Milestone 8 of 8
Write and audit the engine card
Explain searchable terms, both position systems, phrase adjacency, proximity limits, result order, excerpts, missing cases, artifacts, replay evidence, and lexical limits.
The final artifact is an explanation constrained by the evidence, not a claim that lexical matching understands language. Write the engine card after the traces and replay can support it.
Goal
Write and audit report.md as a concise phrase-and-proximity engine card that
explains the representation, matching rules, evidence, replay, and limits.
Inputs
Use the manifest, positional index, query analyses, phrase and proximity traces, ordered results, excerpts, summaries, figures and plot data, read-back checks, and replay agreement or mismatch record. The card must describe the actual versioned configuration and source identities used by the run.
Deliverables
The card must explain, in plain language:
- which scanner records become searchable terms and which remain source-only;
- the difference between contiguous
indexed_positionand complete-streamfull_token_position; - phrase adjacency and offset intersection, including overlaps and repeats;
- proximity direction, the intervening-term formula, boundaries, and same-term pair handling;
- deterministic result order and one-excerpt-per-match construction;
- invalid, missing, empty, and zero-term cases;
- artifact names, read-back checks, plot-data evidence, and replay outcome; and
- what the engine cannot claim from exact lexical positions.
Include a small worked trace that links a query, positions, match, and excerpt to the saved artifacts. Distinguish observations from interpretation and name any first mismatch honestly.
Checks
Audit every statement against a saved record or explicit rule. Reject wording that claims semantic understanding, relevance, probability, popularity, fairness, production performance, or ranking. Check that “exact phrase” means adjacent searchable terms, not raw-character equality; that excerpts retain source tokens without injected markup; and that replay claims agreement only when all required comparisons pass.
Check the card names missing and invalid cases, both proximity directions, zero and exact boundaries, repeated and same-term cases, and the distinction between candidate traces and returned matches. Confirm readers can locate each claim's supporting artifact.
Workspace
Keep the card at the project root as report.md. Do not replace the report
with a presentation or a polished search-results mockup. Preserve all machine-
readable artifacts and figures beside it.
Hints
HintStart with one trace
HintName the coordinate
HintLimits are part of the result
Review
Give the card to a reader who has not opened the source code. Can they explain why one phrase survives and one proximity pair fails, find the supporting trace and excerpt, and tell whether replay agreed? Mark any sentence that outruns the artifacts.
How to check your work
Checks compare report.md with the evidence-limited reference card after completing
your own audit. The supplied fixture explains the contract and cites artifacts; it
does not conceal missing evidence behind confident language.