The oracle boundary is the line between offline calibration and runtime truth ranking.
GCTS may use labels or expert judgment for training, calibration, evaluation, and error review. Runtime truth ranking and posterior mass must come from evidence, access states, rules, and calibrated parameters.
Allowed Oracle Use
- Training labels.
- Calibration labels.
- Evaluation labels.
- Expert review of error cases.
- Human approval of new strict rules.
Forbidden Oracle Use
- Runtime access to gold labels.
- Runtime human or model truth decisions that bypass evidence closure and world ranking.
- Dataset label leakage into retrieval, ranking, or world building.
- Prompting an LLM to decide truth and using that answer as posterior mass.
Promotion Policy
Strict claims require:
- resolvable evidence references;
- zero-temperature proof support;
- proof traces;
- no runtime label access.
Likely-truth claims require:
- posterior calculation over explicit worlds;
- confidence and uncertainty decomposition;
- clear distinction between posterior probability, strict support, and confidence.
Record-contingent claims require:
- identified record dependencies;
- access-state classification;
- an explanation of what evidence would change the ranking.
Main Risks
False certainty: posterior scores can be misread as objective truth. Mitigation: confidence bands, entropy, uncertainty decomposition, estimative language, and explicit caveats.
Source poisoning: manipulated evidence can shift world rankings. Mitigation: source reliability priors, source diversity metrics, adversarial evidence tests, and source-quality uncertainty.
Access overreach: the system may infer withholding or concealment from ordinary missingness. Mitigation: record-duty thresholds, access-path checks, competing missingness worlds, MDL penalties, and conservative confidence.
Access underreach: the system may treat inaccessible controlled records as simple lack of evidence. Mitigation: record-contingency status, expected-record modeling, access uncertainty, and next-evidence requirements.
LLM rendering drift: the renderer may add unsupported details. Mitigation: render from structured payload only, post-render verification, and rejection of unsupported phrases.
Deployment Checklist
- All strict promoted claims have resolvable citations.
- All strict promoted claims have proof traces.
- Runtime labels were unavailable.
- Posterior, strict support, and confidence are reported separately.
- Top alternative worlds are shown.
- Record-contingent claims identify record dependencies.
- Uncertainty decomposition is shown.
- Evidence that would change the conclusion is listed.
GCTS is a decision-support system. It should expose alternatives, likely-truth rankings, access constraints, and uncertainty. It should not replace human judgment in high-stakes domains.