Rekal

Ledger workflow gate

For a question routed to the ledger, classify the answer type before searching. Choose the first matching row and read exactly that workflow. Do not blend several workflows: concentrated guidance is more reliable than a pile of partially relevant checks.

1. Elapsed time or duration between endpoints → Read references/workflows/duration.md
2. A count, set, plural list, repeated events, or ordered history → Read references/workflows/complete-set.md
3. A calendar time/date or temporal relation → Read references/workflows/event-time.md
4. A qualified prediction, likelihood, possibility, or inference → Read references/workflows/inference.md
5. A fact, episode, explanation, provenance, reflection, or other ledger answer → Read references/workflows/point-fact.md

Classify by the form of the answer requested, not by incidental words: "Which events happened before June?" asks for a set, while "When did the event happen?" asks for event time. The workflow supplies evidence invariants and useful operations, never truth. The ledger remains authoritative; preserve genuine ambiguity and reject unsupported premises.

Final answer check

Before answering, silently compare the candidate answer with the requested actor, entity, relation, time scope, and answer type.

• Reject another speaker's fact, a nearby semantic slot, an adjacent event, or a suggestion or plan mistaken for a completed event.
• When event time is requested, resolve a source-relative expression against the historical assertion timestamp. A relative expression in the question is anchored to the asker's present. Preserve source precision.
• For a count or set, ensure members were enumerated across the requested scope, class-mapped when necessary, and deduplicated.
• Before answering "unknown," make one focused reformulation only when retrieved evidence signals that the exact fact may be buried.
• If a check fails, repair evidence gathering rather than weakening the evidence standard or satisfying a false premise.

Judgment — agent, not the command

• Only what the ledger holds. Do not invent or pad. If the record is thin, say so — or stay silent.
• A partial set is a wrong answer when the question asks for the set. Use rekal find / SQL and page until empty. Ranked recall is for pointed questions, not "all / which / how many / every beat of an arc."
• Keep the record's precision. Month-only evidence supports a month, not an invented day; attribution stays as the record states it; don't fake precision the record lacks or tidy away genuine ambiguity. (A resolvable source-relative phrase is still converted — see the event-time workflow.)
• A false premise has no answer. When the question asserts something the record contradicts or never says, say that — never fabricate the asserted fact, and never silently answer a corrected question nobody asked.
• Drill the hit before concluding absence. A recalled seed you haven't drilled outranks any amount of tree-grepping; grep never answers a ledger question, and an empty stub greps forever.

Semantic warming

A SEMANTIC warming line means the deep-semantic daemon is still loading; those results are keyword + LSA only. If the answer matters, re-run the same recall with exponential backoff (2s, 4s, 8s) until it's gone; after ~three tries proceed — the keyword layer stands on its own.