Date: 2026-08-28
User request: continue strengthening MSKazemi/idkmesh with a more solid mathematical/algorithmic background and implement the result through GitHub-native automation.
The preceding Anytime Drift Guard merged in PR #212 as:
32a8848b85b108154b0837ba2a550be5c86891ee
Its canonical default-branch push workflow 33204353541 completed successfully on that exact commit. The same push’s Evolution Loop run was cancelled by the repository’s concurrency lineage rather than failing a test.
Before starting the next layer, main advanced once more through maintenance PR #215 to:
528690273097a4abc926052eec52dba6cde4f28b
A direct GitHub branch read still reported:
protected: false
required status-check enforcement: off
The new branch codex/adversarial-evidence-envelope was created from that exact canonical commit.
Repository code search found no implementation of:
max_faults evidence certification.PR search did find adversarial calibration work (for example inert/Goodhart decoys) and earlier R4 text explicitly warning that contributor-level routing would need Sybil/collusion analysis. That is complementary rather than duplicate: adversarial test fixtures check whether one verifier/evaluator is Goodhartable, while this continuation asks what can still be concluded when some accepted reports themselves may be arbitrary.
scripts/evolution_math.py already contains bayesian_vote_posterior(), which weights votes by declared reliability and discounts same-group reviewers using an equicorrelation effective-sample-size factor.
That model is probabilistic and explicitly says it is not proof of independence. It does not provide a worst-case guarantee if a bounded number of accepted reports are arbitrary.
The new layer therefore stays orthogonal:
correlation model:
how much probabilistic evidence should correlated/reliable reports contribute?
count-contamination model:
what can be certified even if <= f accepted reports are arbitrary?
Neither model is treated as a substitute for the other.
For sorted scalar reports
x_(1) <= ... <= x_(n)
and a declared maximum of f < n arbitrary reports, the unknown honest set has size at least n-f.
Define
L_f = mean(x_(1), ..., x_(n-f))
U_f = mean(x_(f+1), ..., x_(n)).
Then every admissible honest-set mean lies in
[L_f, U_f].
The interval is sharp given only the observed reports and count bound: each endpoint is attained by an admissible assignment that marks the opposite f extremes faulty.
This gives a deterministic, distribution-free report-level certificate. It makes no independence assumption, but it also makes no external truth claim.
For threshold theta and reviewed margin gamma:
support_certified iff L_f > theta + gamma
reject_certified iff U_f < theta - gamma
otherwise observe_adversarial_uncertainty
The module also compares this robust certificate with the ordinary mean. If the naive mean crosses the threshold but the sharp adversarial envelope does not, the output records:
naive_decision_fragile = true
This exposes sensitivity rather than hiding it in a point score.
Hard governance remains conjunctive:
strong robust certificate + hard_guard_ok=false -> guarded
Binary support/reject votes use the same sharp envelope. A strict honest support majority is certified exactly when the lower possible honest support fraction exceeds 1/2.
The implementation additionally records whether at least one honest support/reject is guaranteed under the fault budget.
Because f is a threat-model parameter rather than observed truth, the module emits a sensitivity curve over fault budgets. The certified interval width is required to be non-decreasing as f increases.
While this branch was being assembled, main advanced again through PR #203 to:
bc5b15541963d1ed2f9ce812f1d7f291628d0c2d
That PR adds deterministic validation and descriptive scoring for an individual human review session plus provenance/independence disclosures. Its protocol explicitly preserves disagreement and does not interpret agreement as correctness.
This is complementary to the adversarial envelope:
review-session validator:
is one submitted review structurally/provenance-valid and what descriptive metrics does it contain?
adversarial evidence envelope:
after several accepted scalar/binary reports exist, what honest-report mean range is guaranteed if <= f reports are arbitrary?
The new human-review files do not overlap this branch’s five files and do not implement a multi-review fault-budget aggregator.
The implementation intentionally emits false claims for:
truth_claim
sybil_resistance_claim
byzantine_consensus_claim
The fault budget counts accepted reports. If one adversary can create unlimited accepted identities, the assumption can fail. Network consensus, equivocation, partitions, identity admission, and contributor governance therefore remain separate problems.
The implementation uses explicit left-to-right floating accumulation rather than Python’s version-dependent built-in float sum() behavior. This carries forward the reproducibility lesson discovered by PR #212’s first failed cross-version run.
The companion workflow is designed to run Python 3.11 and 3.13 in one contents-read job, execute the adversarial invariants, generate deterministic JSON twice, compare the payloads byte-for-byte, and publish an explicit content SHA-256.
No issue/PR mutation, merge authority, repository setting change, secret use, or compute activation is introduced.