AMA-TRUST

Ranked by what a user would notice first. The spreadsheet cluster sorts first because Regina asks a question about her own spreadsheets and gets nothing back — the track's own impact statement. Re-read 2026-08-17: that ordering held and the cluster is most of the way down. The render defect (#3193) is fixed and closed, and what a user would notice first is now the client half of that same fix — raw | pipes on the entity screen, reversed columns in Arabic (#3381). The aggregation tool (#3321) is the one rung with nothing left under it, which is why it takes the Next call. Answer-correctness degradations come second (the things that make any answer wrong today), then structural/choke-point work, then decisions and directions last.

What it ignores: priority labels (the live render defect is p2 and sits first), filing order and age, cost or ease, and the needs:ruling flag as a reordering lever — a ruling that blocks work sorts with its work, not above it.

The 2026-08-16/17 batch — five of the six spreadsheet items moved. In order of merge: the entity-projection render fix (PR #3363, server half of #3193, whole-grid not bounded) · the ground-truth corpus and scoring harness (PR #3364, #3318) · heuristic detection plus a structural fingerprint (PR #3365, #3319) · per-cell types served and declared per-column types (PR #3369, Deliverable A of #3320) · the type-free detection fallback and six scored fixes (PR #3370, taking header recall 0.292 → 0.875) · the tier/tenant-scoped fetch-by-id resolver (PR #3380, the disclosure prerequisite for #3321). Reconciled 2026-08-17: #3193, #3318 and #3319 were closed against their own acceptance criteria; #3320 and #3321 stay open because each PR shipped a named half and said which (Deliverable A without B; the resolver prerequisite without the builder). The two things the batch created rather than closed — the client-side raw-markdown render and the list_by_type payload cliff — sat recorded in a panel writeup and unfiled until 2026-08-17, and are now #3381 and #3382. 🔑 Worth naming as a pattern: a panel finding recorded in a merged PR body is invisible to every queue in this repo. Filing is what makes it work.

Merged in the cycle before — all verified as ancestors of main: the plan (PR #3145, 2026-08-07) · the choke-point pass (PR #3171, 2026-08-08, six built / four registered open) · the search-surface split (PR #3287, 2026-08-14, closing #3255) · the tsvector byte budget (PR #3303, 2026-08-14, closing #3288). Closed since the last pass: #3127, #3255 (state:not-in-effect), #3288, #3184, #3181 (ruled), and #3317 (the float binary-repr defect, completed 2026-08-15 — this page last listed it "Live"). Three earlier fixes also landed this cycle: reading the attachment for data questions, sizing the attachment budget, and the declined-salvage citability fix — whose metered-refund half was built, reviewed and deleted, because Regina's enterprise deployment never reaches the meter.

The census that founded the track (2026-08-07). 74 spreadsheets truncated, 51 of 74 lose real content, and 50,244,172 characters dropped — a lower bound, not exact (it subtracts two counts that are not the same unit). Three clippers: the per-sheet row cap (1,000 rows, 36 docs), the workbook character budget (250,000, 21 docs), and the budget dying inside the last sheet (20 docs, silent — the worst class, cut mid-row with no marker and reporting full coverage). The markers cannot be trusted as a census in either direction: the row cap genuinely binds 60 docs but only 36 say so, and the character budget binds 52 but only 21 do; the silent twenty sit between 250,448 and 250,822 characters, the budget expiring on the last sheet so the loop never reaches the marker-writing check the code comment at xlsx_extractor.py:324-330 predicts. Uncapped these render 127.4M characters against 12.4M stored, of which 64.9M (50.9%) is padding from sheets padded to Excel's 1,048,576-row ceiling. The denominator was also wrong — 2,198 spreadsheets, not 6,039 document kinds.

Three fixes, all measured, all dead. Raise the budget: 132,660 substantive rows carrying 33.1M characters sit beyond row 1,000, so no budget reaches them — it plateaus at 45 of 74. Filter the filler rows: 0.7% of the budget actually spent, zero documents completed (an earlier "3 docs" was refuted). Paginate past the cliff: reading is O(rows) and the answer is O(1) — you cannot read your way to an aggregate. One document alone holds 2,337,469 substantive characters against a 250,000 budget.

