2026-08-11-workspace-ethos

2026-08-11 · project workspace · path prompts/2026-08-11-workspace-ethos

Prompt prompts/2026-08-11-workspace-ethos/PROMPT.md

Workspace Ethos Pass — Consolidation Layers, Aspirational/Constructed Status, and Research Debt

Maintainer: A. M. Berns Prompt-author-model: GPT-5.6 Sol (OpenAI)

Mission

Refine the workspace’s research-state architecture so that it remains legible under substantial AI-assisted research without becoming procedurally bloated.

The guiding problem is:

How do we let agents explore aggressively while keeping it cheap to tell what the research program currently aspires to, what it has actually constructed, what remains unresolved, and what has received human judgment?

This is primarily a workspace framing / information-architecture pass, not an object-level research round.

It may run in parallel with active deference/corrigibility research.

Accordingly:

[ \boxed{ \text{do not redesign active research objects underneath ongoing work.} } ]

Prefer small, additive, repo-conformant changes. Do not impose elaborate workflow or mandatory ceremony.

0. First inspect the live workspace

Do not assume this prompt knows the exact current repository structure.

Read the live repository, including as applicable:

root and nested AGENTS.md;
README;
DECISIONS.md;
PROVENANCE;
PRIORITIES.md;
project roadmaps and ledgers;
current consolidation reports;
research-track reports;
canonical specification documents;
current treatment of archived/frozen/superseded work;
any existing review, registration, status, or provenance conventions.

Use the deference/corrigibility project as the main case study because it has undergone several research/consolidation rounds, but design the resulting conventions to be usable across the workspace, including leverage/normative-learning work and future projects.

Before implementing anything, identify what the repository already does well and reuse it.

Do not create parallel machinery for distinctions that are already represented adequately.

1. Governing design principle: discipline without procedural bloat

The workspace is a research lab first.

Adopt this principle:

[ \boxed{ \text{structure should reduce ambiguity and cognitive load, not create ceremony.} } ]

Avoid a system where ordinary research requires:

formal amendment objects;
multi-step approvals;
mandatory status metadata on every note;
redundant registries;
constant manual synchronization;
extensive form-filling before an agent can investigate an idea.

Heavyweight treatment is justified only where coordination cost is actually high—for example:

changing a shared formal skeleton;
changing canonical ontology;
altering trust/build/governance machinery;
registering a major result;
changing something on which many active artifacts depend.

Ordinary research should remain ordinary Git research.

Prefer conventions over mechanisms where conventions suffice.

Prefer one useful file over three registries.

Prefer a compact human-readable table over a schema unless machine structure provides clear immediate value.

2. Establish three research-state layers

The workspace should clearly distinguish:

[ \boxed{ \text{Research Lab} \longrightarrow \text{Agent-Consolidated State} \longrightarrow \text{Human-Canonical State} } ]

Implement this distinction in the lightest useful way.

2.1 Research Lab

Purpose:

maximize exploration.

This includes:

raw research tracks;
speculative theorem shapes;
negative attempts;
finite harnesses;
agent reports;
alternative models;
prompts;
provisional vocabulary;
failed approaches.

The lab may be messy.

Lab artifacts are evidence/history, not automatically the current research position.

2.2 Agent-Consolidated State

Purpose:

maximize coherence, compression, and recoverability.

An agent-consolidated artifact represents:

the best current machine-produced synthesis of the research evidence available in the workspace.

It may:

reconcile several tracks;
identify the current best object;
mark superseded approaches;
compress duplicated terminology;
classify open debts;
distinguish aspiration from construction;
propose canonical changes;
identify the controlling next question.

It is not automatically what the maintainer endorses.

Make this status explicit.

Do not call agent consolidation “canonical.”

2.3 Human-Canonical State

Purpose:

maximize clarity, judgment, and stability.

This layer should remain deliberately small.

It represents the research position that has received human/maintainer judgment.

It may contain:

current core questions;
current definitions;
constructed mathematical results that matter;
current philosophical interpretation;
important impossibility/fence results;
accepted assumptions;
acknowledged debts;
controlling research direction.

Important:

[ \boxed{ \text{agents may propose changes to the canonical state but may not silently confer canonical status.} } ]

Do not infer maintainer endorsement merely because:

a theorem is Lean-verified;
several agents agree;
a parent report recommends it;
it is newer than the existing canonical statement.

If existing repository conventions already have a suitable mechanism for maintainer adoption, use them.

Do not invent an elaborate new approval process.

