Write and audit the search-index card

The final card is a compact explanation of the system you built. It should let another reader understand the boundary without promising more than the artifacts show.

Goal

Write and audit a search-index card that states the exact term rule, postings, query behavior, score, tie order, excerpts, saved evidence, replay result, and limits of the project.

Include

Describe:

• the corpus and tokenizer identities;
• the fact that only text tokens are indexed;
• exact case, spelling, underscore, non-ASCII, and position behavior;
• document-order postings and the count invariants;
• compatible query normalization and scanner rules;
• ignored tokens, missing terms, repeated terms, and empty outcomes;
• query_term_count * document_term_count contributions, including zero rows;
• positive-score filtering and the three ranking tie keys;
• top_k, candidate counts, earliest-match excerpts, and edge metadata;
• serialized artifacts, reload checks, and separate replay evidence; and
• which later retrieval work may replace this score or term representation.

Do not claim

Do not call exact term matching semantic search. Do not claim that repeated terms measure meaning, that a rank is universally relevant, or that this small fixture is fair, popular, representative, fast, production-ready, or suitable for every corpus. Do not imply that replay proves model quality or user satisfaction.

Deliverables

Write output/report.md and audit every sentence against a saved artifact. Include the corpus and tokenizer versions, configuration identity, counts, one traced query result, one excerpt, replay agreement or the first mismatch, and a short limitations section. Link observations to artifact names rather than hiding them in prose.

Checks

Use the prose rubric to reject unsupported claims about semantics, fairness, popularity, universal relevance, speed, and production readiness. Check that the report names zero-term documents, ignored and missing query terms, zero contributions, no zero-score fillers, deterministic ties, and unknown or truncated evidence where applicable.

Read the card after changing one artifact and confirm the audit catches the stale statement. Confirm the report is readable without the interactive editor and that its examples retain document and query identities.

Workspace

Keep the card and its audit in src/report.py or a clearly separated report writer. Do not turn a plot observation into a linguistic or relevance claim.

Hints

Review

Trace one sentence in the card to an index, contribution, result, excerpt, or manifest artifact. Then point to one thing the project cannot establish. If a reader skipped TEXT-02, can the card still identify the fallback boundary?

How to check your work

Checks compare the card with the accepted search-index contract and saved manifest. The card is complete when every public claim has bounded evidence and every important limit is stated plainly.