Codex Formalization Workflow
Install
This is a Codex skill package. Install it by copying the skills/codex-formalize/ directory into $CODEX_HOME/skills/ (default ~/.codex/skills), or load it by naming this SKILL.md path directly in a dispatcher. This repository copy is the single source of truth; any installed copy is a projection of it.
Scope and authority
This file is Codex-specific packaging of repository obligations; it has no authority of its own. docs/develop/spec/golden-ledger-repo-spec.md is the sole normative specification; CLAUDE.md is the invariant frame governing how work is done; and agents/CONTEXT.md is the finite-context map and routing aid, not an authority above the specification. Live harness output is the decisive judge of fact about the current tree. If this file disagrees with any of them, they win and this file is the bug.
Read first
CLAUDE.md- owns the repository's invariant frame and working ethics.agents/CONTEXT.md- owns the finite-context map, routing, and naming guidance.docs/develop/spec/golden-ledger-repo-spec.md- owns the normative repository specification.- The applicable
agents/*.mdrole charter - owns role-specific duties. agents/echo-template.md- owns the statement-echo record.make help- owns the live catalogue of canonical doors.tools/- owns executable admission and repository enforcement.
State machine
Follow these steps in order. Do not pass a step until its postcondition holds.
0. Establish the environment and isolation
Run:
eval "$(sed -n '/^export PATH=/p' tools/scripts/local-harness-gate.sh)"
# If not already in a dispatcher-assigned isolated lane:
make worktree KIND=math NAME=<lane> && cd <created-path-from-output>
pwd -P && git rev-parse --show-toplevel && make -C tools dotnet
Re-read the current export PATH from tools/scripts/local-harness-gate.sh for every task rather than trusting a list quoted elsewhere; its /usr/sbin entry must survive because the report supervisor requires lsof, which lives there. If a dispatcher already assigned an isolated lane, do not create a second one: confirm the existing lane with the same pwd -P / git rev-parse --show-toplevel check. Otherwise, the make worktree output is JSON whose path field names the created path; substitute that value for <created-path-from-output> and work only there.
Build through the canonical make -C tools dotnet door because make show-atom runs the Release CLI with --no-build.
Before any deposit, require git status --short to print nothing except the intended formalization changes. Prefer a fully clean tree before beginning the task. Deposit and cover only update the working tree; inspect their resulting delta and commit it through the normal repository workflow.
Postcondition: the pinned toolchain is on PATH; pwd -P and git rev-parse --show-toplevel agree with the assigned or created isolated lane; make -C tools dotnet has built the CLI so make show-atom succeeds; and no unrelated or unexplained change is present.
1. Choose exactly one open atom
Inspect candidate snapshots in Generated/echo-residuals/<source_id>.md. Treat them only as candidate listings. Obtain the authoritative atom text with:
make show-atom ATOM_ID=<id>
Never quote the projection as authoritative. Prefer an atom with few unresolved subitems and an elementary, self-contained statement.
Triage the claim class before committing to it; each class below is named by landed outcomes, not speculation:
- Best odds — concrete certificate/computation claims whose data is inside the atom text (a walk value, a finite identity, explicit witnesses): these close with
decide/norm_numand have the highest landed success rate. - Good odds — claims whose machinery is already frozen: a bridge, instantiation, or characterization theorem connecting existing declarations.
- Not a standalone target — bind-only corollaries (
CLAUDE.md5⁗).proof_shapeis purely semantic: after inlining local aliases, private helpers, tactic expansions and definitional equalities, a public theorem isbind-onlywhen its conclusion follows from frozen public theorems only by instantiation (substituting parameters — direct instantiations and 第 11 条 thin upstream wrappers are bind-only too), discharging a frozen hypothesis with a frozen witness (de-hypothesization), logical projection or repackaging (.1/.2/.mp/.mpr, conjunctions, TFAE), or normalization (definitional unfolding;simpwith frozen/Mathlib rewrite lemmas only;ring/linarith/omega/norm_num/decideclosing a goal whose atomic facts are all already supplied — a decision procedure that establishes a NEW proposition not supplied by the dependencies yields a candidate witness, not normalization). Non-exhaustive landed examples: one application of a frozen theorem;obtain ⟨x⟩then one frozen theorem; projections or compositions of frozen iffs;Finset.sum_empty/simp/ring/linarith-level rewrites; TFAE/conjunction packaging. Admissibility is a SEPARATE field,admission_basis∈ {escape-witness,rule-11-upstream-wrapper(cite the upstream declaration hit AND the atom clause or concrete repository API/coverage need that makes the wrapper required),atom-required-bridge(cite the atom clause, the previously absent typed edge, and the pre-registered named consumer)}; the latter two are exceptions that permit a deposit without changingproof_shape. A probe may classify a candidate; once it is bind-only with no admission basis it is digestion (if a frozen theorem covers the atom verbatim,make coverit — zero Lean) or a named companion inside a module that has an escape-bearing theorem — never the sole content of a first-freeze module. If every open atom in reach is bind-only, the valuable output isopennaming the missing analytic input, not a deposit. - Definition clauses — only with an earning theorem (see the thin-deposit taxonomy below); a definition alone is not a target.