3. Define precedence for interpreting the workspace

The intended epistemic precedence should be clear enough for both fresh agents and a future public chatbot.

Default current-state interpretation should approximately be:

[ \boxed{ \text{Human-Canonical}

\text{Agent-Consolidated}

\text{Research Lab} } ]

But this is a precedence for representing the current program, not a truth hierarchy.

A lower-level artifact can challenge a higher-level claim.

For example, the workspace must be able to represent:

“The canonical view currently says C, but a new unresolved research result purports to refute C and has not yet been consolidated or reviewed.”

Do not hide conflicts merely because they have not reached the canonical layer.

Historical artifacts should remain retrievable without having equal authority to current artifacts.

4. Introduce the Aspirational / Constructed / Gap distinction

This should become one of the central lightweight framing conventions.

For an important research object, distinguish:

[ \boxed{ \text{Aspirational} \qquad \text{Constructed} \qquad \text{Gap} } ]

Do this on two registers.

4.1 Mathematical register

Aspirational mathematical state

What theorem, object, mechanism, or construction would constitute the hoped-for technical result?

Constructed mathematical state

What has actually been defined, proved, implemented, verified, or formally built?

Mathematical gap

What precisely prevents the constructed state from reaching the aspiration?

4.2 Philosophical register

Aspirational philosophical gloss

What conceptual claim would the intended result vindicate if the full technical program succeeded?

Constructed philosophical gloss

What does the existing mathematics actually warrant saying now?

Interpretive gap

What further mathematics or argument would be required before the aspirational gloss is justified?

This distinction should make it difficult for an agent or reader to confuse:

[ \text{“this is the theorem we want”} ]

with:

[ \text{“this is a theorem we have.”} ]

Likewise:

[ \text{“this is the conceptual picture motivating the work”} ]

with:

[ \text{“the current mathematics establishes this conceptual picture.”} ]

5. Prefer a lightweight recurring research-state format

Do not migrate every existing file mechanically.

Instead, propose and where useful implement a compact convention for canonical and agent-consolidated project summaries.

A good default shape is:

Question

What are we trying to understand?

Aspirational mathematical picture

What would technical success look like?

Constructed mathematical state

What exists now?

Aspirational philosophical gloss

What would the hoped-for result mean?

Constructed philosophical gloss

What does the current work genuinely support?

Fences / negative knowledge

What attractive approaches or stronger claims have been ruled out?

Research debt

What kinds of unresolved work remain?

Controlling gap

What currently blocks the next major advance?

Next question

What should the research attack next?

Do not require every small research note to use this structure.

It is principally for consolidation/frontier/canonical views.

6. Research debt accounting

“Open problem” is too coarse.

Introduce a lightweight vocabulary for different kinds of research debt.

Use only categories that prove useful; do not force every project to populate every category.

Candidate categories:

Model debt

The right mathematical object is not yet adequately defined.

Example shape:

the comparator itself is not yet well-posed.

Theorem debt

The object and statement are sufficiently clear, but the proof is missing.

Assumption debt

A positive result depends on an antecedent whose strength, necessity, or interpretation remains unresolved.

Interface debt

Two coherent components do not yet compose cleanly.

Formalization debt

A mathematical argument exists but has not been promoted into the desired formal substrate.

Verification debt

An artifact exists but lacks the intended independent checking/audit.

Interpretation debt

The mathematics is reasonably clear but its correct philosophical or research significance remains unsettled.

Scope debt

An important omitted phenomenon has been identified, but it is not yet clear whether it should be solved, modeled, or explicitly fenced out.

Compression debt

The live research state has become too complicated to recover cheaply from the canonical/consolidated layer.

This last category is particularly important.

Do not assign numeric debt scores unless the repository already has a useful reason to do so.

The point is classification, not pseudo-precision.

7. Debt can change type

Make explicit that reclassifying debt can itself be genuine research progress.

For example:

[ \text{apparent theorem debt} \rightarrow \text{assumption debt} ]

or:

[ \text{“prove theorem X”} \rightarrow \text{model/interface debt first}. ]

A failed proof round that reveals the target is ill-posed is not merely “no progress.”

The consolidated state should make such transitions visible.

8. Consolidation discipline

The workspace should encourage:

[ \boxed{ \text{explore freely for a bounded interval, then compress aggressively.} } ]

Do not impose a rigid clock.

Instead identify soft consolidation triggers.

