2026-08-11-priorities-and-slop

PATCH ROUND — PRIORITIES rename and slop discipline

Extends the governance architecture; AGENTS.md rules apply. Read from the public repo at 917c7da. Confirm current state yourself; pointers are not a spec.

A. Rename OPEN_PROBLEMS.md → PRIORITIES.md

git mv, then update every live reference. Fourteen tracked files outside frozen/ and prompts/ mention it, including three code paths — checkers/registry.py, checkers/run.py, tests/path_gate.py — where a missed string breaks a gate silently rather than loudly. Also README.md, CONTRIBUTING.md, AGENTS.md, PROVENANCE.md, SETUP_REPORT.md, GOVERNANCE_REPORT.md, .github/PULL_REQUEST_TEMPLATE.md, and projects/leverage/README.md. Files under projects/leverage/forward/attic/ are retired material — leave them.

Rewrite the file's own heading and any prose calling the items "open problems". The framing shifts with the name: this file says what the program wants done next, in its own order, not merely what happens to be unsolved. Difficulty tags stay.

No stub, no redirect note, no "formerly" line — per no-negative-ontologies. Round records under prompts/ keep the old name. DECISIONS.md gets the dated entry; that is where the rename lives.

After the rename, confirm the gates still find the file: run the checkers and path-gate suites and report that they exercised the renamed path rather than passing on an empty match. A gate that silently stops checking looks identical to a gate that passes.

B. Slop discipline — new section in AGENTS.md, summarized in CONTRIBUTING.md

Why it belongs in a verification repository, stated first so the rule is not read as taste: a reader who cannot tell which sentences carry content cannot audit; a document that restates itself three ways hides its own errors; and volume inflates the cost of the maintainer review the whole architecture depends on.

The rule, adapting the wording:

1. Every sentence does work. Cut restatement, throat-clearing, and previews of what the document is about to say. Prefer the shortest form that keeps the content.
2. No padding structure. Headings, tables, and lists are for material that is genuinely sectioned, tabular, or enumerable. A document that would be four paragraphs is four paragraphs.
3. No inflated register. No "comprehensive", "robust", "powerful", "seamlessly", "leverage" as a verb; no assertion of a result's importance in place of stating the result.
4. Hedging is content or it is cut. "May", "could", "arguably" are right when they mark a real epistemic state — and the epistemic classes exist to record that precisely. Hedging that softens a claim the writer will not commit to has no place in a record whose function is to say what is established.
5. No summary of a summary. Each deliverable ships a verification register and a human register; neither gets its own executive summary, and reports do not restate their findings in a closing section.
6. Empty results are reported as empty — cross-reference the existing rule on stating what was not shown; do not restate it.
7. The maintainer may reject on these grounds alone. A pull request whose content is correct and whose prose is padded is a legitimate rejection, said plainly rather than merged and cleaned up later.

Add: agent reports are deliverables under this rule. A 900-line report for a 40-line result is a round done badly.

C. Bounded application pass

README.md (92 lines), CONTRIBUTING.md (207), AGENTS.md (494), PRIORITIES.md (240). Read each against §B and cut what plainly violates it. Report before/after line counts per file with what was cut.

Two constraints. Do not cut content to hit a number — if a document does not violate the rule, say so and change nothing; do not manufacture edits to demonstrate the pass ran. And do not cut anything a gate depends on: several checkbox and field names in CONTRIBUTING.md and the PR template are parsed or enforced, so verify against the scripts before shortening them.

Leave prompts/ (history) and frozen/ (immutable) alone. AGENTS.md is the longest document and the one most likely to hold real duplication after five rounds of amendment; it is also the binding document, so cuts there are removals of restatement only — no rule loses force in this pass.

Report

ROUND_REPORT.md per convention: the rename's full reference list including the three code paths, the gate-exercise confirmation from §A, the AGENTS.md diff, per-file counts from §C with what was cut, and any place where applying §B would have cost content rather than padding — left in place and listed.

Under Outstanding maintainer actions, list anything reserved.

---

Prompt author: Claude Fable 5 (Anthropic), 2026-08-11, in a maintainer-directed session. Dispatched by the maintainer.