- Do not encode — institutional/philosophical prose clauses (governance clauses, postmortem narratives, interpretive premises): they have no mathematical content, and encoding them as generic set/logic predicates is how thin deposits happen. Report them as not-formalizable prose rather than dressing them in Lean.
- Do not attempt without a machinery plan — heavy universal claims (representation theorems, general-dimension obstructions): landed lanes on these either time out or fabricate. If the machinery gap is real,
opennaming the gap is the valuable output.
Postcondition: one atom ID is selected, its claim class is named in the report, and its verified make show-atom output is retained as the statement source.
2. Echo the statement before proving it
Follow agents/echo-template.md. Write a clause-level mapping from every quantifier, domain, hypothesis, conclusion, and generality claim in the authoritative atom text to the intended Lean declaration. Account for every unresolved subitem.
If an ambiguity cannot be resolved without weakening the claim, stop and report the result as open.
Postcondition: every source clause has one intended Lean counterpart, or the task has ended as open with the ambiguity named.
3. Search the library before proving
Apply CLAUDE.md 11, "library before proof." Search pinned mathlib and the repository's D5/ declarations for the complete statement and for lemmas that close its dependencies. Record every query verbatim and record whether it hit.
Check the bind path first: the cheapest faithful discharge is an existing frozen theorem, and dozens of residues have been discharged with zero new Lean. If a frozen declaration already covers the residue leg verbatim (not narrower — watch the grader traps below), the correct deliverable is a bind recommendation naming that declaration and the evidence, not a new module. Ledger surgery itself is dispatcher-owned; your report carries the finding.
If the result exists upstream, import and apply it. Do not reprove it: a reproof of an existing declaration creates a second source of truth.
Search for the statement shape, not the module name. A landed lane died reproving a frozen theorem in strictly weaker form (it added a derivable Odd d hypothesis) even though it had been told to read that exact module: reading a module and greping for the identity are different acts, and only the second one protects you. Grep D5/ for the hypothesis pattern and the conclusion pattern of your target, for example:
grep -rn "≡ -1 \[ZMOD" D5/ --include='*.lean'
grep -rn "jacobiSym\|J(" D5/ --include='*.lean' | grep -i <conclusion-token>
If the identity already exists in equal or stronger form, the correct output is the reuse or an open/abstain naming the frozen declaration — never a new module. A hypothesis derivable from your other hypotheses (parity from a congruence, coprimality from an inverse relation) must be derived, not assumed: assuming it makes your theorem strictly weaker than the source claim and reviewers reject it as a weaker duplicate.
Postcondition: the search trace is recorded and every hit is either reused or accompanied by a concrete explanation of why it is not the same claim.
If the search cannot be completed faithfully, end the task as open with no deposit, carrying the Step 8 evidence: statement echo, search trace, failed approaches with reasons, and machine diagnostics.
4. Learn the current artifact shape
Read the nearest current Lean and Scribe artifacts in the target bucket, then read the live path-policy and header owners. The current working tree is the template; historical commits are not a workflow contract.
git ls-files '<target-lean-directory>/*.lean'
git ls-files '<target-blueprint-directory>/*.scribe.cs'
sed -n '1,12p' <nearest-current-lean-module>
Copy the exact shape from current tracked neighbours rather than from remembered or historical formats. If this file and the current tree disagree, the current tree and its live harness win.
Postcondition: the chosen current neighbours and owner paths are recorded, and the planned artifacts follow that observed shape.
