Forge Intelligence

PhysicsGenesis batch runbook

Record immutable IDs for source, library, prompt pack, model, runtime, world, evaluator, safety policy, clean-room policy, and budget.

Updated

1. Freeze context

Record immutable IDs for source, library, prompt pack, model, runtime, world, evaluator, safety policy, clean-room policy, and budget.

2. Plan

Run 01_partition_and_ingestion_plan; validate against physics-ingestion-plan.schema.json. Stop on missing partition, source metadata, unit route, empirical evidence route, or safety route.

3. Extract bounded source slices

Run 02_source_claim_extractor. Extraction identifies candidates; it does not validate or promote them.

4. Build typed artifacts

Candidate Prompt
quantities, units, dimensions 03
constants and reference values 04
principles, laws, theories, models 05
mathematical formulations and derivations 06
measurements and uncertainty 07
experiments and apparatus 08
datasets and evidence 09
mechanics through quantum domain objects 1017
simulations 18
inverse problems 19
candidate law discovery 20
symmetry, conservation, scaling 21
curriculum and historical pressure 22
cross-domain analogies 23
invention candidates 24

5. Validate shape and registries

Validate JSON/YAML/LMeta and all native objects. Check source, dependency, quantity, unit, partition, and visibility closure.

6. Falsify independently

Run 25 in a separate context. Attach exact bounds, tools, seeds, splits, and counterexamples. A bounded failure to falsify is not proof.

7. Validate independently

Run 26 with authenticated receipts appropriate to the claimed status: unit/dimension, symbolic/formal, numerical, calibration, held-out, intervention, or replication.

8. Deduplicate, safety-check, and contamination-audit

Run 27, 28, and, for clean-room runs, 29. Keep safety and epistemic conclusions separate.

9. Govern promotion

Run 30. A failed hard gate blocks promotion regardless of utility.

10. Plan repository changes

Run 33; review paths and registry mutations. Apply through ordinary version control only.

11. Release audit

Run 34, then repository CI, maturity audits, index generators, wiki projection, and session freeze gate.

12. Historian

After learner freeze, run 35 in isolation. Do not feed modern aliases or comparison results back to the learner snapshot.

max_source_slices: 24
max_objects: 100
max_unresolved_dependencies: 12
max_failed_validations: 3
max_break_in_calls: 6
max_total_tokens: 80000
require_safety_review: true
require_human_release_review: true