A consolidation pass is probably due when:

several agents introduce competing names for the same object;
a result changes the interpretation of multiple earlier artifacts;
live conceptual forks are becoming difficult to enumerate;
fresh agents require large amounts of historical material to understand the present;
agents begin treating superseded reports as current;
the initialization prompt required to state the live research state becomes very large;
the roadmap accumulates layers of exceptions and caveats;
active priorities contain dead or already-resolved routes;
the distinction between aspiration and construction has become blurry.

A strong consolidation success criterion is:

Can a fresh strong agent recover the live research state from the agent-consolidated/canonical layer without reading the full history?

If not, there is compression debt.

9. Consolidation should include garbage collection

A consolidation pass should not merely create one more summary.

It should reduce the live conceptual working set.

Where safe and repository-conformant, consolidation should:

mark superseded definitions clearly;
retire dead routes from active priorities;
collapse duplicate terminology;
identify the current definition owner;
distinguish active from historical questions;
remove stale claims from default current-state views;
compress several exploratory distinctions into the few that survived;
preserve the historical artifacts rather than deleting evidence.

Desired principle:

[ \boxed{ \text{history may remain lossless; the canonical layer should be aggressively compressed.} } ]

Do not rewrite old reports to pretend the research always had its current form.

10. Cheap deprecation

Implement or document a lightweight convention for superseded material.

A minimal form may be enough:

SUPERSEDED BY: ...
STATUS: historical
REASON: ...

or whatever fits existing repo style.

The aim is that a future agent or chatbot can retrieve old ideas without treating them as equally current.

Do not build an elaborate deprecation registry unless the repository clearly needs one.

11. Definition and symbol stability

Consider a very small current-definition/symbol surface for research lines that need it.

Its purpose is to prevent conceptual drift in objects such as:

agents/processes;
policy classes;
execution objects;
quantities;
grades;
settlement;
jurisdiction;
specialized project notation.

The useful information is:

current symbol/name;
current meaning;
where the definition is owned;
whether it is canonical, agent-consolidated, or provisional.

Keep this compact.

Do not create an ontology-management bureaucracy.

12. Module interfaces

Where the repo already has genuinely modular research components, make their interfaces easier to see.

An interface description should answer:

What does this module supply downstream?

and:

What does it explicitly not supply?

For example, a module might supply:

calibration;
a competence condition;
protected execution semantics;
an uncertainty quantity;

without being allowed to silently supply:

moral legitimacy;
jurisdiction;
prediction accuracy;
choice optimality.

Use the deference project to determine whether such interface descriptions would materially reduce confusion.

If yes, add them in the lightest appropriate place.

Do not force every project into formal module contracts.

13. Human review surfaces

Agent consolidation should reduce the amount of material a human must inspect.

At the end of a substantial phase, the consolidated report should be able to identify a short list such as:

Human review surfaces

new shared definition;
important theorem statement;
substantive assumption;
major negative result;
philosophical reinterpretation;
proposed canonical delta.

The goal is:

[ \boxed{ \text{human judgment over compressed consequential changes, not exhaustive rereading of agent output.} } ]

Where possible, distinguish:

Proposed canonical delta

from the full consolidation report.

This can simply be a small section, not a new workflow system.

14. Lab → Consolidation → Brief

Evaluate whether the current repo would benefit from making this architecture more legible:

[ \boxed{ \text{LAB} \longrightarrow \text{AGENT-CONSOLIDATED STATE} \longrightarrow \text{HUMAN-CANONICAL / BRIEF} } ]

The lab preserves detailed research evidence.

Agent consolidation compresses it.

The canonical/brief layer exposes the current human-legible research state.

A future website/chatbot should normally answer from the upper two layers and descend into lab artifacts when:

challenged;
asked for evidence;
asked for history;
asked for a counterexample;
asked how the current state was reached.

Do not build the website/chatbot in this pass.

Just make the workspace structure compatible with that future use.

15. Challenge before canonization

Preserve the workspace’s successful adversarial norm without turning it into ceremony.

For claims that are becoming load-bearing, encourage questions like:

What is the cheapest model in which this fails?
What assumption secretly carries the conclusion?
Does this theorem become uninteresting under its antecedents?
Is the philosophical gloss stronger than the mathematics?
Has a previous witness already killed this route?

Do not require a formal adversarial report for every theorem.

The principle is:

[ \boxed{ \text{important claims should meet serious challenge before becoming canonical.} } ]

16. Do not overbuild a claim graph yet

There are attractive future ideas such as:

machine-readable claim graphs;
theorem-to-paper dependency maps;
objection edges;
formal amendment objects;
public challenge interfaces.

Evaluate them briefly.

But unless the existing workspace has an immediate concrete need, do not implement heavy graph/schema infrastructure in this pass.