5. Write and compile the artifacts
Write the Lean module and its .scribe.cs mirror using the live template. Discover the current path-to-GID rule from that template and the live path-policy owner; do not rely on a remembered grammar.
A new theorem must go in a new Lean module. Before writing into an existing module, check whether it has an active Freeze event:
module_path='D5/path/Module.lean'; grep -l -F "$module_path" Golden/Frozen/accepted/*.json
Exit 0 with an accepted-record path means frozen; exit 1 with no output means not frozen (any other result is a failed check). The frozen ledger pins the module's declaration set, not just its bytes, so it refuses adding a declaration to a frozen module. Reattest covers changed bytes with an unchanged declaration set; it is not an escape hatch for adding a declaration. If the atom genuinely belongs inside an existing frozen module, do not edit it: end open, naming that module and the frozen-ledger constraint.
Before creating the new module, set lean_dir to its target directory and blueprint_dir to the corresponding Blueprint/ mirror directory, then read the live limit from its owner (DirectoryFileLimit in tools/StrataLint.Engine/Rules/RepositoryRules.Structure.cs; never a number remembered from this file), measure both, and confirm that adding one counted file to each stays within that limit:
limit=$(sed -nE 's/.*DirectoryFileLimit = ([0-9]+);.*/\1/p' tools/StrataLint.Engine/Rules/RepositoryRules.Structure.cs); test -n "$limit"; lean_count=$(git ls-files "$lean_dir" | awk 'END { print NR+0 }'); blueprint_count=$(git ls-files "$blueprint_dir" | awk '!/\.md$/ { n++ } END { print n+0 }'); printf 'limit %s\n%s %s\n%s %s\n' "$limit" "$lean_dir" "$lean_count" "$blueprint_dir" "$blueprint_count"; test $((lean_count + 1)) -le "$limit" && test $((blueprint_count + 1)) -le "$limit"
Committed Blueprint .md renderer oracles are excluded from capacity, but .scribe.cs sources count. If the natural target directory is full, do not place the module in a semantically wrong directory to evade the limit: split the bucket, register the new domain in Meta/domains.yaml, place the module in the new directory, and carry the protected-surface cost for the conservative-extension gate to judge. A protected-surface change is priced work, not a stopping condition. Reserve open for what is genuinely unresolvable, such as an ambiguity that cannot be settled without weakening the claim or a proof that will not close, never for work that merely costs more.
Iterate on the source artifacts, then run the canonical door when they are ready:
make lean
Judge completion only by exit code, never elapsed time or quiet output. Full doors cost minutes each; a landed lane died by burning its entire three-hour budget on seventy-two full preflight runs chasing a flaky unrelated test. Iterate scoped, verify canonically once.
Run every shape check NOW, before Step 7: line 6 ends with -/, the generality tag matches the weakest import and the module's nature, the scribe formulas obey the rejection taxonomy, the emitted .md mirrors every conjunct. After a successful deposit the module's bytes are pinned by the frozen ledger, so catch defects before depositing rather than creating a conflicting frozen state.
Postcondition: both source artifacts exist in the observed shape and make lean exits 0.
If a faithful proof cannot be made to compile, end the task as open with no deposit, carrying the Step 8 evidence: statement echo, search trace, failed approaches with reasons, and machine diagnostics.
6. Run the fidelity and non-hollowness gate
Complete every evidence item in the checklist below. A green compiler and harness do not discharge this step.
Postcondition: every checklist item has evidence and none is ASSUMED-UNVERIFIED; otherwise deposit is blocked.
If any checklist item cannot be evidenced, end the task as open with no deposit, carrying the Step 8 evidence: statement echo, search trace, failed approaches with reasons, and machine diagnostics.
