From ad6aa1523dd853cee74c80a76552463fe237cac6 Mon Sep 17 00:00:00 2001 From: Dylan Roy Date: Tue, 6 Oct 2026 14:09:35 +0000 Subject: [PATCH] Add closure-evidence protocol: blind reconstruction, adversarial attack, derived profiles Adopts the useful parts of the closure-superstructure proposal (decision record main#275) without a new status word, register or directory: two new review bases in the review form, a short protocol for blind reconstruction, adversarial attack, numerical reproduction of headline constants, novelty refresh and the human referee packet, and a derived, never-stored evidence profile. Links from the current workflow and the workspace guide. Scientific effect: NONE. Co-Authored-By: Claude Opus 5.5 Claude-Session: https://claude.ai/code/session_017Mi3hxjaxV45x6zo6o1ee3 --- .github/ISSUE_TEMPLATE/review-record.yml | 3 + docs/WORKSPACE.md | 8 ++ governance/OP-CLOSURE-EVIDENCE-20261006.md | 101 +++++++++++++++++++++ governance/OP-WORKFLOW-20260930.md | 3 + 4 files changed, 115 insertions(+) create mode 100644 governance/OP-CLOSURE-EVIDENCE-20261006.md diff --git a/.github/ISSUE_TEMPLATE/review-record.yml b/.github/ISSUE_TEMPLATE/review-record.yml index 5f57d8300..db29f8c31 100644 --- a/.github/ISSUE_TEMPLATE/review-record.yml +++ b/.github/ISSUE_TEMPLATE/review-record.yml @@ -16,8 +16,11 @@ body: id: review_basis attributes: label: Basis of this record + description: Blind reconstruction and adversarial attack follow the closure-evidence protocol (governance/OP-CLOSURE-EVIDENCE-20261006.md). options: - Fresh source-bound technical review + - Blind reconstruction — statement packet only, proof not read + - Adversarial attack — counterexample search - Execution or custody audit - Digest of previously recorded reviews validations: diff --git a/docs/WORKSPACE.md b/docs/WORKSPACE.md index e7b1f54c1..8e39b86ed 100644 --- a/docs/WORKSPACE.md +++ b/docs/WORKSPACE.md @@ -55,6 +55,14 @@ from the relevant work discussion and this guide. The requires affected work owners' acknowledgments before any freeze; it is not a completed audit or declared freeze. +## Closure evidence + +The [closure-evidence protocol](../governance/OP-CLOSURE-EVIDENCE-20261006.md) +adds blind-reconstruction and adversarial-attack review bases and derived, +never-stored evidence profiles. Its +[decision record](https://github.com/d6g8k5htny-coder/main/issues/275) explains +which proposal items were already present, adopted or declined. + ## Tools and execution All participating models have Dylan's permission to download, install, create, and diff --git a/governance/OP-CLOSURE-EVIDENCE-20261006.md b/governance/OP-CLOSURE-EVIDENCE-20261006.md new file mode 100644 index 000000000..b9a792a8e --- /dev/null +++ b/governance/OP-CLOSURE-EVIDENCE-20261006.md @@ -0,0 +1,101 @@ +# Closure evidence: blind reconstruction, adversarial attack and derived profiles + +On 6 October 2026 Dylan forwarded a proposal for a "theorem-closure +superstructure" and pre-approved the team's decision about it. The +[decision record](https://github.com/d6g8k5htny-coder/main/issues/275) maps all +thirteen items onto existing machinery. This protocol adds only the parts that +record did not find here already. It extends the +[current workflow](OP-WORKFLOW-20260930.md). **It starts no review, executes no +register transition and changes no scientific status.** + +## Evidence profiles are derived, never stored + +Show how far a result has been checked as a profile computed from records the +project already keeps: + +| Axis | Read from | +|---|---| +| Source-bound argument and nonauthor review | The result's review records and its `PROVED_REVIEWED` graph node | +| Dependency closure | The downstream hard gate: every required dependency satisfied | +| Provider-distinct review | Recorded author and reviewer providers | +| Blind reconstruction | A review record with that basis (below) | +| Adversarial attack | A review record with that basis (below) | +| Formal evidence | Kernel-checked gate receipt and separate accepted alignment review | +| Numerical reproduction | A second implementation recorded in the packet's review record | +| Novelty | The existing discovery/priority ledger row | +| Human reading | Dylan's personal reading, or an independent human review | + +An axis without a record reads `not recorded`, never a pass. A profile is a +view: it promotes nothing, is not committed as a status