Record them as aspirational workspace capabilities if useful.

The present objective is to improve the working research environment cheaply.

17. Parallel-safety constraint

This pass may occur while active object-level research is ongoing.

Therefore, while that research remains active:

Safe / preferred

documentation;
additive conventions;
small new ethos/frontier files;
clarification of layer semantics;
debt taxonomy;
templates or examples;
non-invasive status annotations;
identifying future cleanup;
documenting consolidation triggers.

Avoid unless clearly safe

renaming active research objects;
moving files used by active agents;
changing formal skeleton semantics;
changing theorem statements;
changing Lean APIs;
changing path gates;
redesigning PRIORITIES.md incompatibly;
introducing mandatory metadata that breaks existing workflow;
reorganizing active project directories.

If an apparently desirable framing change would disrupt active research, record it as a recommendation for the next consolidation boundary instead of performing it now.

18. Use the current deference work as a stress test

Inspect the current deference/corrigibility project and ask:

Can a fresh agent tell what is aspirational versus constructed?
Can it distinguish the mathematical aspiration from the philosophical aspiration?
Can it identify the current human-canonical position?
Can it tell what is merely an agent synthesis?
Can it locate historical failed routes without mistaking them for current ones?
Can it identify the dominant debt types?
Can it tell what the next controlling question is?
Can it tell which results are formal, finite-verified, report-level, architectural, or assumed?
Can it identify the major human review surfaces without reading every track?
Can it reconstruct why the program changed its mind on major concepts?

Use the answers to determine what implementation is actually needed.

Do not merely instantiate every idea in this prompt.

19. Likely implementation shape

The exact repository-conformant solution is yours to determine after inspection.

A good outcome might be surprisingly small—for example:

one concise workspace ethos/research-state document;
one lightweight consolidation template or convention;
clarification in agent instructions about lab / agent-consolidated / human-canonical status;
debt vocabulary incorporated into consolidation reports;
a small current-state/frontier convention;
one example application to the deference project;
minor deprecation/status cleanup where obviously safe.

This is preferable to creating seven new registries.

If the repository already has documents that can absorb these functions, update those instead of multiplying files.

20. Human-canonical safety rule

This is load-bearing:

[ \boxed{ \text{do not mark new substantive research claims as human-canonical merely because this pass recommends them.} } ]

You may:

define what human-canonical means;
identify which existing artifacts already have clear maintainer adoption;
propose canonical changes;
create a section listing proposed canonical deltas.

You may not manufacture maintainer review.

When status is unclear, say so.

21. Deliverables

Produce:

A. Implemented framing changes

Make the smallest set of repo changes that materially improves the architecture.

B. Workspace ethos / architecture explanation

Persist a concise description of:

[ \text{Lab} \to \text{Agent Consolidation} \to \text{Human Canonical} ]

plus:

[ \text{Aspirational} / \text{Constructed} / \text{Gap} ]

and research-debt vocabulary.

C. Deference case-study audit

Briefly demonstrate how the distinctions apply to the live deference/corrigibility work.

Do not redo the deference research.

D. Proposed canonical delta

List any changes that would require maintainer judgment.

Keep this short.

E. Deferred structural ideas

Record promising but currently unnecessary infrastructure rather than implementing it.

F. Procedural-bloat audit

Explicitly answer:

What process did you consider adding but decline because the coordination benefit did not justify the ceremony?

This is an important deliverable.

22. Final evaluation

Conclude with answers to:

What was structurally confusing before this pass?
What did you change?
What remained intentionally lightweight?
How is agent-consolidated now distinguished from human-canonical?
How is aspirational now distinguished from constructed?
How are mathematical and philosophical versions separated?
What research-debt categories survived implementation?
What triggers consolidation?
How does consolidation reduce rather than add artifacts?
What is the intended precedence for a future chatbot?
What still creates compression debt?
What proposed changes require maintainer review?
What did you deliberately refuse to proceduralize?
Did any change interfere with active research?
Is the resulting setup simpler to reason inside than the prior setup?

The success criterion is not “more structure.”

