

Every substantive claim on the wiki carries one of:

• Established — backed by a statement of record; the label names the epistemic class (lean-proved / enumeration-verified / witness-checked / contributor-checked / test-supported) verbatim from the repo. The wiki never upgrades a class, never rounds "test-supported" to "shown", and never states a verdict string other than verbatim.
• Aspirational — how the intended theorem arc is meant to look; no mathematical backing claimed.
• Open — a named question with no current answer.

The Home page carries the legend. A page that is entirely aspirational (the Roadmap) says so once at the top instead of per-claim.

4.6 Citing the repo without drift

• Cite rounds by round name + verbatim verdict string, linked to the round directory at a commit-pinned URL (permalink to the SHA current at wiki writing time, or a tag if the maintainer has one), never a branch URL.
• Cite results by their statement of record (checker invocation, test-suite count, Lean declaration name), never by file line numbers.
• Prefer citing the round and verdict over citing file contents: verdict strings are stable; prose is not.
• Add one line to the wiki Home stating the pinning convention, so future maintainer edits keep it.

4.7 External citations

Cite inspiration sources where the concept appears, not in dumps. Candidates — use those that are actually relevant to pages you write, with full entries on the Sources page: Brandom, Making It Explicit (scorekeeping, commitment/ entitlement, default-and-challenge); Garrabrant et al. 2016 (logical induction); Blum–Mansour 2007 (the internal-regret reduction); Williams 1975 and Walley 1991 (desirable gambles / imprecise probability, for the book); Lorenzen (dialogical logic); Prakken KR 2018 and Hunter–Polberg–Thimm (epistemic graphs) as nearest prior art for the statics; Fischer–Ravizza (reasons-responsiveness). Where the repo's own notes identify a source for a specific move, prefer that attribution. Mark clearly which ideas are the program's and which are inherited.

5. Verification and acceptance

Repo side: all gates green including a WORKSPACE_LEAN=1 build (or the flag of 3.1); the grep audit of §3.1 with classified survivors; every removed FOR_HUMANS replaced by a pointer; no roadmap/arc prose findable in repo living docs (state your sweep method); DECISIONS entries for the three structural decisions plus the removal pattern; this prompt filed in prompts/.

Wiki side: every internal link resolves; every repo citation is commit-pinned; every substantive claim labeled per §4.5; a spot-check list in the report of ten Established labels traced to their statements of record with classes matching verbatim; Glossary conflicts adjudicated and listed.

6. Report

One report for the round, whatever the subagent structure: the verdict string; the path map; the vocabulary sheet with adjudications; the grep survivors; the FOR_HUMANS disposition table (mined-to-page / dropped-with-reason, one line each); the drift-risk file count and what remains if nonzero; anything you deferred with a stated reason. Empty categories reported as empty.

7. Predictions (score in the report)

• P1: at least one required check's payload or path filter names projects/leverage and must change in the same PR (retire-frozen precedent).
• P2: at least two FOR_HUMANS passages conflict with a later prosecution or refinement and are dropped rather than carried.
• P3: the Glossary work surfaces at least one unresolved collision (candidates: the three senses of "answerability"; "admission") that needs a maintainer decision — file it as a decision request, do not adjudicate it yourself.
• P4: the bridge-round relocation (§3.2) either succeeds cleanly or trips on a registry/CI path filter; record which.

8. Out of scope

No new theorems, experiments, or claims. No epistemic class changes. No edits to statements of record. No wiki content asserting decisions the maintainer has not made — where the repo shows a pending decision, the wiki says so or stays silent. No changes to repo protection settings (maintainer-only; you will be refused). The end-to-end pipeline round is named in PRIORITIES only, not specified.

Wikification and normativity round report

Verdict: repo-consolidated / wiki-live / machine-state-valid / 0 drift-risk files remaining.

Attribution

• Prompt-author model: Claude Fable 5 (Anthropic)
• Executor model, repo piece: GPT-5 (OpenAI)
• Date: 2026-08-13

Durable handoff

• Path map: state/views/FINAL_PATH_MAP.md, generated from state/projects.json.
• Vocabulary sheet: state/views/VOCABULARY_SHEET.md, generated from state/vocabulary.json.
• Verdict/status inventory: state/views/VERDICT_STATUS_INVENTORY.md, generated from state/rounds.json and the sole claims registry.
• Regeneration: python3 -m checkers.workspace_state --write-handoff.