7. Deposit and cover, then early local feedback
Only after Step 6 passes, run:
make deposit ATOM_ID=<id> GID=<D5/Path/Module.theorem_name>
make cover ATOM_ID=<id> GID=<gid>
make preflight BASE="$(git rev-parse HEAD^1)" # early feedback only; NOT a gate (CLAUDE.md 器律②)
Deposit and cover operate on the same working tree and may land in the same pull request. Neither command creates an intermediate commit.
make deposit and make cover must exit 0. make preflight BASE="$(git rev-parse HEAD^1)" is local early feedback: report its exit code and every rejected rule verbatim, but a nonzero local preflight does not stop the lane — the three remote required checks are the only judgment (CLAUDE.md 器律②, user 2026-08-26; ②′ push and local verification run in parallel). This macOS host has known preflight noise (Perl C.UTF-8 locale crashes in script tests, ENGINEERING_TEST_EVIDENCE_FAILED without TRX, an SL-022 observation line); stop only on a content-level red you caused (Lean error, SL-008 frozen-surface diff, SL-010 generality, digestion coverage), and then end as open. Judge them only by exit code, and never pipe a judgment command: cmd | tail -1 reports the pipe's exit status, not the command's, and three landed incidents (a merge that silently failed, validation run on a stale base, a cover failure read as success) trace to exactly this. Run the command bare, or capture $? on the command itself before any formatting.
If make deposit or make cover exits nonzero, stop and end as open. Report the failed command and exit code, machine diagnostics, touched paths, and the actual resulting tree state; deposit and cover do not commit their changes.
Before pushing, inspect the complete deposit-and-cover delta and create one builder-owned commit:
git status --short
git diff --check
git add -A
git commit -F <commit-message-file>
Postcondition: deposit and cover exited 0; preflight exit code and diagnostics reported; the complete intended delta is in one explicit builder commit; and the worktree is clean. A failure is captured as an open outcome with diagnostics and no push.
8. Push and open the pull request, or report open
After Step 7 completes (local preflight reported, not gated), push the committed current branch and use the repository door:
git push -u origin <branch>
make pr-open HEAD=<branch> MESSAGE=<message-file> # deposit PRs: no AUTO_MERGE — the review triplet must return first
# The message file's first line is the PR title; the rest is the PR body.
AUTO_MERGE=1 explicitly opts an invocation into auto-merge; without that option, the door does not arm auto-merge. Deposit lanes must NOT pass it: the sshx review triplet (tests / quality / architecture) runs after the PR opens and must be able to block the merge (memory review-must-precede-automerge, #2337 precedent). After it opens the pull request, do not push further changes to that branch: the pull request may already have merged, in which case a later successful push does not put that commit on dev. Any further change requires a new branch and a new pull request.
If the dispatched sandbox forbids git writes, state that constraint explicitly and hand the exact git push and make pr-open invocations above, with substituted arguments, to the caller; do not report success as though the work landed. Otherwise report success only with the opened pull request, touched paths, door-produced commit subjects, every relevant exit code, completed fidelity-gate evidence, and, for every public theorem, proof_shape (content | bind-only), its direct frozen public dependencies (GID + statement_id), its escape_witness (or bind-only) and the module's admission_basis. Or report open, naming the stopping step and carrying every evidence class reached; mark each unreached class not run and explain why. There is no third outcome.
Postcondition: the task ends with an opened pull request, or with evidence-complete open.
Fidelity and non-hollowness gate
Before Step 7, the producing seat must answer every item with concrete evidence. This checklist collects producer-side evidence; it is not machine-verified and does not itself prove non-hollowness. Independent adversarial review, not producer self-attestation, turns this evidence into verification.
- Conclusion substance: show that the conclusion is not
True, not definitionally equal toTrue, and not a restatement of a hypothesis. - Hypothesis satisfiability: exhibit a Lean term witnessing the hypotheses that elaborates in the pinned toolchain, such as a checked
examplein the module or a term the seat states and checks, and carry that term in the report. Prose asserting or naming a witness does not discharge this item; if no compiling witness can be produced, the outcome isopenand deposit is blocked. - Domain inhabitance: exhibit a Lean term inhabiting the domain that elaborates in the pinned toolchain, such as a checked
examplein the module or a term the seat states and checks, and carry that term in the report. Prose asserting or naming an inhabitant does not discharge this item. - Proof substance: show that the statement carries content beyond unfolding a definition the producing seat itself introduced, whatever tactic closes it.