It is:

[ \boxed{ \text{a fresh human or agent can recover the live research state more cheaply and with less risk of category error.} } ]

And the central ethos is:

[ \boxed{ \begin{array}{c} \textbf{Lab: explore freely}\ \downarrow\ \textbf{Agent-consolidated: compress aggressively}\ \downarrow\ \textbf{Human-canonical: exercise judgment sparingly} \end{array} } ]

with every serious research line making it easy to distinguish:

[ \boxed{ \text{what we hope to show} \quad \text{from} \quad \text{what we have actually shown}. } ]

Prompt-author-model: GPT-5.6 Sol (OpenAI)

Report prompts/2026-08-11-workspace-ethos/REPORT.md

Report — workspace ethos pass

Prompt author: GPT-5.6 Sol (OpenAI). Executor: Claude Opus 5 (Anthropic). Date: 2026-08-11. Write scope: granted for framing, extended twice by maintainer instruction during the round — to PRIORITIES.md and to two prose rules in AGENTS.md. No research claim, theorem, epistemic class or definition was changed. Eight item status marks were added on that authorization; no item was filed, retired or renumbered.

What the repository already did, and was reused rather than rebuilt

Most of what the dispatch asks for exists here in substance. Reusing it was the larger part of the work.

Adoption mechanism — a dated DECISIONS.md entry, with self-review by a co-equal maintainer explicitly sufficient. No approval process was invented.
Precedence — already stated three times: ledger over roadmap on whether something is established, registry over prose on what is established here, PRIORITIES.md over GitHub issues on what is wanted.
Deprecation — no negative ontologies plus DECISIONS.md plus the registry's superseded-by. A superseded round record is not annotated by design, and that design is right; no deprecation registry was added.
Evidence strength — the epistemic classes, inherited-established versus workspace-established, architected, unverified-nonvacuous, and the generator/review-status split. Nothing here needed a parallel scheme.
Per-object gap lists — FINITE_MODEL_SKELETON.md §8, "what this version deliberately does not fix", eight numbered holes each marked as a hole rather than an oversight. This is the aspiration/construction distinction already being done well, one object at a time.
Provisional naming — skeleton §9 lists twenty-two working names, none proposed for permanence.
Retirement with its reason — every major reversal on the deference line carries the mechanism that forced it. That is the property this pass most wanted and least needed to add.

A. What was changed

**RESEARCH_STATE.md**, new, at the root. The three layers mapped onto the artifacts that already carry them; precedence including the rule for a lab result that refutes a canonical one; aspirational/constructed/gap on both registers with the fully-updated-deference movement worked through; seven research-debt kinds each anchored to a live instance; consolidation triggers and the recovery test; two report conventions; the supplies/does-not-supply interface examples. Its closing section states its scope: vocabulary and precedence, read by people and agents and by nothing in CI, with no list that needs synchronising.

**AGENTS.md**, four changes. agent-consolidated now says explicitly that it is a handling status and does not say whose judgment a document carries, with the two questions it is confusable with routed to DECISIONS.md and PROVENANCE.md.

Three further changes were made on maintainer instruction mid-round, which is the authorization for them: two prose rules, and standard 14 below.

No negative ontologies gains its stronger form: do not define the present by narrating an absence. The existing rule banned residue in names and structure; this bans the sentence-level version — "we no longer do X", a section explaining what a document is not because it once was — which compounds faster because prose has no length limit. It ships with a test that keeps the two legitimate cases: does the sentence work for a reader who has never heard of the absent thing? Disambiguating two currently live senses passes; so does a superseded-by pointer, which carries current epistemic content.

Slop discipline gains a seventh point, no rules-perseveration: follow the standards, do not narrate following them. Its test — is this a fact a later reader needs about the work, or a fact about the writing of it? — is what keeps it from eating the declarations the standards require, since deviations, what was not shown, outstanding actions and provenance are all facts about the work. The stated reason for the rule is that compliance narration reads as evidence of rigour and is not.

Both were applied to this round's own output: RESEARCH_STATE.md's closing section was rewritten from a list of what it is not into a statement of scope, and its agent-consolidated note now leads with what the status is.

**prompts/README.md**, one section. Round records are evidence; a later ruling governs; a report being newer than the ledger row it bears on does not reverse that, it means the row is not updated yet.

**README.md**, one sentence naming the new document in the layout section.

Four dead pointers repaired, left by the frozen/ retirement. README.md and CONTRIBUTING.md instructed readers to run tests/check_frozen.py, which does not exist; both claimed eight gates where seven run; CONTRIBUTING.md and projects/leverage/CLAIMS.md named the retired job foundations-verification rather than consolidation-verification; AGENTS.md's gates table closed on "eight jobs". .github/branch-protection.json is the source of truth and lists seven.