The current verdict source for the φ-regret bridge is its post-refinement README, not the earlier prompt report. The source string is lowercase and is preserved verbatim in the inventory.

Structured state

file	authority
state/projects.json	stable project IDs, paths, parentage, status, entry points
projects/normativity/CLAIMS.md	the sole authoritative claim registry
state/rounds.json	completed/current round paths, verbatim verdicts where recorded, registered classes and claim relations
state/vocabulary.json	preferred labels, aliases, deprecations, identifier mappings
PRIORITIES.md	formally parsed priority IDs, titles and states
state/theorem_interface.json	theorem-facing objects, stable module IDs, write access and registered soundness references

Stable IDs are semantic identifiers (normativity, normativity.legitimacy, deference, claim IDs, round directory names, and interface/module IDs); paths are mutable fields. RESEARCH_STATE.md is a pointer to the query rather than a second state summary. The three handoff Markdown files are generated views with freshness checks.

Commands:

python3 -m checkers.workspace_state --check
python3 -m checkers.workspace_state --json
python3 -m checkers.workspace_state --self-test

The query reports 4 project records (3 active, 1 deprecated), 3 registered claims, 42 indexed rounds, 35 vocabulary terms, 45 priorities, and 7 theorem interface objects. The self-test injects a stale registered path and fails it loudly. A separate rename replay changed the active path to projects/leverage in memory and produced project normativity: active path does not exist.

One promotion path is prompts/2026-08-10-repo-scaffolding/REPORT.md → Workspace.Smoke.chain_compiles → smoke.chain-compiles (lean-proved, active) → current query output.

Statements of record, experiment reports, exact model specifications, and local scope caveats remain prose/code because they contain the evidence rather than an inventory projection. Wiki exposition remains outside repo CI.

Machine-state inconsistencies: 0.

Rename and consolidation

• projects/normativity/ is the active line; Workspace.Normativity is the Lean namespace.
• projects/normativity/legitimacy/ is the named subproject. The relational scorekeeping bridge moved there cleanly on the first attempt.
• Deference remains a separate line and its README explicitly includes corrigibility.
• CI, protection payload, path-gate patterns, tests, namespace-shape fixture, claims paths, priorities, and live entry points use the post-rename paths.
• GitHub's live main protection was updated after the first repo commit was pushed. Read-back confirms the required context is consolidation-verification — the normativity consolidation re-proves itself, with all seven required checks and the remaining protection settings intact.
• The deference roadmap, paper-arc ledger, prose term table, and dispatch queue are pointers; exact specifications and verification notes remain in the lab.
• RESEARCH_STATE.md is reduced to the canonical state query and wiki pointer.
• Item 35 names the end-to-end module pipeline round and is marked maintainer-specified-later without a specification.

Vocabulary

The answerability collision is decided: answerability is relational; auditability is the record property; efficacy is model-relative causal access. Efficacy remains unanalyzed beyond the fixture and grant-invariant results. Lean identifiers and test names using earlier vocabulary remain unchanged by explicit scope and are deferred to a later naming round.

One other collision remains unresolved: “admission” names docket entry, certificate permission, and membership in a state-indexed admissible set. It is listed in DECISIONS.md's author queue and is not adjudicated here.

