Handbook

Training records and historian comparison (M08)

Governed training-record batches and post-freeze historian comparisons without storing chain-of-thought.

Updated 2026-07-26

Training records

Schema: training-record-batch.schema.json
Runtime: training_records.py
Example: training_record_batch_example.json

Rules (prompt 18):

No chain_of_thought, hidden_reasoning, or scratchpad fields
authenticated records require receipt_ids
Every record carries a checksum and library_snapshot_id on the batch

Historian comparator

Schema: historian-comparison.schema.json
Runtime: historian_compare.py
Example: historian_comparison_example.json

Rules (prompt 21):

Runs after learner freeze only
modern_mappings remain historian-only
Novelty labels bounded (new_to_agent, new_to_snapshot, …) — no humanity novelty from local miss alone

Handbook parity

Paired wiki stubs under wiki/semantic/math-*.md and wiki/execution/math-*.md link here; rebuild forge-intelligence-website after doc changes.