PRIORITIES rename and slop discipline — report

Dispatch: PROMPT.md beside this file, verbatim.

Stacked on prompts/2026-08-11-attribution-provenance-names/, which was open when this round was dispatched and touches the same four documents. Line counts below therefore distinguish the dispatch's baseline at 917c7da from the state this round actually started from.

Outstanding maintainer actions

1. **Decide whether ## PRIORITIES item answered should stay a template field name.** It is prose, not parsed — no script reads it — so it is free to change, but it is the field contributors fill in and renaming it twice is worse than renaming it once.
2. Carried from the previous round and still open: the scrub-round-2 dispatch and the contributor-checkers dispatch are both missing from prompts/, and whether the program gets a name is reserved.

A. The rename

git mv OPEN_PROBLEMS.md PRIORITIES.md. Fifteen live references in eleven files:

file	references
checkers/registry.py	the module docstring, the path lookup, the failure message
checkers/run.py	the self-test's temporary-tree fixture
tests/path_gate.py	the specification-path list
.github/PULL_REQUEST_TEMPLATE.md	the section heading and its comment
AGENTS.md	the specification-layer enumeration, the demand-gating rule, the gate table
CONTRIBUTING.md	the demand-gating rule, "where to find work"
README.md	the layout description, the contributing pointer
GOVERNANCE_REPORT.md	the demand-gating line, the specification-path listing
PROVENANCE.md	the table row
SETUP_REPORT.md	the item-count finding
PRIORITIES.md	its own heading and framing

Two references were deliberately not changed. PRIORITIES.md line 32 and projects/leverage/README.md line 40 both point at frozen/consolidation_aug9/OPEN_PROBLEMS.md — the frozen tree's own list, which is immutable and keeps its name. Renaming those would have produced two dangling paths.

Report references were changed, unlike the provenance vocabulary in the previous round. The distinction: a report saying "OPEN_PROBLEMS.md has eleven items" is a path reference that dangles once the file moves, while a report saying its output is llm-unreviewed is a statement of state at the time. Paths get updated so they resolve; vocabulary statements stay as records.

Framing

# Open problems → # Priorities, and the opening now says the file is the maintainer's ranking of what would move the work rather than an inventory of what is unsolved — so an item's absence means nobody has asked for it. Difficulty tags unchanged. No stub, no redirect, no "formerly".

The gate-exercise confirmation, and a bug the rename exposed

The dispatch asked for proof the gates exercise the renamed path rather than passing on an empty match. They did not, and would not have:

open_items = set()
if open_path.exists():
    open_items = set(re.findall(r"^###\s+(\d+)\.", ...))
...
if open_items and item and item not in open_items:

A missing file left open_items empty, and the guard then skipped the answers-a-filed-item check on every claim while the gate reported green. The rename would have silently disabled a check rather than breaking anything. Both a missing file and an empty item set are now hard failures, and the check no longer guards on the set being non-empty.

Two permanent self-test cases, taking the harness from 9 to 11:

ok: claim answering an unfiled priorities item is rejected
ok: missing PRIORITIES.md fails rather than skipping the check

The second is the one that matters — it fails if anyone reintroduces the skip.

Path gate, checked directly rather than inferred from a green run:

PRIORITIES.md          spec=True   proof=False
OPEN_PROBLEMS.md       spec=False  proof=False

The new name is in the specification set and the old one is in neither, which is what a live rename should look like from the gate's side.

One rename bug of my own, caught before commit. Renaming the variable open_path → priorities left two uses of the old name behind. Python would have raised NameError on every registry run — loud, not silent, so CI would have caught it — but it is the same class of error the dispatch warned about.

B. Slop discipline

New ## Slop discipline section in AGENTS.md, seven numbered rules as dispatched, with the rationale first so the rule does not read as taste. Rule 6 cross-references §9 rather than restating it, per the dispatch. Agent reports are named as deliverables under the rule.

CONTRIBUTING.md gains a ## Prose section: the short form, the reason it is not style policing, and that a padded pull request can be rejected on that ground. The ## Review list gains "prose".

