MathGenesis batch runbook

1. Freeze the run context

Create immutable IDs for:

• source snapshot;
• target library snapshot;
• prompt-pack version;
• model and runtime configuration;
• clean-room policy;
• token and tool budget.

No batch should mix outputs from different hidden reference or learner snapshots without explicit lineage.

2. Plan

Combine the system governor with 01_partition_and_ingestion_plan.md. Validate the result against schemas/ingestion-plan.schema.json.

Stop when source metadata, partition, visibility, or verification routes are missing.

3. Extract small source slices

Use 02_source_claim_extractor.md on bounded source spans. Extraction identifies candidate objects; it does not prove or promote them.

A Forge LCDL adapter can reuse extract_schema_from_text for the structured response and deterministic JSON Schema validation around it.

4. Build typed objects

Route candidates by type:

Run 10 only after base objects exist and IDs are stable.

5. Validate machine shape

Validate every object against math-knowledge-object.schema.json and each batch against library-batch.schema.json before semantic review.

Example local validation:

python3 -m jsonschema \
  -i examples/gcd_definition.json \
  schemas/math-knowledge-object.schema.json

Use a registry-aware resolver when validating schemas that reference sibling files.

6. Falsify independently

Run 12_counterexample_falsifier.md in a separate context from the builder. Attach exact finite bounds, solver versions, seeds, and failing cases.

A missing counterexample is not proof.

7. Verify independently

Run 13_validation_tribunal.md with authenticated receipts from the applicable checker:

• proof kernel for formal theorems;
• CAS or substitution checker for symbolic identities;
• numerical residual/convergence harness;
• geometry constraint checker;
• dimensional/unit checker;
• held-out or intervention evidence for physics.

The tribunal issues a bounded status and cannot promote.

8. Deduplicate and govern

Run 14_ontology_deduplicator.md. For clean-room discoveries, compare only with the learner snapshot at this stage.

Run 19_contamination_auditor.md before clean-room promotion.

Run 15_promotion_gate.md with validation, falsifier, ontology, utility, and contamination reports. Promotion is a separate artifact, not a field the generating model may set on its own.

9. Assemble a repository plan

Run 16_repository_assembler.md. Review every proposed path and registry mutation. Apply writes through ordinary version-controlled tooling, never through prompt text alone.

10. Audit and merge

Run 20_batch_release_auditor.md, then execute repository CI:

pytest -q
python3 tools/audit_capability_maturity.py --fail-on-missing --fail-on-orphan-files
python3 tools/generate_intelligence_maturity.py
python3 tools/generate_node_catalog.py
python3 tools/project_concept_wiki.py

Add MathGenesis-specific schema, dependency, partition, and contamination checks to CI before the first native batch.

11. Generate the session dual-wiki

Store route trace, token ledger, ambiguity, artifact IDs, validation receipts, and promotion decisions in machine JSON. Generate the human report from machine truth and run the existing freeze gate.

12. Historian comparison

After a clean-room learner generation is frozen, run 21_historian_equivalence_and_novelty_assessor.md in an isolated context. Do not feed its aliases or mappings back into the learner snapshot.

Recommended first batch limits

max_source_slices: 20
max_objects: 80
max_unresolved_dependencies: 10
max_failed_validations: 3
max_break_in_calls: 4
max_total_tokens: 60000
require_human_release_review: true