**PRIORITIES.md**, a standing section — Where ingenuity is the bottleneck — added on maintainer instruction mid-round, which is the authorization for a change to this file that the round otherwise had no scope to make. It is the one part of the file that is not a work order: a numbered item is a request for execution, and these are requests for an idea. Seeded with four questions read off the existing record and no new research — the near-indifference leakage, the grade-to-quantity relation, the register that would price who holds authorization, and the time-indexed A-valuation family with the two holes beside it. The section's own rule is that an entry graduates by becoming a numbered item, which is why the leverage line has no entries: its hard problems are open with known shapes and are filed as items 1 and 2. Headings are ### Q<n> — rather than ### <n>. so that checkers/registry.py, which resolves answers_item against ^###\s+(\d+)\., cannot mistake a question for a filed item. Verified: the registry gate still resolves three entries and the item set is unchanged.

**RESEARCH_STATE.md**, a matching section stating the convention and its diagnostic — most entries are model or assumption debt, because theorem debt is dispatchable and debt that is not yet theorem debt is usually waiting on an idea rather than on labour.

**Status marks on PRIORITIES.md items 14–21**, on maintainer authorization, which discharges what had been proposed delta 2. All seven wave-1 items are marked returned wave 1; item 20 additionally superseded by item 26; item 21 answered by Stage II; the magnitude target is retired. The section gains a paragraph saying what the mark means and does not mean — it records dispatch history, not that the science is settled, and several returned negative or partial. No item was renumbered or removed, so answers_item resolution is unaffected; item 13's existing *satisfied, kept open* supplied the form.

**A second standing section, *Workspace friction***, and AGENTS.md standard 14 obliging a round to use it: structural defects are reported, not worked around, because a workaround leaves nothing behind and the next round pays the same cost unknowingly. Reporting is the obligation and fixing is a maintainer decision, with one exception — a dead pointer is a fact rather than a design question and is repaired in place. Seeded with five defects this pass hit: the path gate's permissive default for new root documents; the absence of any check that a documented command names a file that exists; and three consequences of maintainer throughput, below.

**RESEARCH_STATE.md**, the attention paragraph. The maintainer writes in few places and does not read most of what is produced here. That is recorded as a design parameter rather than a backlog: ci-only is the standing condition of almost everything, maintainer-reviewed is rare and deliberate, and the judgment is spent on naming and on what is worth proving rather than on reading prose for approval. Two standing rules were written on the opposite assumption, were filed as friction, and were ruled on by the maintainer during the round — see the next section.

Three rulings landed during the round, and the documents were changed to match rather than left describing rules that no longer hold. The flagship rule is retired from AGENTS.md and PROVENANCE.md; external citation is restated to stand on its own, requiring that a citation carry the status the thing actually has; and a maintainer-dispatched round may now file PRIORITIES.md items within its own scope, with demand-gating and reserved naming otherwise intact. The dated entry is in DECISIONS.md, the friction entries were retired or narrowed, and no residue of the retired rule was left in a living document.

One queue. AGENTS.md §10 now requires a reserved item to be appended to DECISIONS.md's Awaiting the author rather than left in the round's own report; the ledger's header says the same; and the section is populated, at six entries, each with what deciding it costs. Four competing sources of reserved items existed and none answered "what needs me?".

Six further rulings, taken against the populated queue and performed, which emptied it to one entry: prose is not externally citable and asking the maintainers is the mechanism; RESEARCH_STATE.md becomes specification layer; the governance report moves to prompts/2026-08-10-contribution-architecture/; the deference line gets TERMS.md; the leverage forward tree keeps its name; and further leverage trees are frozen at the next leverage round rather than now. Two of those had live pointers to repair — the moved report was both dual-register documents of a registered claim — and no claim changed class.

Four more rulings, on the questions carried out of the contribution-architecture round: auto-merge on full green; the checker harness ships unread, recorded as a decision rather than an omission; the resource budgets confirmed; and the permissive path-gate default confirmed, with the enumeration itself explicitly not re-approved, since it has changed repeatedly since that report proposed it.

Auto-merge is GitHub-native rather than a workflow, and that is forced rather than chosen: a merging bot needs write scope, and CI holds zero secrets, permanently is not a rule to spend on convenience. Implementing it surfaced a stale literal — .github/apply-branch-protection.sh required exactly eight required checks and the payload has carried seven since frozen-integrity was retired, so the verifier would have reported correct protection as wrong. It now counts what the payload declares.

**PROVENANCE.md**, rows for the above and a round-attribution line.

C. Deference case study — the ten stress-test questions

Answered against the line as it stood before this pass. Where the answer is now different, that is said.