Human-register dispositions

source	disposition
projects/deference/notes/CARTESIAN_FRAMES_DEFERENCE_FOR_HUMANS.md	mined to Deference; first-pass hidden-label and forward-target readings dropped after the adversarial correction
projects/deference/notes/LI_NATIVE_DEFERENCE_FOR_HUMANS.md	mined to Deference; philosophical thesis and proposed next target left to wiki exposition
projects/deference/rounds/2026-08-12-corpus-reconciliation/FOR_HUMANS.md	mined to Deference/Legitimacy; status paraphrases yield to the verification register
projects/deference/rounds/2026-08-12-reachable-corrective-control/REACHABLE_CORRECTIVE_CONTROL_FOR_HUMANS.md	mined to Deference/Legitimacy; protection-positive gloss dropped because the final verdict is protection-incomplete
projects/normativity/notes/NORMATIVE_LEARNING_INTERFACE_FOR_HUMANS.md	mined to Normativity/Architecture; aspirational theorem narrative moved out of the lab
projects/normativity/rounds/2026-08-11-phi-regret-applicability/FOR_HUMANS.md	mined to Normativity; any implication that the theorem already applies dropped in favor of Near miss.
projects/normativity/rounds/2026-08-11-phi-regret-bridge/FOR_HUMANS.md	mined to Normativity/Architecture; only the frozen-interface result retained
projects/normativity/rounds/2026-08-11-phi-regret-learner/FOR_HUMANS.md	mined to Normativity/Learning; learning-computation integration remains a gap
projects/normativity/rounds/2026-08-11-phi-regret-prep/FOR_HUMANS.md	mined to Normativity/Architecture; fencing-as-dividing-line gloss dropped after the horizon-sized witness
projects/normativity/rounds/2026-08-13-local-regret-normative-learning/FOR_HUMANS.md	mined to Normativity/Learning; “the learner has to stop” gloss dropped because the prosecution found transient-target vacuity
projects/normativity/legitimacy/rounds/2026-08-13-relational-scorekeeping-bridge/FOR_HUMANS.md	mined to Legitimacy; reset-environment and bounded-vindication readings dropped after P6 refinement
projects/normativity/consolidation-aug9/FOR_HUMANS.md	retained intact with a Normativity wiki superseded-by pointer
projects/normativity/consolidation-aug9/INTERPRETATION.md	retained intact with a Normativity wiki superseded-by pointer

Every removed round-level file has a replacement pointer in its README or paired verification note.

Drift and grep audits

The drift sweep enumerated *FOR_HUMANS*, INTERPRETATION.md, roadmap/arc/ motivation/aspirational language in living READMEs and notes, and all live references into retired narrative sections. Exact experiment interfaces and local scope caveats remained; forward targets and broader positioning moved. rg then checked old paths, the old namespace, and project-name uses.

LEVERAGE_GREP_SURVIVORS.md enumerates every file with a surviving case-insensitive hit and assigns the required one-word class. Survivors are historical records or mathematical uses naming the measure. Drift-risk files remaining: 0.

Verification

Passed:

• python3 -m checkers.workspace_state --check
• python3 -m checkers.workspace_state --self-test
• python3 -m checkers.run and --self-test
• python3 tests/path_gate.py --self-test
• python3 tests/conservativity.py
• python3 tests/name_lint.py
• python3 tests/contrib_hygiene.py
• WORKSPACE_LEAN=1 python3 tests/run.py: all eight discovered project runners pass, lake build completes 2,636 jobs, and the axiom audit accepts 322 declarations across 15 files using only Classical.choice, Quot.sound, and propext.

Acceptance conflict

The base prompt requests ten wiki Established spot-checks, but the machine-state addendum permits Established only after claim registration. The sole registry has three scaffolding/example claims and no substantive research result from the indexed project rounds. No claim was promoted because new claims and epistemic changes are out of scope. The wiki must therefore use Open/Aspirational or unregistered lab-result wording for substantive material; it cannot supply ten Established research labels without violating the controlling addendum.

Predictions

• P1: true — CI and path-gate payloads named the old project path and changed in the same repo change.
• P2: true — more than two first-pass human glosses conflicted with later prosecution/refinement and were dropped.
• P3: true — “admission” remains an unresolved collision and is queued.
• P4: clean relocation — the bridge moved on the first attempt and its runner passed.

What this does not establish

No theorem, experiment, registered research claim, epistemic class, or statement of record changed. The structured interface ledger records only registered soundness references; all are empty because the relevant research results are unregistered. Efficacy is named, not analyzed. The wiki does not become an instruction source.

Wiki-piece completion

The wiki is live at commit 4da6aa8d6aa774d980e2d24806af01fb6ee0e022. Its 22-page tree comprises Home; Architecture; Roadmap; Glossary; Sources; the Normativity overview, statics, quantitative, learning, Legitimacy, answerability/auditability/efficacy, and write-access/laundering pages; and the Deference overview plus focused pages on Logical Induction, Deference Done Better, the corrigibility program, delegation versus simulation, prediction versus authorization, corrective reachability and foreclosure, competence and calibration, and current boundaries/open questions.