file and adds no status +word. When an axis later fails, report a finding; the existing loss-only +transition rules decide what is held or revalidated. Proof bytes and earlier +verdicts are never edited to match a profile. + +## Blind reconstruction + +The packet contains the statement, hypotheses, definitions, notation and an +allowed-literature list. It contains no proof steps, intermediate lemma names or +review text. Anyone may prepare it; record its exact identity. + +The reconstructor must not have read the proof or its reviews. Shared accounts +and repository access make this hard to guarantee, so state exactly what was +available. The credible form is a fresh session with no access to the project's +repositories or files. Record that exposure in the review form. The +reconstructor publishes a complete derivation, or the precise point where it +fails, **before** seeing the original. + +A third participant then compares the two dependency graphs. Record whether the +mechanisms converge, whether the routes differ, and any gap. A reconstruction +that reuses an original lemma must justify it independently. A blind +reconstruction earns no organizational-independence credit by itself; provider +and account rules still apply. + +## Adversarial attack + +Name the exact target (repository, commit, path and blob) and assume it is +false. Look for the weakest counterexample or failure. Check at least: missing +hypotheses, non-uniform error terms, illegal limit interchange, boundary +contributions, degeneracies, multiple witnesses, candidate/elder confusion, +conditioning errors, hidden independence assumptions, exceptional +configurations, scaling regimes where the asymptotic changes, and +dimension-dependent failure. End the record with: + +```text +ATTACK STATUS: counterexample found | counterexample not found +Target: +Attempts: method -> result> +Remaining attack surfaces: +``` + +"Not found" is not acceptance. A counterexample is a finding for the owner; it +is not a register edit. The attacker should not be the target's author, and the +record states the attacker's prior exposure. + +## Numerical reproduction for headline constants + +For headline constants, such as the SIDE24 coefficients, `C_U(1)` and certified +tail constants, a second implementation uses a different method and is written +by a nonauthor lane. Record input identities, code commit, runtime, precision +and the resulting enclosure in that packet's review record. Disagreement beyond +the stated enclosures is a finding. Running the same program twice is not a +second implementation. + +## Novelty and the human referee + +Novelty audits amend the existing +[discovery/priority ledger](../docs/DISCOVERY_PRIORITY_LEDGER_20260925.md). +Use two independent searches, each told to find an earlier proof rather than +support. Its wording rule is unchanged: "to our knowledge", never "first". + +A referee packet runs to 20 pages at most: the theorem, definitions, the critical +lemmas, a dependency diagram, an assumption ledger, the open interfaces, the +certificates and any blind reconstruction. It asks one question: *identify the +earliest unsupported inference*. Choosing and contacting a human referee is +Dylan's decision. Delegated AI review never fills a human-review requirement. + +## Not adopted + +No `CLOSED` status, passport file, or new `closure/` or `reproduction/` +directory. These would duplicate the scientific-status register, which +[AGENTS.md](../AGENTS.md) and the current workflow forbid. Results stay in the +existing review, certificate, coefficient, formal and graph locations. diff --git a/governance/OP-WORKFLOW-20260930.md b/governance/OP-WORKFLOW-20260930.md index 32842fb0d..ead2832aa 100644 --- a/governance/OP-WORKFLOW-20260930.md +++ b/governance/OP-WORKFLOW-20260930.md @@ -71,6 +71,9 @@ argument. Changed reviewed bytes, dependencies or scope invalidate the affected review and require a successor disposition. Regenerating an inventory cannot transfer an old review to new proof bytes. +Blind reconstruction, adversarial attack and derived evidence profiles follow +the [closure-evidence protocol](OP-CLOSURE-EVIDENCE-20261006.md). + Record actual author/reviewer exposure and separate technical verdict from organizational independence. Same-provider review can be useful technical work but earns zero organizational-independence credit. An author replay is not a