- Deposit substance: for a module introducing a new definition, name the theorem(s) that make the definition earn its freeze (instantiation against existing repository structures, an independent-sides characterization, or a citable property), and show the module's connection to the existing machinery its vocabulary names. Anonymous
examples do not discharge this item; see the thin-deposit taxonomy above. - Duplicate search: cite the Step 3 trace showing this is not a renamed duplicate of a mathlib or
D5/declaration. - Clause fidelity: place the authoritative atom clauses beside the Lean clauses one-to-one, mapping every clause to an exact Lean binder, hypothesis, or conclusion. The dropped-or-weakened set must be empty; any weakening, omission, or unresolved ambiguity forces
openbefore deposit. - Rendered-statement fidelity: read the emitted Blueprint
.mdfor this document and compare its displayed statement against the Lean declaration symbol by symbol; use a neighbouring landed mirror as a shape check. The formula DSL and writer own tokens that can be valid LaTeX and structurally accepted yet mean something different from the theorem, soemitexiting 0 is not evidence that the rendering is faithful. A mismatch blocks deposit; resolve it againsttools/StrataLint.Scribe/Ast/FormulaDsl.csandtools/StrataLint.Scribe/Writers/LatexWriter.cs, or end the taskopen.
Finally, run the grader-trap checklist against your own work before sign-off — witness-vs-universal, instance-vs-general, conditional-vs-unconditional, pointwise-vs-operator, proof-internal-vs-addressable-statement, multi-clause residue names, mechanism-vs-outcome — and record for each either "not applicable" or how your statement clears it. Reviewers will run exactly this list; a mismatch you find yourself is a free fix, one they find is a blocked lane.
Any item without evidence blocks deposit. Mark an unverified fact exactly ASSUMED-UNVERIFIED; never replace measurement with hedging language. Compilation, deposit, and cover do not by themselves certify clause fidelity, so the evidence checklist remains mandatory.
Earned hard gates (precedent taxonomy)
Every entry below names a failure class that actually occurred in landed rounds of this repository. These are not hypothetical defenses; they are the recurring ways seats fail review. The owners named in "What this skill does not own" still win on current thresholds — this section tells you where lanes die.
Mathematical content
- Definitional tautology (the most common kill, seven landed cases). Never define a thing as the formula you then "prove"; never install the conclusion by definition; never prove an iff whose two sides you defined to coincide. Both sides of every equation and iff must have independent anchors: a frozen declaration, a mathlib declaration, or data present in the authoritative atom text.
- Invented-classifier variant. Defining your own key or equivalence (for example a
Setoid.kerquotient of a tuple you chose) and then proving bijectivity or a count over its quotient proves nothing: the result is true by construction for any injective key, and the arithmetic coordinates you computed go unused. If the subject's own orbit/class notion is not in the repository, the honest outcome isopennaming the missing carrier — reviewers reject the invented one 2-1 or worse. - Fabrication ban. If a concrete datum your theorem needs — a case list, a walk, a witness pair, the meaning of a symbol — is not derivable from the authoritative atom text plus in-repo frozen definitions, do not invent it. Exhibiting concrete numerals you computed yourself from in-repo definitions is derivation and is fine; conjuring a list the source only attests is fabrication. End
opennaming precisely the missing datum. - A stronger variant does not excuse a duplicate. Twice now a lane found the frozen theorem covering its residue leg and wrote a new module anyway, rationalizing the addition as "exposing hidden facts" or "removing a hypothesis" — a strictly stronger or nicer-shaped variant of an already-covering frozen statement. If the frozen statement covers the leg, the deliverable is the bind report; strengthening it is gold-plating unless a named consumer needs the stronger form, and reviewers discard the module either way.
- Witness-vs-universal honesty. An
∃-witness does not discharge a universal claim; an instance (dim 4, a fixed modulus, an 18-ray set) does not discharge a general claim; a conditional theorem does not discharge its unconditional attestation. State exactly what you proved; graders run a named-trap checklist (witness-vs-universal, instance-vs-general, conditional-vs-unconditional, pointwise-vs-operator, proof-internal-vs-addressable-statement, multi-clause residue names, mechanism-vs-outcome) and a mismatch is a blocker, not a nuance. - Unused hypotheses are dishonest signatures. A binder the proof never uses gets stripped at collection; write the minimal true signature.
