idkmesh

R1 Real-Result Replay

Issue: #30
Purpose: Bridge the synthetic R1 experiments to actual IDKMesh worker results without creating a second competing result format.

Inputs

The replay adapter consumes the existing Phase 0 contracts:

The Phase 0 schema validator remains the authoritative full JSON-Schema validator. randomness_lab.r1_replay performs additional R1-specific normalization and conservative evidence filtering after schema-valid records are produced.

Trust boundary

A ResultManifest is not truth. The replay code does not infer correctness from:

A candidate becomes conclusive for R1 only when at least one VerificationResult declares independence.independent_from_worker: true and the available independent evidence resolves to acceptance or rejection.

Conclusive good

A candidate is counted as independently verified good only when:

Conclusive bad

A candidate is counted as bad when any independent verifier supplies a clear rejection signal:

Inconclusive

No independent verifier, conflicting evidence, escalation, or insufficient evidence is excluded, not guessed. The report publishes exclusion counts.

This is intentionally conservative: unknown is not converted into failure or success merely to make a comparison easier.

Structural diversity signature

R1 needs to compare repeated instances of one structure with candidates from different structures.

By default a structural signature is derived from the worker metadata:

worker.type
+ worker.adapter
+ model.provider
+ model.name
+ model.version

A real experiment can override this with:

{
  "extensions": {
    "r1_structural_signature": "planner:gpt-x|coder:model-y|tools:test-first"
  }
}

The override exists because meaningful structural diversity may include prompt strategy, role composition, toolchain, organization, or orchestration topology that cannot be inferred from the base v0.1 worker fields.

The replay report records both the signature and how it was obtained.

Fixed-budget comparison

For swarm size N, a work unit is eligible only when it has:

  1. at least N conclusive candidates from the baseline structural signature; and
  2. at least N distinct structural signatures with conclusive candidates.

Each bootstrap sample compares equal candidate counts:

replication:
  N candidates from the baseline signature

structural diversity:
  N distinct signatures
  one candidate sampled from each

This prevents an apparent diversity advantage caused solely by giving the diverse condition more attempts.

If --baseline-signature is omitted, the adapter chooses the signature covering the most work units with at least N replicas, then the most total candidates, with a deterministic tie-break. For a serious experiment, record and normally specify the intended baseline explicitly.

Real metrics

The replay computes from actual supplied records:

Missing resources are not zero

If a selected candidate omits compute_units or human_minutes, the cost-based metric for that bootstrap trial becomes unavailable. The replay does not silently treat missing resource measurements as free work.

--human-minute-cost-weight is an explicit analysis coefficient in:

normalized cost = compute_units + weight * human_minutes

It is not a currency conversion or economic claim. Reports must preserve the configured value.

Bootstrap and classification

Eligible work units are resampled with replacement. Candidate selection inside each work unit is also sampled reproducibly from the recorded candidate pools.

For each bootstrap trial the replay stores signed deltas:

structural diversity - replication

For verified success and, where resource data is complete, verified utility per normalized cost.

The empirical 2.5th and 97.5th percentiles form a descriptive bootstrap interval:

interval entirely > 0 -> helps
interval entirely < 0 -> hurts
otherwise              -> uncertain

Raw bootstrap trials remain in the report.

Failure correlation

For each pair of structural signatures, the replay reports:

This is observational. A low measured correlation can make diversity more useful, but correlation itself does not prove causal independence.

Run

Inputs may be JSON arrays, single JSON objects, JSONL files, or directories recursively containing .json/.jsonl files.

python -m randomness_lab.r1_replay \
  --results path/to/result-manifests/ \
  --verifications path/to/verification-results/ \
  --swarm-size 2 \
  --trials 500 \
  --seed 42 \
  --baseline-signature 'agent|adapter|provider|model|version' \
  --output results/r1-real-replay.json

Before replay, validate records through the Phase 0 schema workflow/harness.

For a BenchmarkCohort-backed real experiment, also run the fail-closed readiness audit documented in docs/research/R1_CORPUS_READINESS.md. The readiness command validates the cohort/evidence bindings and checks the held-out task count, exact signature budget, independent-test coverage, negative retention, and cost completeness without changing this replay’s decision rules.

What replay can and cannot establish

Replay is stronger than synthetic probability assumptions because it uses measured outcomes from actual candidate artifacts and actual independent verification records.

But it remains observational. It can still be confounded by:

Therefore replay should be followed by a prospectively designed, held-out experiment where candidate generation budgets, task assignment, and verification policies are fixed before results are seen.

Minimum useful real corpus for R1

For swarm size 2, a practical first corpus should include multiple held-out work units where each has:

The project should publish the normalized replay report and the underlying non-sensitive evidence/provenance needed to reproduce it.

Next step

Collect the first real R1 corpus from bounded coding tasks produced by the Verified Swarm Runner / Phase 0 harness, validate the manifests, run this replay unchanged, and publish the resulting help/hurt/uncertain map. Do not tune the replay rules after seeing the benchmark outcome without recording the change as a new analysis version.