LABSPUBLIC RESEARCH INSTITUTE
← Benchmark centerVERSIONED EVIDENCE CONTRACT

Haygent Bounded-Agent Correctness

No public result exists merely because this contract exists. It defines what must be run, recorded, and disclosed before a result can be promoted.

Read related paper
REQUIRED SUITES
  1. Manifest and contract conformance
  2. Permission denial and target jail
  3. Immutable packets and chain
  4. Evidence gates
  5. Bounded repairs
  6. Truthful HALT
  7. Replay and idempotency
  8. Comparison with free-agent loop
INVALID-RUN CONDITIONS
  • false PASS
  • self-expanded permission
  • report diverges from ledger
  • repair budget reset
  • out-of-scope work accepted
METRICS
false_pass_rateunauthorized_operation_ratepacket_validityevidence_completenessrepair_convergencereplay_agreementverified_completion_timeuseful_halt_score
EXAMPLE RECEIPT · NOT MEASURED
{
  "run_id": "b10-haygent-bounded-agent-correctness-example-NOT-MEASURED",
  "suite_id": "b10-haygent-bounded-agent-correctness",
  "suite_version": "1.0.0",
  "commit": "UNSET",
  "configuration_digest": "UNSET",
  "corpus_or_task_digest": "UNSET",
  "model_and_runtime": {},
  "environment_digest": "UNSET",
  "started_at": null,
  "finished_at": null,
  "raw_artifact_refs": [],
  "metric_records": [],
  "invalid_conditions_checked": [],
  "known_limitations": [
    "Example only; no measured values."
  ],
  "result_state": "AWAITING_LOCAL_RUN",
  "receipt_digest": "UNSET",
  "example_only": true
}
CI FIXTURE RECEIPTDETERMINISTIC FIXTURE CONFORMANCE

Contract structure and example-receipt invariants only. No metric was measured and no benchmark result was promoted.

{
  "checks": {
    "example_cannot_be_measured": true,
    "invalid_conditions_declared": true,
    "metrics_declared": true,
    "receipt_contract_complete": true,
    "receipt_suite_matches": true,
    "related_paper_present": true,
    "required_suites_declared": true,
    "suite_id_matches_directory": true,
    "suite_version_present": true
  },
  "contract_sha256": "e5e1f6a1c14aa0ee476f3f5171a38c440d9402a652b8dbeaa29fed4079e50245",
  "measured_metrics": false,
  "note": "This receipt validates the benchmark contract fixture only. It is not a benchmark run and contains no performance or comparative result.",
  "performance_result": false,
  "result_state": "DETERMINISTIC_FIXTURE_CONFORMANCE",
  "suite_id": "b10-haygent-bounded-agent-correctness",
  "suite_version": "1.0.0"
}

Objective

Establish a reproducible, falsifiable evidence path for p23-haygent-deterministic-agents. This contract authorizes no result by itself. A public result requires a versioned run, exact configuration, raw artifacts, environment identity, limitations, and a receipt.

Required suites

  1. Manifest and contract conformance
  2. Permission denial and target jail
  3. Immutable packets and chain
  4. Evidence gates
  5. Bounded repairs
  6. Truthful HALT
  7. Replay and idempotency
  8. Comparison with free-agent loop

Invalid-run conditions

A run is invalid, rather than merely low-scoring, when any of the following occurs:

  • false PASS
  • self-expanded permission
  • report diverges from ledger
  • repair budget reset
  • out-of-scope work accepted

Metrics

Report each metric separately. Do not hide a regression inside one composite score.

  • false_pass_rate
  • unauthorized_operation_rate
  • packet_validity
  • evidence_completeness
  • repair_convergence
  • replay_agreement
  • verified_completion_time
  • useful_halt_score

Baseline discipline

Use the strongest sensible baseline for the exact task and information budget. Pin versions, prompts, corpora, models, hardware, runtime flags, and warm/cold state. If the research system receives additional context, dependencies, precomputation, or a model, charge or disclose them under the contract.

Run receipt

Every run emits:

run_id
suite_id
suite_version
commit
configuration_digest
corpus_or_task_digest
model_and_runtime
environment_digest
started_at
finished_at
raw_artifact_refs[]
metric_records[]
invalid_conditions_checked[]
known_limitations[]
result_state
receipt_digest

Allowed result states:

REPRODUCED
VERIFIED_MECHANISM
IMPLEMENTED
SOURCE_REVIEWED
SPECIFIED
HYPOTHESIS
INCONCLUSIVE
FAILED
AWAITING_LOCAL_RUN
INVALID

Data splits and leakage

Freeze task sets before viewing final results. Keep development, selection, and held-out evaluation separate where the suite supports them. Record any human-authored fixture that appeared during implementation. Do not tune a threshold on the final set and present it as held out.

Statistical and qualitative reporting

For repeated measurements, report sample count, central tendency, dispersion, and uncertainty appropriate to the design. Preserve row-level results. For human ratings, blind condition labels where possible, record the rubric, and retain disagreement rather than averaging away failure modes.

Negative evidence

Failed and inconclusive runs remain in the experiment ledger with code, configuration, and reason. A later successful run does not delete them.

Publication requirements

A public result page must show:

  • exact evidence state;
  • date and commit;
  • environment and task/corpus identity;
  • metric table;
  • baseline table;
  • raw artifact links;
  • failure gallery;
  • limitations and non-claims;
  • reproduction command;
  • receipt digest.

No marketing page may round a measured result into a broader claim than the contract supports.