Artifact shape
No new top-level domain for one module (three landed rejections in one day). When your module's natural parent directory is at its file limit, the repository convention is a SPLIT BUCKET under that parent (
Parent/NewBucket/, split-only, no moves), recorded in the stratum's MAP file (split-history entry AND bucket-catalog line; create the stratum MAP and register it as a governance document if the stratum has none), with any new group name registered where the precedent registers it. Reuse an existing bucket whenever ownership genuinely fits (a symmetry theorem belongs in the existingSymmetry/bucket, not a new sibling). Registering a fresh top-level domain inMeta/domains.yamlfor a single module is the anti-pattern: every instance was struck down by review and relocated.FromLean is only legal for projector-registered shapes (five-statement landed rejection). A lane once used
StatementSource.FromLean()for all five of its Describes; the kernel projector returned Unprojectable for every one (Finset, a repo predicate,LT.lt,DirichletCharacter,Nat.Primeare all outside the registered shapes), the emitter exited 1, and no Blueprint mirror existed for the fidelity gate.FromAuthorwith a typed Disp formula is the DEFAULT; before even considering FromLean, verify the statement's every constituent is projector-registered — when in doubt, FromAuthor. Then run the emit and READ the produced.md: a missing mirror blocks deposit by itself.Header law (three landed violations; corrected 2026-09-06 for the
utility:line). The header sits at byte zero anddigest:is a SINGLE line ending-/— it is the LAST header line, not necessarily line 6. The owner isRepositoryRules.HeaderPattern, whose exact shape isGID/generality: [GIE]/mirror-B/mirror-E/anchors: [...]/ optionalutility:/digest: … -/. So the header is SIX lines withoututility:and SEVEN with it, andutility:sits betweenanchors:anddigest:— that position is fixed by the pattern, not a convention. A wrapped digest is a violation;-/alone on its own line is the same violation. Copy the shape from a CURRENT tracked neighbour (an older neighbour may predateutility:), then verify your file's last header line ends with-/.A
FromLiteratureDescribe drags in a Library note, and that note has a machine-checked shape (incomplete-library-locator, RED; landed 2026-09-06). Whenever a Describe carriesAssessedProvenance.FromLiterature(<bibkey>)(or an acknowledgement reference), the note atLibrary/<Area>/<bibkey>.mdis pulled into the check. Owner:tools/StrataLint.Scribe/Verification/DescribeContentGovernance.cs:212-237. It is a RED insideDESCRIBE_STATUS, so it fails theCanonical Lean report productionrequired check — not the engineering one, which is where you would look first if you only read the check name. The three conditions, verbatim from the source: the note text must contain\n## Verified locator\n(or\n## Locator\n); the section body, taken up to the next\n##, must be non-blank; and that body must literally contain the frontmatterdoi:value (case-insensitive). Putting the DOI only in the frontmatter is not enough — the frontmatter is what the checker compares against, not what satisfies it. Frontmatter is a closed alphabet (LibraryNoteCatalog.cs:45-53): exactlybibkey,authors,year,title,doi,claim,strata_touched,license,triage. Required lines:bibkey,title,authors,claim,license,triage, andyearas an integer in 1000–9999. Copy a current tracked neighbour in the sameLibrary/<Area>/and change the values. Minimal passing tail (this exact shape landed):## Verified locator - DOI: https://doi.org/10.2307/1990319Landed precedent (2026-09-06, #5827). A deposit whose Scribe used
FromLiterature(AlaogluErdos)shipped a note withdoi:in the frontmatter and no## Verified locatorsection.DESCRIBE_STATUScame backred=1 observe=3865— the 3865OPEN projectionlines are pre-existing observations, not the failure; the single RED was this locator. Cost: one full CI round trip. Verify the three conditions mechanically before you push, not by eye.utility:is mandatory on a first-freeze module and is machine-judged (SL-031, merged 2026-09-06, Block, Active). Scope: a frozen-state path that the candidate ADDS (present inCurrent, absent from the protectedBaseline) — i.e. exactly the first-freeze case of section 5⁗. Absent line, or the placeholderEDIT-ME, yieldsUTILITY-MISSING module=<path>; the finding carries no explicit effect and therefore takes the rule's registered default, andRegisterdefaults toAdmissionEffect.BlockwithRuleLifecycle.Active. Not a warning. Grammar (ownerUtilitySyntax.cs; fields are separated by exactly"; "):utility: none— the whole declaration, for a module with no computational content. Legal and common; five modules on dev use it. It is NOT a way to skip the judgment: review still asks whethernoneis honest (section 5⁗ 用途准入 is soft there).utility: kind=<K>; basis=<B>[; instance=<GID>][; premises=<GID>[,<GID>…]][; result=<GID>]—kindandbasisMUST be the first two fields in that order; the optional three, when present, must appear in the orderinstance,premises,result; 2–5 fields total, no duplicate keys.