GCTS MVP Build

The GCTS MVP can be built without full custom model training. The first target is an auditable likely-truth engine that accepts a bounded corpus, extracts evidence and claims, models access states, enumerates worlds, and emits ranked reports.

Phase 1: Local Prototype

Use existing models and explicit schemas:

• LLMs for candidate extraction, latent-context suggestions, access-hypothesis suggestions, and rendering.
• Retrieval plus citation validation for evidence grounding.
• NLI or entailment models for claim-evidence scoring.
• A small evidence-access model for expected record existence, availability, control, and non-production.
• A rule compiler for a monotone tensor-logic core.
• Candidate-world enumeration or beam search.
• Calibration data to map evidence, access, incentive, and contradiction signals to probabilities.
• A dashboard to expose world rankings, proof traces, record-access states, uncertainty, and next evidence.

Fine-tuning is optional in Phase 1. If used, it should target extraction, evidence linking, access-state classification, and calibration rather than direct runtime truth judgment.

Runtime Data Products

The MVP should persist:

• evidence atoms;
• record-access states;
• institutional incentive profiles;
• claims and relations;
• rules and proof traces;
• world views;
• posterior and confidence reports;
• rendered synthesis reports.

API Surface

The first API should expose:

• POST /runs to create a synthesis run from a corpus manifest;
• GET /runs/{id} for run status;
• GET /runs/{id}/evidence for evidence atoms;
• GET /runs/{id}/access for record-access states;
• GET /runs/{id}/worlds for top-K possible worlds;
• GET /runs/{id}/claims for claim rankings and statuses;
• GET /runs/{id}/report for the rendered report.

MVP Gates

The first build succeeds only if it reaches:

• 100% resolvable citations for promoted strict claims;
• zero promoted zero-temperature claims without proof traces;
• calibrated claim probabilities with ECE at or below 0.10 on held-out verification tasks;
• top-3 world coverage at or above 85% on synthetic latent-context tasks;
• measurable chirality correlation with synthesis difficulty;
• measurable access-state calibration on adversarial record-suppression tasks;
• explicit distinction between unsupported, record_contingent, conflicted, and rejected;
• ablation evidence that multiverse/proof/access scoring beats simple RAG and LLM debate baselines on grounding, uncertainty quality, and likely-truth ranking.

Implementation Boundary

The MVP is a bounded, auditable decision support prototype for testing GCTS under controlled evidence and access conditions. It should say when evidence is insufficient, when a claim is record-contingent, and what record production would change the result.