GCTS Experiments

The project must test the theory and expose failure modes. Every GCTS experiment should have pre-registered hypotheses, baselines, ablations, held-out data, no runtime oracle, and failure criteria.

Baseline Families

Core Experiments

EXP-001: Synthetic Latent-Context Resolution

Test whether GCTS recovers hidden modifiers that explain contradictions: time, subgroup, measurement method, population, jurisdiction, or mechanism.

Success criteria:

• top-3 world coverage at or above 85%;
• latent predicate recovery F1 at or above 0.70;
• calibration ECE at or below 0.10.

EXP-002: Fact-Verification Grounding Without Runtime Labels

Test claim-status assignment with benchmark labels withheld at runtime. Candidate data sources include FEVER, SciFact, FEVEROUS, and AVeriTeC.

Baselines:

• RAG plus direct answer;
• RAG plus NLI verifier;
• claim-verification classifier;
• multi-agent debate;
• GCTS without multiverse;
• full GCTS.

Success criteria include zero strict promoted claims with unresolved citations, improved calibration over baselines, and higher abstention precision.

EXP-003: Multiverse Calibration

Test whether top-K possible worlds are calibrated and informative.

Metrics:

• top-K world coverage;
• Brier score;
• expected calibration error;
• entropy versus error correlation.

EXP-004: Oracle-Boundary Ablation

Compare three conditions:

1. No labels.
2. Labels for offline calibration only.
3. Illegal runtime oracle upper bound.

Condition 2 should improve calibration over condition 1. Condition 3 is an upper bound and must never be treated as deployable.

EXP-005: Chirality Predictiveness

Test whether chirality predicts synthesis difficulty beyond embedding distance, claim-graph conflict, and graph cycle count.

Dependent variables:

• convergence iterations;
• contradiction residual after synthesis;
• human uncertainty rating;
• false synthesis rate;
• abstention correctness.

EXP-006: Adversarial Record Suppression

Test whether GCTS distinguishes absent evidence, evidence of absence, inaccessible evidence, sealed evidence, likely withheld evidence, destroyed evidence, and not-generated evidence.

Synthetic planted states include:

• expected record exists and is produced;
• expected record exists but is inaccessible;
• expected record exists but is sealed;
• expected record exists but is withheld;
• expected record was destroyed;
• record was never expected to exist;
• produced record affirmatively refutes the claim.

Success criteria:

• access-state F1 at or above 0.75 on planted cases;
• improved likely-truth Brier/ECE over no-access ablation;
• lower false rejection rate where decisive records are inaccessible;
• lower false promotion rate where evidence of absence is available;
• lower missingness-overreach rate on not-generated records.

Test Layers

Unit tests cover schema validation, citation resolution, tensor rule firing, proof trace emission, posterior normalization, entropy/confidence formulas, and access-state invariants.

Property tests enforce that posterior mass sums to 1, strict promotion requires proof traces, invalid citations force strict status to unsupported, soft rules cannot promote strict truth, and absence of evidence cannot become evidence of absence without record-duty and access-path basis.

Integration tests cover ingestion, extraction, grounding, access modeling, closure, world ranking, and report generation.

Red-team tests include citation hallucination, semantically similar negations, unsupported paraphrases, misleading lexical overlap, narrow evidence inflated into broad claims, false suppression inference, selective disclosure, and strategic partial production.