<K>∈bounded-enumeration|checker|numeric-reduction|certified-instance.<B>∈consumer=<declaration GID>|refutes=<target>|terminal=<target>, where<target>∈atom:<atom_id>|task:<TASK code>|gid:<declaration GID>.kind=checkerREQUIRESinstance=;kind=numeric-reductionREQUIRESpremises=AND a basis ofconsumerorrefutes(terminalis refused for it). Landed example on dev (b00aeab4a8):utility: kind=certified-instance; basis=terminal=atom:271aa869d74ec58b…. Landed precedent for ignoring this (2026-09-06, #4971 closed under 第 12 条). A deposit branch written 2026-09-03 carried two first-freeze modules withutility:line count 0. It had three other blockers as well — its freeze output wrote onlyGolden/Frozen/accepted/**and noGolden/Frozen/state/**slice (so it does not even match the current deposit contract, and note SL-031 keys on the state path, so it would NOT have fired on that branch as it stood), a named engineering red (FormulaCorpusInventoryTests.InventoryAllLegacyLatexStatementsAndSyntaxFamilies), and a base 1118 commits behind. The single lesson worth carrying: redo the deposit on current dev, and theutility:line is then required before the state slice can land. Rebasing an old deposit branch is not enough.
Generality tag follows the weakest import and the module's nature (two landed blockers). A concrete-instance module (fixed modulus, fixed witness set) is
generality: I; a general theorem module isG; aGtag on a file importingI-level facts is a violation. Compare your nearest landed neighbors before writing the tag.Scribe formula rejection taxonomy (dozens of mechanical rejections; owner
FormulaDsl.cs/LatexWriter.cswins on current tokens):F.Idarguments are strictly alphabetic;D()takes one digit per argument (D(2,3)neverD(23));Spis required after macro and relation tokens (Neg,Neq,Vert,Lvert,Rvert,Forall,InMacro,ExistsafterNeg, …) beforeF.Id/Operatorname;StarnotAst;NegnotNot; theFormulaDslusings are required; private structural helpers in.scribe.cs(Call/Bound/Seqwrappers overFormulaDsltokens) are the landed convention (600+ Blueprint sources on dev use them) — keep them minimal, typed, and never hand-assemble LaTeX strings; the displayed formula must mirror every conjunct of the Lean statement — mirror-value swaps (two constants exchanged between clauses) are a landed reviewer catch, so read your emitted.mdvalue by value.
Deposit substance (thin-deposit taxonomy)
A landed pull request consisting of one generic one-line def plus two simp-trivial anonymous examples, with an empty pull-request body, is the canonical thin deposit — it froze notation, not truth. The gates below name what was missing.
- Encoding is not formalization. Strip your Lean statement of its prose-borrowed names and read what remains. If it is a generic set/logic triviality (
∀ x ∈ S, f x ⊆ {c}under an "observer" vocabulary), you have renamed the source clause, not discharged it. Before depositing a definition, answer in writing: what named theorem would make this definition earn its freeze? If no such theorem is in reach, the honest outcome isopenwith the definition offered as a sketch, not a deposit. - Definition-only modules are presumptively insufficient. A new
defmust ship with at least one NAMED theorem that either instantiates it against existing repository structures, characterizes it (an iff with independently defined sides), or proves a property that some later consumer can cite. Anonymousexamples are fidelity evidence, never deposit content: they have no GID, cannot enter coverage, and cannot be cited by anything. - No island modules. A module whose only relationship to the repository is its directory path is a second framework wearing repository vocabulary. If your definition names a concept the repository already has machinery for (observers, windows, channels, walks), it must import and connect to that machinery — an "observer" predicate that touches none of the existing observer declarations belongs to nothing and discharges nothing. If the concept genuinely has no repository counterpart yet, say so explicitly and justify why a floating generic definition is worth freezing now rather than when its first theorem arrives.
