{
  "schema": "gyre-prospective-research/1",
  "status": "draft specification; not a registered or executed confirmatory trial",
  "purpose": "Specify evidence that could support or refute a stronger claim than the existing single-task feasibility pilot.",
  "candidate_primary_claim": "Under a fixed external model and all-in search budget, inherited whole-program search with co-evolving mutation lenses improves independently audited success relative to otherwise matched inherited search with a fixed lens.",
  "primary_estimand": "Paired difference in probability that a candidate selected and frozen before final audit satisfies both the original-use-case contract and the predefined extension or optimization criterion at budget B.",
  "secondary_estimand": "Benefit of inherited fixed-lens search relative to independent best-of-N proposals at the same budget.",
  "separate_operational_claim": "Content-addressed source frames improve inspectable reconstruction, corruption detection, lineage preservation, and controlled transfer relative to equivalent unframed state. This is not a claim that hashing improves solver intelligence.",
  "conditions_required": [
    {
      "id": "independent-best-of-n",
      "role": "Fresh proposals from the same task/initial program and fixed mutation policy. No earlier candidate source or cross-candidate feedback is given to later proposals. Selection uses development evaluation."
    },
    {
      "id": "fixed-lens-inheritance",
      "role": "Inherited program state, an explicitly specified archive/selection rule, and development feedback, with a strong fixed mutation policy tuned only on separate development tasks."
    },
    {
      "id": "evolving-lens-inheritance",
      "role": "Program and lens changes, charging meta-lens proposal/evaluation to the same total budget."
    },
    {
      "id": "inheritance-ablation",
      "role": "Remove or reset inherited state under a prespecified information policy, without falsifying any original lineage."
    },
    {
      "id": "equivalent-unframed-state",
      "role": "Match carried information and compare operational reliability rather than assuming a solver-performance advantage from the envelope."
    }
  ],
  "prompt_policy_factor": {
    "role": "Separate success-conditioned speculative wording from the search architecture.",
    "design_if_claimed": "Cross the three principal search architectures with conventional-repair and success-conditioned speculative policies, yielding six cells; otherwise limit the prompt-specific claim.",
    "cost_boundary": "Charge all calls and context to the relevant treatment. A lens-evolving condition does not receive free extra inference."
  },
  "lens_credit_assignment": {
    "requirement": "Compare evolved and baseline lenses on common, previously unseen parent programs with matched budgets.",
    "reason": "A lens can survive by being attached to a good program without producing better future offspring.",
    "controls": "Use prespecified reset or reassignment controls without falsifying original frame ancestry. Record each actual parent/lens intervention.",
    "accounting": "If these probes influence selection, include their cost and data access inside the search budget."
  },
  "measurement_requirements": [
    "Independent task/run replications across more than one family; do not count correlated siblings or case rows as independent trials.",
    "A task-development/calibration set separate from the locked primary evaluation set.",
    "Hidden or newly prepared task instances, explicit leakage controls, and fixed public task contracts.",
    "A protected external evaluator whose implementation and expected outcomes cannot be weakened by candidates.",
    "Original-use-case retention, new-task success, failure modes, cumulative performance curves, and transfer outcomes.",
    "Measured model input/output/cached/reasoning tokens where exposed, actual spend, retries, compilation, evaluator time, archive/context handling, and all meta-level costs; do not substitute only client request count for total compute.",
    "Record final audit and replication costs separately and in total, plus human setup/tuning and interventions.",
    "Predeclared practical effect size, sample-size justification, uncertainty estimates, exclusions, and missing-result treatment.",
    "Fixed stop rules, no optional stopping after a favorable result, and preservation of all failed or blocked attempts.",
    "A clean-process transfer phase with declared runtime/model dependencies and no undisclosed conversation state.",
    "Independent replication and adversarial re-evaluation before broad generalization."
  ],
  "lessons_from_the_pilot_review": [
    "Passing the frozen cases did not establish complete original-contract preservation.",
    "Include independently derived metamorphic properties such as leading-zero invariance before primary evaluation.",
    "Exercise worst-case permitted original inputs against the same execution/resource limits, not just their output values.",
    "Check interpreter semantics explicitly; do not call a restricted execution profile generally CPython-equivalent.",
    "Keep targeted post-hoc counterexamples separate from the original denominators and from unbiased arm comparisons."
  ],
  "disconfirming_outcomes": [
    "Matched independent attempts perform equally well or better.",
    "Resetting inherited state does not reduce performance on the stated task distribution.",
    "Lens evolution adds no benefit after its inference and evaluation costs are counted.",
    "New-task gains do not survive the original-use-case floor or a fresh-process transfer.",
    "Independent graders overturn the reported successes.",
    "Apparent gains depend on visible-test exploitation, undisclosed interventions, selective reporting, or changing the denominator."
  ],
  "registration_blockers": [
    "Executable task generators and evaluator implementations for the new study are not yet frozen.",
    "Exact model versions, resource ceilings, treatment schedules, and information supplied to each control are not yet fixed.",
    "A practical effect size and confirmatory sample size have not been justified from independent calibration data.",
    "The complete protocol has not been independently reviewed and publicly committed before primary data collection."
  ],
  "uncertainty_policy": "Justify the minimum worthwhile effect and sample size using paired or hierarchical simulations with separate calibration data. A wide interval is inconclusive; failure to find significance is not equivalence. Do not infer population confidence from many correlated case rows.",
  "replication_levels": [
    "Artifact replay: independently reconstruct the disclosed programs, decisions, and observed outputs.",
    "Fresh-search replication: an independent group repeats the registered comparison with new stochastic runs and separately authored audit tasks."
  ],
  "rule": "Do not label this draft as preregistered, and do not collect or advertise confirmatory results until every registration blocker is resolved.",
  "current_pilot_boundary": "The existing all-success matched-control tie cannot be retrospectively promoted into evidence for the stronger causal claim."
}