C. The application pass

file	at 917c7da	entering §C	after §C
README.md	92	94	92
CONTRIBUTING.md	207	218	223
AGENTS.md	494	556	565
PRIORITIES.md	240	242	240

The pass net-added lines, and that is the honest result. Reading these four documents against §B turned up less padding than it turned up factual defects, and correcting a stale table costs more lines than deleting a restatement saves. The cuts and the corrections are listed separately below so the two are not confused.

AGENTS.md — cut

• **The doubled --- before No negative ontologies.** Structural residue of an insertion.
• **### 9b.** A section number with a letter is a visible amendment marker. Renumbered 9b→10 and 10–12→11–13, with the four cross-references updated (§12→§13, §13→§14, and the gate table's write-scope row).
• The contrib-checker asymmetry, stated then restated. "That asymmetry is why one is gated and the other is merely labelled" repeated the preceding paragraph; cut, keeping the sentence after it, which is new content.
• The Lean regime's opener, which restated standard 4 in full before pointing at it. Now points at it.

AGENTS.md — corrected, not cut

• The gate table was wrong. It listed four numbered gates and closed with "Six gated standards decide correctness" while CI runs eight jobs. Nine rows added: path-gate, conservativity, the registry and the contributor-checked ceiling, contrib hygiene, DCO, model attribution, the name lint, reserved-items, and slop discipline. This is the document's map of what is enforced, and it had drifted from the workflow.
• Numbered gate references retired. "Gate 1/2/3" in standards 1, 4 and 5 named a numbering the table no longer uses. They now name the CI job, which is also what branch protection matches on.

CONTRIBUTING.md

• Corrected: "Four gates run in CI and all four run locally" — there are eight. The local-run block listed five commands under the old gate numbers; it now lists all eight commands the jobs run, labelled by job, and says plainly which two gates have no local form (dco reads commit sign-offs; attribution reads a pull-request body that does not exist yet).
• Cut: one stray double blank line.
• Otherwise unchanged, deliberately. It is contributor-facing and restates AGENTS.md by design — that is the dual register, not duplication within a document. Nothing in it plainly violates §B.

README.md

• Cut: "That is not a slogan about openness; it is what the house discipline is for" — a negated strawman ahead of the list that makes the point.
• Corrected: the PROVENANCE.md description still described the three-class scheme ("what was written by a person and what by a model, and which of the latter the author has reviewed"), superseded by generator + review status in the previous round.

PRIORITIES.md

• Cut: the ## Infrastructure, continued heading — padding structure, three items split from the section above it for no reason. Merged.
• Otherwise unchanged. The item bodies are dense and the framing was rewritten in §A.

Where §B would have cost content — left in place

• **CONTRIBUTING.md's "This is an invitation, not a consolation."** Reads as inflated register, but it is doing real work: it tells a contributor that contributor-checked is not a demerit, which the surrounding mechanics do not say.
• **AGENTS.md's "Everything below follows from that."** A preview sentence by §B.1, and load-bearing: it is what makes the four-part premise binding on the rest rather than decorative.
• The trust chain's closing "If you are auditing this repository, audit that list." Restates the list's purpose. Kept — it is the one instruction an external auditor needs, and it is where they will look.
• **README.md's two-line description of each research line.** Longer than strictly needed; it is the only place a stranger learns what the program is.
• **CONTRIBUTING.md's restatement of the contrib-checker rules** at more length than AGENTS.md states them. Cross-document restatement for a different audience is the dual-register design.

Not touched, and why

Field names and checkbox text in CONTRIBUTING.md and the pull-request template were verified against the scripts before this pass: tests/attribution.py parses ## Model attribution, and nothing parses any other heading or checkbox in either document. The attribution heading was left exactly as the parser expects.

Gates

Green locally: name lint (37 files), attribution self-test (7/7), checkers self-test (11/11), registries (3 entries), contrib hygiene, path gate, conservativity, check_frozen, repo runner, py_compile. The Lean gate is unaffected — this round touches no Lean — and runs in CI.

Attribution

prompt author	Claude Fable 5 (Anthropic)
executor	Claude Opus 5 (Anthropic)
date	2026-08-11