The design. Lossless capture: one row per worksheet in a side table, no embedding, no tsvector, not new nodes — both caps, the truncated suffix and every marker deleted. The search profile (sheet names, headers, labels, row counts) carries further than intended: searchable_text has four downstream consumers — the full-text generated column, the node embedding, the sensitivity tier classifier, and the fallback arm of two content chains. The split (PR #3287): searchable_text = profile + summary + roster, the embedded surface = profile + summary. The "length dilutes the embedding" theory is refuted — the variable is TOPICALITY. The profile is written at extraction — sheet names, column headers, label values, row counts — and ranking should improve, because 250k of numeric cells today dilute term weighting and the code already calls the current value "an unparseable digit run". No migration and no materializer change — both already land in node properties. A summary stage ships with it, because a spreadsheet's mechanical title was "Spreadsheet: Sheet2" — which is why images, whose vision model writes them one, outranked spreadsheets on every topical question.

Re-extraction is lazy, not a sweep. The bump ships with the code (PR #3287, xlsx-extractor v2 → v3); a workbook re-extracts on the next enrichment pass of its own source, so a dormant tenant's sheets stay at v2. Re-measured 2026-08-15: 1,940 of 2,672 nodes at v3, 572 at v2, 60 at v1, one hundred unversioned; 2,507 documents hold lossless grids (6,373 sheets, 28 MB). The bump also turned a latent defect live: 2,612 of 2,672 nodes now carry an empty body_text — the inline-render row above. Median indexed characters recovered from the v2 trough of 313 to 1,707 (v1: 2,278). The embedding half needs no re-extraction: v2 nodes already carry a profile, so the read-time composer took the profile branch the moment this deployed.

Table detection. Header found on 94.7% of sheets, but only 39.6% on row 0; label column first 73.7%; both 53.5% — detect, don't assume. Prior art is not shippable: TableSense publishes annotations only (no model, no weights), ExcelTableCNN self-reports 40.3% recall. The "~71%" coverage figure is plausibility, not accuracy. Typing is nearly free once boundaries are known — 0.998 median column-type purity, 76% median of columns numeric — so labels really are the small minority worth indexing.

Storage reality. Node table 2,317 MB: embedding 469 MB, HNSW indexes 1,085 MB, trigram 57 MB, full-text 39 MB, all searchable text 26 MB. A vector is 6 KB whether the node holds 500 characters or 250,000, so node count is the expensive axis — why sheets get a side table, not nodes. Regina's rows sit in two indexes, ~21.9 KB per node marginal.

The tsvector budget (PR #3303). Ceiling 1,048,575 bytes; to_tsvector raises rather than truncating. A character budget never overflowed — it peaked at 89.0% (CJK 2-char tokens) — so the case for UTF-8 bytes is the margin (89.0% → 45.8%), not a demonstrated overflow. Cost re-measured in bytes: 26 rows exceed 200,000 bytes (2 pdf, 24 xlsx), one more than by characters. The largest production PDF carries 959,521 characters and builds only 81,068 bytes (0.0845 per character) because prose repeats, while the worst real ratio is 2.1193. Truncation is not free even for surviving terms: ts_rank_cd cover density and an absolute-score filter both move.

Choke-points and the second census. Eleven questions named; six have a choke-point. The gates: two rulings — the document-body expressions (#3185, open) and the turn ceiling (#3181, ruled unbounded); the citation regex made unbounded (#3184); the prefix→media_type table deleted rather than guarded; the budget rule made name-blind and given its first test file. The pass itself refuted six claims (citability was decided from per-turn tool identity but gated on thread-scoped text; the per-turn character figure was 120,000, not 84,000). A second census found the fallback-order expression across five sites — document_scope_classifier, stage_inputs._cached_node_text, node_document_repository, entity_projection_service and sensitivity_classification_phase — each choosing among a node's content fields; two return nothing for a spreadsheet, and one of those gates conversation analysis — latent until the bump ran, live the moment it did. The one-looking-defect that is not: two chains are exact reverses, and the property whitelists are disjoint so the order cannot change the answer.

What this page does not claim. Post-merge production telemetry has not been measured. Corpus-wide re-extraction cost is estimated from 74 documents, not the 2,306 on production. The structural census covers Regina only. The template-signature count keys on sheet names (collisions unchecked). The search index is configured for English on Arabic content — pre-existing, out of scope. Corrected 2026-08-17: this section used to say the one-attachment-per-message limit had never been sized. It has — 57% of Regina's document-carrying maildir messages lost at least one file, ≥71,440 documents recoverable — and the model fix has since shipped forward-only. What is still unclaimed is any measurement of the 2026-08-16/17 batch on production: the detection scores come from a 22-sheet labelled sample, and no post-merge telemetry has been read for the render fix or the typed read path.