Structuring rules

• All ids are kebab-case slugs of ASCII letters, digits and underscores.
• Each section has a prefix: ent_, con_, pri_, heu_, dec_, proc_, pat_, ant_, cau_, ctx_, ift_, exc_, mm_, pb_, qa_, chk_, atm_.
• Ids must be unique within their section and stable across re-compilations of the same source.
• The package_id is global; recommended pattern is <source-slug>-<version>.

Confidence is a float in [0, 1] with two decimal places. The protocol defines five bands; see also the Protocol page.

• 0.90 – 1.00 — the source states the claim directly and unambiguously.
• 0.75 – 0.89 — the source supports the claim with mild interpretation.
• 0.50 – 0.74 — inferred by combining two or more explicit statements.
• 0.25 – 0.49 — weak inference; agents should treat as a hypothesis.
• 0.00 – 0.24 — extractor uncertain; emit only if useful, mark source_basis: uncertain.

• Strings are NFC-normalized; whitespace collapsed to single spaces.
• Languages follow BCP-47 (en, pt-BR, …).
• Domains use the closed enum from the schema; novel topics go in subdomains.
• Singular labels for entities and concepts; verb-led labels for procedures and playbooks.

When two extractions produce overlapping items (e.g. one item per chunk in the reduce phase):

• Merge by canonical id. If ids differ but labels are aliases, prefer the higher-confidence record and add the other label to aliases.
• When confidences disagree, keep the higher value and lower it by the standard deviation between sources, capped at the higher of the two.
• When source-basis disagrees, downgrade conservatively: explicit + inferred ⇒ inferred.
• Rich sections (procedures, playbooks, if_then_rules) are deduplicated on a normalized fingerprint of their action sequence, not just on label.

After reduce, the compiler walks atomic_units and retrieval_chunks and promotes statements into higher-order sections:

• Conditionals ("if X then Y", "when X, do Y") → if_then_rules.
• Multi-step actions (numbered or sequenced) → playbooks.
• Stated failure modes ("avoid", "do not", "common mistake") → anti_patterns.

Promotion is rejected when the candidate fails the language filter, completeness check or the truncation filter. Rejections are counted in result.promotion.rejected.

Every non-synthesized item must have at least one matching source_traceability entry whose extraction_type equals the item's source_basis. Synthesized items (e.g. retrieval chunks) may omit it. The v1.3.1 pipeline rebuilds this section automatically after sanitation, so stale entries pointing at removed items never reach the final package — and now propagates source_record_id when the source is record-oriented (JSONL, json_array, FAQ, normative texts), so consumers can pinpoint the exact record an item came from.