Aspirational versus constructed — yes, and better than the dispatch assumes. The ledger says "not derivable", "retired", "empty rather than false", "not well-posed as attempted". The cost is that it is carried in prose across three hundred-odd lines and has to be reassembled on each read.
Mathematical versus philosophical aspiration — yes, and this is the line's strongest habit. Fully updated deference carries a deflationary gloss with the stronger reading marked unsupported; the non-preemption impossibility is labelled fatal read as jurisdiction and correct read as autonomy. It was never named as a convention, so a new track has nothing prompting it to do the same.
The current human-canonical position — no, not reliably, and this was the largest category error available. The roadmap and the ledger open by calling themselves canonical; PROVENANCE.md records them ci-only. Both are true and they answer different questions — which document governs, versus whether anyone read it — and nothing said so. Fixed.
What is merely an agent synthesis — same defect, same fix. The layer table now says it; the four line documents still do not carry the pointer themselves, which is proposed delta 2.
Historical failed routes — yes within the ledger, which records each retirement with its reason. The exposure is prompts/: a reader arriving at phase-ii-certificate/REPORT.md by path meets a strict-minority gloss that a later ledger entry refutes with an exact counterexample, and the report is correctly not edited. prompts/README.md now states the rule.
Dominant debt types — no, because everything unresolved was open or blocked. Now nameable: this line's dominant debts are assumption (item 25's near-indifference leakage, which makes the surviving competence candidate's bound vacuous unbounded) and model (the skeleton's FU[g] hole, and no operation reassigning the authorization relation at a later index), with interpretation debt on three reserved maintainer-decisions.
The next controlling question — yes, unambiguously, and it moved during this pass without becoming ambiguous: the Stage III ruling withdrew the round's positive reading, so the recommended dispatch went from item 25 alone to items 25 and 27, and both the ledger and DECISIONS.md say which and why. Best-signposted thing on the line, and the reversal is the evidence for that rather than against it.
Which results are formal, finite-verified, report-level, architectural or assumed — yes, mechanically. "155 theorems kernel-verified and unregistered, and those are different things" is the model sentence.
Human review surfaces without reading every track — partly. They are in four places: DECISIONS.md's Awaiting the author, GOVERNANCE_REPORT.md's own list, each report's Outstanding maintainer actions, and the maintainer-decision rows in the ledger. No single view. Collecting them into one would be a fifth list, so this is left as a convention rather than a file.
Why the program changed its mind — yes, and this is the repository's best property. Every reversal carries its mechanism: regret bounds are equivalent to the conclusion they were meant to buy; trader net worth is affine in settlement and absolute value is not; the valuation register prices whose selection authorises and never who holds the right.

Worked debt-transition example, all three already on the record. Choice-level competence was theorem debt, and the finding that every regret-bound form is equivalent to its conclusion makes the request ill-posed — model debt, and the vocabulary had to change. The certificate's non-preemption impossibility moved from theorem debt to interpretation debt: same mathematics, reread as a bound on autonomy. Fully updated deference moved from theorem debt to apparent assumption debt and then, under adversarial review during this pass, to model debt: the transferred arm contained no distinct future agent, so there was no question for an assumption to answer. All three read as failure under open/blocked and as progress under the debt vocabulary.

D. Proposed canonical delta

Three of the five were ruled on during the round and performed; see the dated DECISIONS.md entry. Two remain proposed.

**One line at the head of each deference notes/ document** separating "canonical for this line" from review status. Not done: those files carried in-flight edits from two deference rounds throughout this pass.
Whether the ledger absorbs the state shape as a header section, so the live deference state is recoverable in forty lines rather than six hundred. The cheapest available reduction of the line's compression debt.

E. Deferred structural ideas

Recorded, not built. None has a present concrete need.

Machine-readable claim graph, objection edges, theorem-to-paper dependency map. The hook exists — CLAIMS.md entries already carry answers_item and the schema has room for superseded-by — so an edge is a field away when demand appears. It has not: the deference line has no CLAIMS.md at all, so a graph would currently be over an empty set while the line's 155 kernel-verified theorems sit unregistered. Registration is the prerequisite, not the graph.
Formal amendment objects for canonical changes. A dated ledger entry already does this and reads better.
A public challenge interface. CONTRIBUTING.md already specifies the format — a counterexample, a failing test, or a precise objection against a named item — and GitHub issues carry it. A dedicated surface waits for traffic.
A per-line frontier document. Would desynchronise with the ledger, which wins on precedence. Delta 6 proposes absorbing it instead.

F. Procedural-bloat audit

Considered and declined, with the reason.

