You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Proposal: per-repo kernel-receipt records so the census can show kernel results beside text matches #28
The census counts holes, axioms and unsafe markers by text matching (scripts/census/proof-holes.sh). The catalogue says so: "textual candidates, not kernel verdicts". Two cases show where text matching and the kernel disagree:
hyperpolymath/ephapax had 15 "holes" in the catalogue. A Print Assumptions census of its whole Coq development finds 0: 257 of 257 named results are closed under the global context, and coqchk finds no ephapax axiom. fix(census): comment-aware proof-holes.sh, Idris2 undefined-declaration axioms; refresh 13 rows #27 closes most of that gap by making the scanner skip comments, but a text scan still cannot see a proof's assumptions.
PROOF-POLICY §8.4–§8.5 says grades are derived from receipts and that mathesis "may catalogue grades; it derives them from those receipts". Today mathesis has nowhere to put a receipt, so it cannot derive or show a single Proof Witness Grade, even ungraded.
Proposal
The home repo owns its proofs and evidence (§2.2, §8.5). Mathesis would only ingest a small, machine-readable kernel receipt from each repo that publishes one, and show it next to the text census. It would never assign anything itself.
A receipt is one I-JSON/JCS record per checker run, with:
repo and the full 40-character commit;
checker and its version (coqc 8.18.0, agda 2.6.3 + stdlib 2.3, lean, idris2);
invocation, including flags (for Agda, the §2.1 --safe --without-K --cubical-compatible);
results: the count of named results checked and the count that are closed or assumption-free;
axioms: the reported axioms by name, from Print Assumptions, #print axioms, or the Agda import-cone postulate scan;
independent_recheck: for example coqchk -o or agda --ignore-interfaces, with its own result;
result: pass or fail, plus a receipt state of current, superseded or regressed (§8.4).
Mathesis would then:
read receipts from a known path in each repo (to be agreed) or from a nomination PR;
add kernel columns to the catalogue (results closed / checked, axioms, recheck);
derive a PWG per theorem only where an Agda receipt and RECONCILED correspondence records exist, and otherwise show ungraded. That last case is everything today.
Worked example already available
hyperpolymath/ephapax#425 adds formal/scripts/no-false-gate.sh. It already produces every field above for Coq: a census of 257/257 closed results, the axioms reported, and a coqchk -o recheck. Emitting a receipt from it is a small change, and that would make it the pilot.
Open questions for the owner
Where receipts live: a fixed path in each home repo, or nomination PRs into mathesis.
Whether CI-produced receipts only count as PWG-1 evidence when they come from the pinned, headless runs §8.2 describes.
Whether the schema belongs in standards (for example next to proof-witness-grades.k9.ncl) rather than in mathesis.
One input for the receipt's checker and version fields: hyperpolymath/bofj-kitt is building the estate's proof-toolchain bundle.
tools/proof-toolchain/toolchain.kyaml declares one entry per prover, translator and solver, each with its PROOF-POLICY role.
Its planned lock pins each installed tool's version and the SHA-256 of its executable.
A receipt could name its checker by lock entry and hash, for example agda-2.6.3 with its stdlib entry. Mathesis would then not have to trust a free-text version string. It would also answer, for each transported leg, "which agda2dedukti, which Agda".
Status, so this isn't read as more than it is:
The manifest exists. The lock and installer (provers.sh install and doctor) are not built yet.
bofj-kitt#103 adds to the design that receipts cite the lock, and that the installer writes the PROOF-POLICY §6 deviation records for anything installed outside Guix or the container.
The lock is planned as toolchain.lock.json. Under JSON-POLICY it would be I-JSON/JCS, which matches this proposal's receipt format.
This is offered as a source to cite, not as a schema decision for Mathesis.
Closed by #68 (re-anchoring plan WO 2.2), on the owner's word. What happened to each part of this proposal:
The receipt format and where it lives. PROOF-POLICY §8.5.1 settles both. A receipt is a praxis-deed with :profile proof-receipt in the home repository (verification/receipts/*_praxis.deed). The canon reader is readProofReceipt in metadatastician/standards 1-formats/deed/tools/deed_lint.js.
Grades derived from receipts. Done. The claims registry pins the reader (commit be39f90 and the sha256 of the file), and the reader derives each grade. hyperpolymath/mathesis#THY-1 derives PWG-1. A receipt that the reader refuses shows as ungraded, with the reader's reason.
Kernel columns in the catalogue. Not added, by owner ruling (2026-11-11). Kernel results belong to claims, so the registry view carries them. The catalogue stays a per-repository text census, and it says so ("textual candidates, not kernel verdicts").
The open questions: receipts live in the home repository at a fixed path; the §8.2 rules for PWG-1 (current receipts from pinned CI runs on the exact commit) are in the reader; and the schema is in standards. Experiments E1 to E6 stay open in docs/proposals/proof-receipts-as-data.adoc.
The current receipts of hyperpolymath/secret-types and metadatastician/lintel do not yet pass the reader ("unexpected clause (identity) in the receipt"), so the registry pins none of their claims yet.
Activity
hyperpolymath commentedon Nov 9, 2026
One input for the receipt's
checkerand version fields:hyperpolymath/bofj-kittis building the estate's proof-toolchain bundle.tools/proof-toolchain/toolchain.kyamldeclares one entry per prover, translator and solver, each with its PROOF-POLICY role.A receipt could name its checker by lock entry and hash, for example
agda-2.6.3with its stdlib entry. Mathesis would then not have to trust a free-text version string. It would also answer, for each transported leg, "which agda2dedukti, which Agda".Status, so this isn't read as more than it is:
provers.sh installanddoctor) are not built yet.toolchain.lock.json. Under JSON-POLICY it would be I-JSON/JCS, which matches this proposal's receipt format.This is offered as a source to cite, not as a schema decision for Mathesis.
Generated by Claude Code
hyperpolymath commentedon Nov 11, 2026
Closed by #68 (re-anchoring plan WO 2.2), on the owner's word. What happened to each part of this proposal:
praxis-deedwith:profile proof-receiptin the home repository (verification/receipts/*_praxis.deed). The canon reader isreadProofReceiptin metadatastician/standards1-formats/deed/tools/deed_lint.js.be39f90and the sha256 of the file), and the reader derives each grade.hyperpolymath/mathesis#THY-1derivesPWG-1. A receipt that the reader refuses shows asungraded, with the reader's reason.The open questions: receipts live in the home repository at a fixed path; the §8.2 rules for
PWG-1(current receipts from pinned CI runs on the exact commit) are in the reader; and the schema is in standards. Experiments E1 to E6 stay open indocs/proposals/proof-receipts-as-data.adoc.The current receipts of
hyperpolymath/secret-typesandmetadatastician/linteldo not yet pass the reader ("unexpected clause (identity) in the receipt"), so the registry pins none of their claims yet.Generated by Claude Code