Part of CNS 8.0 / Grounded Dialectical Orthesis

11 — Implementation Plan

11 — Implementation Plan

MVP objective

Build a CNS 8.0 prototype that can process a small evidence corpus, produce SNOs, identify productive contradictions, perform proof-constrained synthesis, recover simple latent predicates on synthetic tasks, and emit an orthesis candidate with evidence, proof traces, residuals, and uncertainty recorded.

Phase 0 — Repository skeleton

Deliverables:

  • cns8/ Python package;
  • tests/ with deterministic toy cases;
  • configs/cns8_mvp.yaml;
  • JSON schemas;
  • run manifest format.

Phase 1 — Evidence and SNO extraction

Components:

  • EvidenceStore
  • EvidenceAtom
  • SNO parser
  • Claim/Relation extractor interface
  • citation validator

Acceptance criteria:

  • every evidence atom has stable ID and hash;
  • missing evidence IDs reject invalid inputs;
  • parser handles valid and invalid SNOs;
  • citation validity measured per claim.

Phase 2 — Grounding critics

Components:

  • entailment scorer;
  • source quality scorer;
  • access-state validator;
  • proof status assigner.

Acceptance criteria:

  • strict claims require valid citation and entailment;
  • invalid citation causes rejection;
  • access states do not masquerade as truth values.

Phase 3 — Chirality and entanglement

Components:

  • evidence overlap;
  • graph chirality;
  • evidence-polarity chirality;
  • round-trip language–logic residual;
  • Productive Conflict Score.

Acceptance criteria:

  • pair selector ranks synthetic productive conflicts above unrelated conflicts;
  • high overlap/agreement is not misclassified as synthesis target;
  • high contradiction/no overlap is downgraded.

Phase 4 — Tensor proof closure

Components:

  • rule registry;
  • zero-temperature closure;
  • proof trace recorder;
  • ZTHR metric.

Acceptance criteria:

  • strict claims carry proof traces;
  • unsupported claims cannot be promoted;
  • closure produces expected atoms on toy rules.

Phase 5 — Residual tensor and predicate invention

Components:

  • residual tensor builder;
  • factorization sketch;
  • latent predicate candidate generator;
  • predicate grounding gate;
  • PIU metric.

Acceptance criteria:

  • synthetic hidden contexts recovered above baseline;
  • spurious predicates rejected by grounding/complexity gates;
  • residual energy decreases when correct latent predicate is accepted.

Phase 6 — Synthesizer and orthesis loop

Components:

  • structured synthesis planner;
  • LLM renderer with bounded prompt;
  • re-grounding loop;
  • round-trip residual scorer;
  • orthesis report.

Acceptance criteria:

  • synthesized SNO preserves evidence provenance;
  • round-trip residual decreases over iterations;
  • final strict claims have proof traces;
  • unresolved contradictions are reported, not hidden.

Phase 7 — Multiverse and audit layer

Components:

  • possible-world generator;
  • posterior scoring;
  • calibration report;
  • audit renderer.

Acceptance criteria:

  • worlds include access assumptions;
  • posterior report does not replace synthesized SNO;
  • final output separates strict, likely, hypothesis, unresolved, and rejected claims.

Engineering stack

Recommended:

  • Python for MVP proof algorithms;
  • Pydantic or dataclasses for schemas;
  • NetworkX for topology;
  • NumPy/PyTorch for tensor operations;
  • sentence-transformers or equivalent for embeddings;
  • NLI model for entailment;
  • optional LoRA for extraction;
  • simple CLI before dashboard.

CLI sketch

cns8 ingest corpus.jsonl --out runs/evidence.jsonl
cns8 propose runs/evidence.jsonl --out runs/snos.jsonl
cns8 critique runs/snos.jsonl --out runs/critic.jsonl
cns8 select-pairs runs/snos.jsonl --out runs/pairs.jsonl
cns8 synthesize runs/pairs.jsonl --out runs/synthesized_snos.jsonl
cns8 orthesis runs/synthesized_snos.jsonl --out runs/orthesis_report.json
cns8 report runs/orthesis_report.json --format markdown

Build order warning

Do not build a dashboard first. Do not build a large multi-agent runtime first. Build the proof-bearing SNO loop first.