Freeze the evidence and define one vote
This workflow is for editors preparing a book or product recommendation table from an authorized HN export. It does not fetch comments, identify every entity automatically or measure public opinion. Record source links, capture date, thread scope, exclusions and reviewer. Keep personal information out of submitted material unless necessary and authorized.
Extract relevant arrays with JSONPath or parse a CSV, then explicitly map entity, context, id and commentId. Extraction is not semantic field mapping. Keep source URLs and reviewer notes separately; CSV input accepts only the named input columns. The context must contain the entity name literally, ignoring case. Split a comment mentioning two entities into separate evidence rows with distinct ids and the same commentId, without discarding qualifying context.
Preserve original text and a versioned alias ledger. The classifier does not merge aliases automatically. Do not silently rewrite a title inside a quotation to force a match; if a canonical display name is needed, maintain its connection to the actual evidence spelling.
Validate structure without erasing disagreement
Use a supplied compatible JSON Schema to check the mapped array and field types, but remember that structural success does not prove a quote belongs to the source. Each run accepts 1–20 records before grouping. Give ids stable meanings rather than depending on row-position defaults across batches.
Exact export duplicates can be inspected on a copy with fuzzy matching off, case-sensitive matching on and whitespace trimming off. Use the full evidence fields when defining equality and retain a removal log. Do not remove repeated entity names or similar-sounding comments: they may represent different sources or disagreement. The classifier separately groups by entity and commentId, falling back to an entity/context fingerprint when the comment id is absent. A shared comment containing conflicting labels stays in review; hiding one side can manufacture an apparent recommendation.
Review meaning, mode and corrections
For a concrete check, take three distinct comments about the same title: one explicitly recommends it, one criticizes it and one only plans to read it. If all are resolved correctly, the grouped before count is three and after is one. Criticism is not a negative vote subtracted from praise; reading intent is a separate category. Mixed, low-confidence and unresolved cases do not become automatic recommendations.
Rules-only performs no model calls, but it is not a promise of browser-local processing or complete language coverage. Its narrow rules can leave ordinary prose unresolved. Hybrid sends unresolved records to an external model provider. Approve that disclosure first, preserve the mode and threshold and do not treat a confidence score as a measured accuracy rate. Model evidence is checked against the supplied context, not the original website; humans must still verify interpretation and provenance.
Download the correction template from the current report. Its id and fingerprint bind each decision to the entity/context pair. Record a reason, set confirmed only after review and preserve mixed status when the evidence really is mixed. Confirmed non-mixed praise can qualify without a numeric model confidence; unresolved conflicts within a group still block inclusion. If text changes, generate a fresh template rather than bypassing stale-correction rejection.
Reconcile batches before publishing the table
Compare initial and reviewed JSON using stable row ids and consistent ordering. A path-level diff does not match records semantically. Retain original reports and correction files so a reviewer can explain every included or removed vote.
The tool's before counts are grouped entity/comment observations within one run, not raw mentions, authors or unique readers. after is eligible praise groups. Never add batch rankings blindly: reconcile repeated global entity/comment keys and conflicting judgments first, using a documented ledger or separately reviewed aggregation. Keep unresolved-group counts distinct from row-level review counts. State alias policy, sample scope and missing-comment-id limitations beside the final table.
Select ranking explicitly before JSON-to-CSV conversion; passing a whole report can select a different nested array. Keep the classifier's native mention CSV for detailed evidence and preserve JSON as the unaltered reference. Generic CSV quoting is not spreadsheet formula protection: import untrusted text columns as text and check formula-like cells before sharing. Verify totals and filenames, retain local copies and hash the actual final files. Use related data-quality or privacy workflows for broader cleanup or redaction; this page is about evidence-bound recommendation counting.