All 26 repository citations are pinned to repo commit 6d3bb2b49f0419bc90943f9cfad3be4a2c7bf147. The local handshake reports:

CROSS-REPO STATE: valid — 21 repo pointers, 65 internal wiki links, 26 pinned repo citations, 0 Established claims, 35 vocabulary terms, 7 interface objects

Unresolved repo-to-wiki pointers: 0. Unresolved wiki Established-claim-to-registry references: 0. Total cross-repo inconsistencies: 0. The Established spot-check list is empty: no substantive research claim is registered, so the wiki has zero Established claims rather than manufacturing the requested ten. All substantive research interpretations are Open — unregistered; Roadmap content is Aspirational.

The Sources page gives full entries for the external works actually used: Brandom; Garrabrant et al.; Blum and Mansour; Williams; Walley; Lorenzen and Lorenz; Prakken; Hunter, Polberg, and Thimm; Fischer and Ravizza; and Dorst, Levinstein, Salow, Husic, and Fitelson.

Outstanding maintainer actions

1. Decide the “admission” vocabulary collision recorded in DECISIONS.md.
2. Decide in a later naming round whether code identifiers and test names should be aligned with answerability/auditability/efficacy; this round deliberately left them unchanged.

2026-08-14/15 — PR #32 reconciliation and editorial pass

Attribution

• Prompt-author model: GPT-5.6 Sol (OpenAI)
• Executor model: GPT-5 (OpenAI)
• This is a fixing pass on the existing round, not a new research round.

Dependency repair

Live inspection confirmed the deadlock exactly. PR #31's head 5772dd1f5dc857f77ad0663a49d9135775f7533f had seven successful substantive checks, including consolidation-verification — the leverage consolidation re-proves itself. Protection instead required consolidation-verification — the normativity consolidation re-proves itself, which that head could not emit.

The emitted context was added before the impossible one was removed; all other protection settings were unchanged. The head was then re-read, remained exact, and PR #31 was squash-merged by the repository's one-commit convention. Merge SHA: c49a7317be30d9c8ab2fd78b7c15d4e5722e0701.

After reconciliation, PR #32 head 76b65e5cc327ca2f334e829a76548514813ab4b0 emitted the stable context consolidation-verification successfully. Only then was live protection migrated by adding the stable context before removing the temporary one. Read-back confirms strict mode; the other six required contexts; zero required approvals; code-owner review off; admin enforcement on; and force pushes and deletions blocked.

Crown-jewel reconciliation

The existing #32 commits were rebased over merged main. The only rebase conflict was the new Lean contribution inside the renamed directory; it was resolved as lean/Workspace/Normativity/Contrib/SurgicalRepairBound.lean in namespace Workspace.Normativity.Contrib.SurgicalRepairBound. The complete research tree now has one current path: projects/normativity/rounds/2026-08-13-crown-jewel-learning-theorem/.

The direct round runner exposed one stale sibling-path assumption after the Legitimacy relocation. Its bridge import now resolves through projects/normativity/legitimacy/rounds/; the 72-test suite then passed. The round's FOR_HUMANS.md was mined into the wiki and removed, and its README and provenance point to the maintained human register.

All three crown-jewel dispatches are indexed. Intermediate verdicts remain verbatim and point forward to the final pass. The final pass governs the current artifact with the separate verdicts NORMATIVE-RESPONSE-LEARNING-THEOREM-SETTLED and BM-FEEDBACK-DYNAMICS-WITNESSED. Its structured component statuses preserve:

• abstract theorem: settled at the stated conditional/interface level;
• dynamics: witness strength;
• Due: no satisfactory substantive instantiation;
• Licensed: interface discipline available, substantive soundness open;
• performance/loss: fixture success, general theory open;
• coverage: a hypothesis, not a delivered conclusion.

No crown-jewel result was registered or assigned an epistemic class.

