Scientific record
Epistemic label: NEW PROPOSAL
These human construct/signal gates operationalize the handover's preference for cheap falsification. No human or empirical capability gate has passed. Offline checker verification is separately scoped. Numeric thresholds, domains, modality, sample sizes, effect sizes, and budgets must be specified before the relevant data are collected or examined; this document does not invent them.
| Gate | Evidence required for continuation | Failure · modification trigger | Next rational decision |
|---|---|---|---|
| G0 — Construct and information boundary | A defensible transformation criterion, unit of observation, and documented external information constraint. | Definition admits paraphrase/randomness without meaningful transformation, or "fixed" conditions cannot be stated. | Narrow/split the target or redefine the boundary before detector work. |
| G1 — Label feasibility | Independent rubric, documented evidence access, reliability estimate with uncertainty, and distinct validity checks for process and product criteria. | Labels rely on telemetry/features, condition instructions, Aha alone, or superficial output cues. | Revise the evidence/rubric; consider dimensions/coarser windows. Do not train on invalid labels. |
| G2 — Directional local signal | Participant/session-disjoint evidence that candidate features add useful information beyond agreed nuisance/output baselines and negative controls. | Leakage, nuisance-only equivalence, unstable labels, or no useful incremental signal under adequate measurement. | Revise signal/instrument or stop the detector branch. An inconclusive small test is not proof of no capability. |
| G3 — Same-modality domain transfer | A model and threshold locked using Domain A, tested on untouched Domain B with clear domain difference and participant controls. | Transfer is absent, uncertain, or explained by shared cue/identity/difficulty distributions. | Narrow the transfer claim or redesign; retain a runtime proposal only with an independent rationale. |
| G4 — Confirmatory study | Preregistered construct, analysis, outcome, and meaningful effect; sample justified for power/precision and clustering; independent confirmation data. | Frozen criteria are not met or alternative explanations survive. | Modify/abandon the specific claim; preserve negative and inconclusive evidence. |
| GR — Runtime branch | Independent model-task improvement from detector-guided intervention beyond matched interventions, with false-trigger costs measured. | No incremental outcome benefit or unacceptable tradeoffs. | Revise or abandon runtime use independently of G3. |
| GW — Reward branch | Independent capability improvement, robustness to noise/shift, and explicit tests for feature gaming and proxy optimization. | Higher reward without target improvement, exploitability, or unstable out-of-distribution behavior. | Reject/narrow reward use; accuracy alone never licenses production RL. |
G0-G3 are directional feasibility/continuation gates, not confirmation. Passing a gate changes the next research decision, not the evidential label of untested downstream claims. GR and GW are separate branches; human signal evidence also does not establish human-to-model transfer. Modality translation requires a later separate hypothesis and design.
For each gate, record the hypothesis, operational definition, alternative explanation, control, success criterion, failure criterion, and next decision before execution. Link any actual evidence to the experiment registry and hypothesis map.
Complementary Fiction-CWR gates (NEW PROPOSAL)
These do not replace or imply passage of G0–G4, GR or GW.
| Gate | Evidence · decision |
|---|---|
| CWR-0 — Reference validity | Finite semantics; independent witness/solver; alternate solutions, invalid mutants, exhaustive and metamorphic checks. Token engineering evidence is in VERIFICATION; no across-family completion inferred. |
| CWR-1 — Composition manipulation | One public token manifest and bounded necessary type-DAG certificate audited; target/exposure matching checked. Source mismatches, protected main stimuli and empirical freeze remain unresolved; no model interpretation licensed. |
| CWR-2 — Directional model evidence | Frozen settings/budget/P qualification; unconditional and qualified paired results; all family/task outcomes. Proposed, zero model runs. |
| CWR-3 — Confirmation and scope | Independent private bases, prospective power/precision, preregistered substantive criterion and clustered uncertainty. Unresolved; representation/budget diagnostics before persistent claims. |
Upper interval bound below a frozen deficit rejects that scoped size; crossing is inconclusive. Candidate δ=0.20/τ=0.90 are proposed, not passed or universal thresholds. Primitive failure blocks the understood-primitive attribution. EXP-0002 records the seven design statements.
Separate EXP-0003 workflow diagnostic gates
NEW PROPOSAL: Protocol (not included in this public selection: evaluations/fiction-cwr-workflow-v0.1/PROTOCOL.md). No empirical gates passed.
| Gate | Required next decision |
|---|---|
| WF-0 — Candidate measurement | Review all-solution motif/copy/exposure certificates and residual source differences; retain frozen checker/oracle; candidate engineering verification alone is insufficient for unrestricted validity claims |
| WF-1 — Execution freeze | Fresh protected materials, counterbalanced assignment/order, pinned model/settings, measurable resource/usage convention, matched stage/cumulative ceilings, P and numerical continuation rules, transport policy and authorized spend |
| WF-2 — Directional attribution | All task outcomes/P/resources, C gain and shrink with bounded F loss; no F-only deterioration or hidden resource advantage; finite related panel only |
| WF-3 — Independent confirmation | Justified independent private tasks and prospective power/precision; no repetition-as-task counting or human-mechanism inference |
Failing measurement/resource/prerequisite gates blocks the intended attribution. No baseline gap means the deficit-explanation premise is untested. Equal validity at lower cost supports efficiency only. Runtime/reward/transfer require separate tests; success authorizes none automatically.
Token scope update — 2026-10-12
CWR-1 engineering audit is partial: it verifies the original showcase and one imported public pair without rejecting valid alternatives. Strict illegal-prefix agreement precedes enumeration pruning. F/C source worlds differ in useful stock/trajectories and feasible-plan counts. Address or narrow these residuals and freeze independent evidence/provider settings before model testing. No human, runtime/reward or empirical gate is implied.