- The external-referee test. Ask whether the module would survive review as a standalone library contribution: a one-line definition with
simpexamples would be rejected anywhere as content-free. The cost of freezing and covering a deposit is justified by content, and "it compiles and is honest" is the floor for a report, not a reason to deposit. - Bind-only modules are thin deposits (escape witness). Scope: any descriptor path with no active Freeze event on the protected base for which the candidate adds the first Freeze event; classify the COMPLETE declaration set that event records — the
make depositanchor GID does not change the classification. For every public theorem, record in the Step 8 report:proof_shape: content | bind-only; its DIRECT frozen public dependencies as GID +statement_id(pinned-Mathlib declarations are not frozen dependencies, but a conclusion obtained from a Mathlib lemma by direct instantiation is still not a witness); itsescape_witness— one precisely stated new intermediate proposition whose proving declaration lies in the theorem's elaborated transitive constant-dependency closure, is NOT obtained from those dependencies by instantiation, projection or normalization, is neither definitionally equivalent to nor an alias/restatement of the theorem's conclusion, and lies on a LIVE derivation path to that conclusion (after ζ/β/ι reduction and removal of dead or projection-discarded proof terms;(And.intro h frozen).2-style smuggling is not a witness — if the conclusion is reachable by bind-only operations without it, it is not a witness); andadmission_basis. A witness may take two forms: (1) a distinct intermediate proposition satisfying all four conditions, or (2) the public conclusion itself when it is produced on the live proof path by a named non-bind-only computation or construction (adecide/norm_numdecision of a new numeric fact, an explicit witness construction, a new estimate chain) and the same counterfactual holds (bind-only operations cannot reach the conclusion without it) — do not manufacture an otherwise unneeded intermediate proposition. No witness ⇒bind-only;escape-witnessmay be the admission basis only when at least one public theorem is classifiedcontentwith such a witness. Source-line counts are non-load-bearing readings, never the criterion. A first-freeze module with no admission basis must not be deposited: stop beforemake depositand reportopenwith reasonproof_shape: bind-only. Each bind-only companion in an admitted module must name its atom/obligation and a directed edge with both endpoints and consumer→prerequisite direction (the companion is a prerequisite of the escape-bearing theorem, or the escape-bearing theorem is a prerequisite of the companion, or a pre-registered named use); unrelated bind-only declarations may not ride along. If the witness observed after the probe differs from the one preregistered, re-preregister a new version; never relabel after the fact. Landed precedent (2026-09-04, itemized): across five RH-route addenda (W-10…W-21, PRs #5177/#5186/#5207/#5219), W-12 (an integrability proof) was the analytic escape; W-18's helperactivePrimePowers_eq_empty_of_exp_lt_twoand the W-14/W-15 companionexists_supportRadiuswere small new lemmas whose sufficiency as witnesses is judged by the four conditions above while the obligations W-14/W-15/W-18 themselves arebind-only; W-10/11/13/16/17/19/20/21 were de-hypothesizations, projections,obtain ⟨Z⟩-then-one-line or compositions of frozen iffs with no admission basis. The owner judged the sequence 平推 and the fifth same-shape lane was stopped unlanded. - The pull-request body is evidence, not decoration. Carry the clause-mapping echo, the search trace, and the explicit honest-partial disclosure (what is asserted, what remains open) in the pull-request body. An empty body on a deposit pull request hides exactly the thinness these gates exist to catch.
Cover and quarantine (digestion-side precedents)
Every entry below names a failure class that actually occurred on 2026-09-05/06, each caught by independent review after all three required checks were green. The machine gates judge admission compliance; they do not judge coverage faithfulness.
- A cover door has four preconditions, not three. Beyond (1) the covering module is frozen,
(2) the atom is still under
residual-open, and (3) the atom is not multi-clause by the purely syntactic test, the covering GID's theorem must be referenced by its Scribe declaration:grep -c "<theorem>" "Blueprint/<module>.scribe.cs"must be at least 1. A zero there makes the door answerpartial-closed deletable=false gaps=scribe-declaration-reference-missing. Becausemake cover-batchis strictly sequential and aborts on the first failure, one such pair blocks every later pair in the batch. - A theorem that is strictly weaker than the atom is not a cover. Check the theorem's every
explicit hypothesis against the atom text and find the sentence that supplies it; a hypothesis the
atom neither states nor implies makes the bound statement weaker than the source claim. Landed
case: an owner requiring
0 < regularizerwas bound to an atom that only gave p
…(truncated)