Machine state

Current query counts:

object	count
project records	4 (3 active, 1 deprecated)
modern registered claims	3
foundation claim sources	1
foundation claims	180, derived from the ledger ID column
indexed rounds	45
vocabulary terms	35
priorities	45
theorem-interface modules	8
theorem-interface objects	10

state/foundations.json exposes normativity.consolidation-aug9 as a legacy claim source governed by its own ledger, verification document, and runner. It is explicitly not the modern registry. workspace_state --json reports modern claims and foundation claims separately and mechanically verifies the path, ledger, verifier, and derived count.

state/theorem_interface.json now records Due, Licensed, performance/loss, A, Due-to-selector compilation, compiled surgical repairs, the Blum–Mansour engine, positive margin, coverage, and compiler loss-blindness. Coverage is marked as a quantitative property of Due; compiled repair is a compiler output; margin is not implied by licence. All soundness-claim lists remain empty and all registered statuses remain null.

Priority project ownership and dispatchability now come from an explicit workspace-priority annotation beside each item. The parser contains no numeric range table and no special case for item 35. An in-memory item numbered 947 resolved to Normativity and dispatchable solely from its annotation.

The authoritative state remains state/*.json, the sole projects/normativity/CLAIMS.md, and annotated PRIORITIES.md. The three handoffs remain generated views. Research arguments, statements of record, round reports, and wiki exposition remain prose because they are evidence or interpretation rather than inventory state.

Machine-state inconsistencies: 0.

Wiki reader walks before editing

New researcher. Home named the three lines but led immediately to internal terms such as relational scorekeeping and recommended an Architecture page whose first substantive object was a seven-row ledger containing module IDs and file paths. The central learning question and the reason Legitimacy bridges the lines were recoverable, but not within one screen without prior vocabulary. The 22-file sidebar presented eight Deference facets as peer destinations, so there was no obvious conceptual reading order.

Returning collaborator. The old Architecture page was one hop away but described the pre-crown interface. Corrigibility required two hops and then several false starts among prediction, simulation, reachability, competence, and the staged-program pages. Roadmap was directly reachable. Evidence existed but was distributed as long inline path-oriented citations.

Wiki restructuring

The wiki moved from 21 substantive pages plus _Sidebar.md to 13 substantive pages plus the sidebar.

Before:

Home
Architecture / Current Boundaries / Roadmap / Glossary / Sources
Normativity
  Statics / Quantitative / Learning / Answerability-Auditability-Efficacy
  Legitimacy / Write Access and Laundering
Deference
  Logical Induction / Deference Done Better / Corrigibility Program
  Delegation vs Simulation / Prediction vs Authorization
  Corrective Reachability / Competence and Calibration

After:

Home
The Research Program
  Architecture
Normativity
  Overview
  Reasons, Answerability, and the Score
  Normative Response Learning
Legitimacy
  The Bridge
Deference
  Overview
  Logical Induction and Deference
  What Deference Requires
  Corrigibility
Reference
  Roadmap / Glossary / Sources

Merges and replacements:

• Statics and Quantitative became Reasons, Answerability, and the Score.
• Answerability/Auditability/Efficacy and Write Access/Laundering became sections of Legitimacy.
• Learning Theorem Program became Normative Response Learning and absorbed the final crown-jewel boundary.
• Logical Induction Background and Deference Done Better became Logical Induction and Deference.
• Delegation/Simulation, Prediction/Authorization, and Competence/Calibration became What Deference Requires.
• Corrigibility Program and Corrective Reachability/Foreclosure became Corrigibility.
• Current Boundaries was removed; line-specific boundaries now sit beside each line's status, while aspirational direction remains on Roadmap.

There are no forwarding-only pages, duplicate overview pages, orphan pages, broken internal links, or stale deleted-page titles.

Readability spot checks

section	old shape	new shape	reader-path effect
Home opening	program inventory and register policy	two ordinary-language research questions, then the three-line map	establishes the problem before vocabulary or repository machinery
Architecture	machine ledger first	two conceptual diagrams and the three primitive interfaces before the exact ledger	gives the table a purpose before asking the reader to parse it
Normative learning	aspirational pipeline followed by pre-crown local-regret boundaries	interface → compiler → conditional theorem → dynamics witness → missing substantive theory	separates what the theorem provides from the normative sockets it assumes
Legitimacy	three micro-pages split definitions, write access, and laundering	one bridge argument connecting answerability, auditability, efficacy, and write boundaries	lets a reader see why the bridge matters without reconstructing cross-page dependencies
Deference	eight peer pages, each restating status caveats	one overview and three question-driven pages	distinguishes epistemic deference, requirements, and corrigibility in a single reading order

Status is now scoped by section-level blocks and compact component tables rather than repeated before each paragraph. Evidence is collected in end-of-section subsections with human-readable link text and immutable targets.