A status header on every research note. The dispatch's own prohibition, and it would tax exactly the exploration the lab layer exists to make cheap. Layer is derivable from location.
**A FRONTIER.md per line.** A fifth list to keep in sync with four that already exist. Absorbed into the ledger as a proposal instead.
**A deprecation registry, and SUPERSEDED BY: headers on round records.** No negative ontologies deliberately keeps supersession in the two places history lives. Annotating reports would put it in a third and re-open the question of who updates them.
Debt words for formalization and verification. Both already have names here — the standing Lean-port item family and the class ordering; and unverified-nonvacuous, unregistered, ci-only. A second vocabulary for the same facts is the parallel machinery the dispatch warned against. Seven kinds survived, not nine.
Numeric debt scores. Classification is the point; a number would imply a comparison nothing supports.
A CI gate for layer or aspiration metadata. There is no rule here to enforce, and this repository's own record shows a gate that matches nothing is indistinguishable from one that works.
A maintainer-review queue file. Four sources of review surfaces exist; a fifth that must be synchronised by hand is the one that goes stale.
An adversarial-review requirement per theorem. Kept as a norm. The line already runs independent adversarial contexts where it matters — Stage III's Track F ran with no access to the round report — and mandating it would price out cheap rounds.

Deviations from the dispatch

The dispatch's layer name "Agent-Consolidated State" was not adopted as a label. agent-consolidated already means something else here — a handling status attaching to material at any layer, including the leverage line's authoritative record — and importing the dispatch's sense would have created the confusion the pass exists to remove. The layer is called consolidated view, and the collision is stated explicitly in both documents.
§5's state shape, §11's symbol table and §12's interface table were not installed as line documents. The four deference notes/ files carry uncommitted Stage III edits; adding this pass's changes to them would make neither diff attributable, which §17 forbids in substance. §12's content is demonstrated in RESEARCH_STATE.md from material quoted out of the roadmap; §5 and §11 are deltas 6 and 5.
**tests/path_gate.py was not edited**, though the new document currently lands in the proof layer. §17 names path gates as avoid-unless-clearly-safe and it is a trust-chain file. Outstanding action 1.
The pass went slightly beyond framing to repair four dead pointers, on the ground that a command naming a deleted script is a defect in the contributor onboarding path rather than a matter of framing.
Nothing was committed. The working tree already carried the Stage III round's uncommitted output when this round began.
The case study was rewritten mid-round against a moving record. Between the first and second readings of the deference ledger, an independent adversarial review overturned Stage III's positive reading: the comparator was found to contain no distinct future agent, FUD_COMPARATOR_SPEC.md v1 became a corrected defective record, and item 27 was filed. The worked example in RESEARCH_STATE.md had been written from the withdrawn reading and was replaced. It is kept as the worked case rather than swapped for a stabler one, because the aspiration survived intact while the constructed state shrank and the debt changed kind — which is the distinction the document exists to make. It is also the pass's only live test of the precedence rule: a lower-layer artifact refuted a consolidated reading, and the ruling landed in DECISIONS.md rather than as an edit to the round record.

What this round does not establish

No research claim, theorem, class, item status or definition changed. The case-study section restates the ledger and roadmap; where it appears to characterise the science it is quoting them, and they govern.
The layer semantics are a reading of AGENTS.md, DECISIONS.md and PROVENANCE.md, not a legislation. That DECISIONS.md is the whole adoption mechanism is what those documents currently say; whether the maintainer wants it to remain the whole mechanism is itself a maintainer question.
The category error is asserted as available, not observed. No agent is on record having read a ci-only canonical document as endorsed. The claim is that the documents permit it, which is checkable; that it has happened is not claimed.
Whether the debt vocabulary reduces confusion is untested — no round has used it, and the honest test is whether the next consolidation reaches for it.
The line's compression debt is named, not discharged. Recovering the live deference state still needs the roadmap, the ledger, the dispatch queue, the skeleton, the comparator spec and eight DECISIONS.md entries. Delta 6 is the cheapest available reduction and it was not performed.

Outstanding maintainer actions

The queue is DECISIONS.md's Awaiting the author; this is the round's own record of what it reserved. Everything this round raised has been ruled on except the two deltas above and one carried entry.

Rule on the two remaining deltas. Neither has a cost to waiting.
Commit. Every uncommitted file in the tree is now this round's — the deference rounds committed theirs at bb8b351. The move shows as a rename, so git add -A preserves it.

Claims it registered

None.

Items it filed or answered

None recorded in the index.

Nodes pointing at it

enabling-work