idkmesh

Project Turn: Real two-attempt convergence and benchmark cohort contract

Date: 2026-08-28
Repository: MSKazemi/idkmesh

User message

Okay, go ahead and continue.

Live repository reassessment

This turn began by checking the current repository state rather than assuming the previous controlled-Docker gate was still open.

The critical path had advanced:

The first implementation target was consequently the same task through two real worker attempts rather than another verifier protocol.

Independent real two-attempt implementation

A branch and PR #118 were built to compose existing canonical pieces:

exact accepted node SHA
 -> real attempt 1 -> independent VerificationResult
 -> real attempt 2 -> independent VerificationResult
 -> existing idkmesh-two-attempt-run shape
 -> existing non-selecting Evidence Report
 -> replay report from saved run metadata

PR #118 introduced no new worker/verifier/result protocol. Its real GitHub Docker workflow ran successfully on head:

8f86355f7822935e9f7bf29d3a6dab0ad57ddc97

Workflow run:

33187069657

Job:

98902714723

The job passed all real-runtime steps:

Convergence with concurrent PR #117

Before PR #118 could be merged, main advanced with PR #117:

Prove real two-attempt node evidence and fault isolation

merge commit:

35f427359b3bf38419c6028ac9d24a08e68269d4

PR #117 proved the same core real two-attempt path and additionally exercised a real worker process error while preserving the independently verified peer attempt in the combined run/report.

Therefore PR #118 was deliberately not merged. It was closed as successful corroborating evidence because merging a parallel integration bridge/workflow would create duplicate canonical surfaces.

This is repository convergence, not experiment failure.

Issue synchronization

Issues #4, #5, and #16 were updated with the merged PR #117 evidence and PR #118 corroboration.

The important state transition is:

single real attempt proof       -> complete
real two-attempt proof          -> complete
real worker fault isolation     -> complete
combined non-selecting report   -> complete
saved-run report replay         -> complete
first small benchmark cohort    -> now unblocked

PR #91 remains separately human/reviewer-gated. The real evidence loops intentionally use its exact accepted SHA without claiming that this separate review gate is satisfied.

Benchmark infrastructure gap

The repository tree was then inspected for a reusable benchmark-corpus contract.

Existing components include:

There was no general index representing:

ExperimentManifest should not be overloaded to carry these semantics because it answers a different question: experiment hypotheses/configurations/repetitions rather than corpus membership and evidence binding.

New contract: Benchmark Cohort Index v0.1

A new branch was created from the merged real two-attempt baseline:

feature/benchmark-cohort-contract-v1

The branch adds:

The cohort contract is deliberately a thin index, not a new task or verification protocol.

Pre-Outcome Definition Commitment

The main new idea is a content commitment over the benchmark definition before analyzed outcomes are attached.

The definition projection commits:

cohort identity
+ required families / minimum target size
+ task family / difficulty / split
+ immutable source snapshot
+ WorkUnit id/version/digest
+ EvaluatorPlan visibility/id/digest/backend
+ predeclared structural signatures
+ seeded-negative expectation
+ accounting requirements
+ no-auto-authority policy

Then:

D_definition = SHA256(canonical(pre_outcome_definition))

Evidence/result paths and outcomes are excluded from this projection.

Therefore adding later ResultManifest/VerificationResult evidence should not change the definition digest, while changing task labels, source revision, WorkUnit, evaluator, negative expectation, or structural-signature taxonomy must change it.

This is intended to reduce post-outcome relabeling, evaluator drift, and benchmark overfitting/Goodhart pressure.

The digest proves content identity, not chronology by itself. Git history, PR review, signatures, or a future transparency log are needed to prove that the commitment existed before final outcomes were inspected.

Cross-object validation

tools/benchmark_cohort.py validates the cohort schema plus canonical referenced objects.

For public evaluator plans it checks:

For verified attempts it additionally checks:

Hidden evaluators expose only a digest/id/backend commitment; their plan path is forbidden from the public index. A later VerificationResult must still bind to the hidden plan digest.

Seeded-negative rule

Every task defines a meaningful negative expectation before final evidence, such as:

With --require-evidence, an included task must have analyzed real attempts plus verified negative evidence. Excluded tasks retain explicit exclusion reasons rather than disappearing from the corpus.

First cohort direction

The first Phase B2 cohort remains intentionally small: 5–10 repository-level tasks across bug fix, test/failure reproduction, bounded feature, refactor/code consistency, and documentation/code-contract work.

That number is an engineering bootstrap target, not a statistical-power claim.

The same cohort index is designed to scale into:

The immediate next gate is CI/review of the cohort contract. After that, the project should instantiate the first real tasks rather than adding more synthetic infrastructure.