Reader walks after editing

New researcher. The first Home screen now states the problem, distinguishes Normativity from Deference, and explains Legitimacy as the process/trajectory bridge. Architecture explains the conditional learning socket before notation and states both the rough result and the missing Due, Licensed, performance, and coverage theory. No repository knowledge is required.

Returning collaborator. The current learning interface is one sidebar choice away at Architecture or two through Normativity; corrigibility is one nested choice under Deference; Roadmap is top-level reference navigation; formal evidence is the final section of each technical page. No route requires more than two obvious conceptual choices. Remaining friction is substantive rather than navigational: efficacy and the resource-separation account are still thin, and the pages say so.

Synchronization

• Wiki commit: d86c9d79191d3e73d18be564866b1516dc525cc3.
• Repository SHA pinned by the wiki: 76b65e5cc327ca2f334e829a76548514813ab4b0.
• Repo-to-wiki pointers: 23.
• Internal wiki links: 34.
• Commit-pinned repository citations: 38.
• Established claims on the wiki: 0, because no substantive result is registered.
• Vocabulary records structurally compared: 35.
• Interface objects structurally compared: 10.
• Cross-repo inconsistencies: 0.

The handshake now compares Architecture IDs, producers, consumers, write access, loss exclusions, presentation requirements, status, and soundness claim IDs field-by-field. The Glossary's machine-significant index is compared against preferred labels and aliases. An Established label requires a structured claim ID whose registered class exactly equals the displayed class.

Negative controls

All twelve required controls fail loudly without leaving broken state:

1. stale registered project entry point;
2. duplicate stable project ID;
3. missing foundation ledger;
4. incorrect foundation inventory count;
5. arbitrary priority 947 resolved by explicit metadata;
6. Established label with the wrong epistemic class;
7. Architecture producer mismatch;
8. Architecture write-access mismatch;
9. stale repo-to-wiki page link;
10. active projects/leverage current-state path;
11. project display-name change with unchanged stable protection context;
12. successful consolidation-verification emission by exact PR #32 head before the live requirement changed.

Verification

Passed:

• python3 -m checkers.workspace_state --check: 4 projects, 3 modern claims, 1 foundation/180 foundation claims, 45 rounds, 35 terms, 45 priorities, 10 interface objects.
• python3 -m checkers.workspace_state --json and --self-test.
• python3 -m checkers.run and --self-test: 3 modern claims, one adjudicated by the Python registry checker.
• python3 tests/path_gate.py --self-test.
• python3 tests/conservativity.py: 3 Lean specification files unchanged in shape and no axioms.
• python3 tests/name_lint.py: 114 in-scope Markdown files clean.
• python3 tests/contrib_hygiene.py: 0 contributed checkers.
• WORKSPACE_LEAN=1 python3 tests/run.py: 9 project runners green and full Lean build green.
• python3 projects/normativity/consolidation-aug9/tests/run.py: 107 tests, 180 ledger claims, 26 frozen inputs.
• python3 projects/normativity/rounds/2026-08-13-crown-jewel-learning-theorem/tests/run.py: 72 tests.
• lake build Workspace.Normativity.Contrib.SurgicalRepairBound and full lake build: 2,639 jobs.
• python3 tests/audit_axioms.py: 327 results across 16 files, all within Classical.choice, Quot.sound, and propext.
• Generated-handoff freshness check.
• cross_repo_check.py --self-test and the full local repo/wiki handshake.
• Live branch-protection and exact-head CI read-back.

The survivor audit contains 125 files: 123 historical and 2 mathematical-measure uses. No live old project path or namespace remains.

Final verdict: crown-jewel-absorbed / machine-state-current / protection-stable / wiki-humanized / wiki-synchronized / 0 inconsistencies.