Verdicts and routing states

Penrose separates an evidence judgment from an inability to test. That distinction is essential.

Research verdicts

Verdict	Meaning	What it does not mean
kill	The adequately powered claim failed one or more falsification gates.	The underlying market can never exhibit the effect.
underpowered	Available evidence cannot resolve a realistic effect from noise.	The claim is false.
watch	The claim survived important gates but remains provisional or capped, including a cost_contingent maker-only survivor.	Trade it.
research-supported	The claim survived the full configured process, including confirmation requirements.	Guaranteed profit or permanent truth.

Two current outcome reasons deserve explicit treatment:

• no_oos_edge is a kill: the out-of-sample score is below the configured decision band after search deflation.
• cost_contingent is a watch cap, not a clean kill: the claim clears the 2× cost margin only at maker cost, not at the authoritative taker cost, and remains capped pending forward fill validation.

Execution costs are venue-aware and all-in, resolved through the frozen, versioned config.EXECUTION_COST_SCHEDULE. Taker is the default because a daily bar cannot prove a passive fill. A maker assertion therefore adds an availability hazard rather than earning the cheaper price.

Routing states

State	Meaning	Operator response
duplicate	The same claim already has an adjudication; this is provenance, not a verdict.	Read the named prior decision or use --force to re-run. It never enters adjudication counts.
needs_data	Required data is unavailable or cannot satisfy the contract.	Acquire the named series or narrow the claim.
needs_implementer	No trusted executor currently implements the claim. This is the current primary state.	Implement or review the generated specification.
pending_module	Legacy alias for an implementation still awaiting a trusted module or review.	Prefer needs_implementer in new integrations.
needs_kind_review	The extracted claim kind is ambiguous, so adjudication is held.	Resolve whether the claim is a strategy or relationship before running it.
cannot_replicate	The available reconstruction is not faithful.	Correct the spec or provide code-complete logic.
needs_review	Binding, plausibility, or another integrity condition is ambiguous.	Review; never treat as a survivor.
engine_error	The system failed before producing a trustworthy judgment.	Diagnose and rerun after fixing the engine.
off_domain	No supported quantitative performance claim was found.	Reformulate only if a real performance claim exists.
cannot_confirm	Confirmation lacks adequate power or eligibility without spending the reserve.	Acquire a valid confirmation cohort.

Precedence

Structural integrity failures take precedence over flattering statistics. A Sharpe ratio of 10 does not rescue an unfaithful reconstruction, a venue mismatch, look-ahead, or an implausible payoff. Such a run routes to review or cannot-replicate; it does not receive a positive verdict.

Reading counts

Always distinguish:

• active-epoch decisions - results produced under the current corpus contract;
• legacy decisions - preserved historical results from prior epochs;
• physical rows - every version and supersession marker in the append-only ledger;
• eligible corpus evidence - the smaller subset permitted to inform learning.

These numbers are not expected to match.