Research-home descriptions: 511 student-ready-blocked entities have no stored prose to synthesise from (scraper gap)

[quntao-z]

Context

Re-scoped 2026-09-22. The split of this cohort by cause, with current counts and the host ranking, is in this comment. The original numbers below are kept as the record of what was believed at the time.

This issue now tracks only the rows with no stored prose to synthesise from: 511 rows, 110 of which carry no hard blocker outside the description family. The card-derivation half of the original cohort was a distinct root cause and is tracked in #2959.

---

While repairing the thin/missing-description student-visibility blocker class (442 non-archived research entities blocked only by description-family hard reasons: missing_card_description + thin_description/missing_description), the grounded LLM-synthesis lane could only promote the subset that carries real source material.

Applied result (Development): 42 promoted to student_ready; 400 still blocked.

Root cause

The remaining entities have no usable source prose to synthesize from, even after the NIH/NSF grant-abstract fallback. They were never captured with a research description at scrape time. Source-availability breakdown over the class:

• ~209 had stored text or a grant abstract (synthesizable); of the candidates actually attempted, 42 passed grounding + quality + genuine-prose gates.
• ~233 were source-empty (no full description, no corroborating profile-synthesis text, no NIH/NSF grant abstract).

Dominant source hosts among the source-empty entities:

• medicine.yale.edu - 71
• orcid.org - 27
• campuspress.yale.edu - 13
• som.yale.edu - 7
• wti.yale.edu - 6
• long tail of departmental hosts

Proposed follow-up

Investigate the medicine.yale.edu faculty/lab scraper first (largest single gap): confirm whether the source pages carry a research-focus blurb that the scraper is not capturing into a fullDescription observation, and if so, capture it. ORCID-only entities likely need a different source (they carry no narrative research description on ORCID itself).

No entity should be promoted without genuine source-backed prose; these stay correctly blocked until re-scraped.

Related: #1876 / #1877 (record-scoped, durable synthesis tooling used for the 42 promotions).

[quntao-z]

Instance with a transferable-sibling lever (found auditing the visibility-gated shard, t2402):

6a8ba34e2fc9821108f8e590 "Yansheng Liu Lab" (LAB, PI yl994) is source-empty (no short/full description, empty websiteUrl) and sits in operator_review - a member of this cohort.

But the SAME PI has a content-rich FACULTY_RESEARCH_AREA sibling, 6a057e5c13fc60d57ec2bc46 "Yansheng Liu Research", carrying a real source-backed description (proteostasis / DIA-MS / cancer, from medicine.yale.edu/cancer/profile/yansheng-liu/). That shell is held as duplicate_risk because the gate suppresses the profile-area shell in favor of the concrete lab - so BOTH are held and the professor has zero student-visible cards.

For source-empty concrete homes that have a same-PI content-rich profile-area shell, the recovery lever is a description TRANSFER/merge (seed the empty concrete home from the shell's already-source-backed prose + sourceUrls, then let the shell defer), not just fresh scraper acquisition. Corpus-wide this exact strict pattern (content-rich FACULTY_RESEARCH_AREA shell + same-PI concrete home stub with no description, both non-ready) is currently a singleton (this pair), so it is a narrow lever, but worth encoding in the empty-home repair since it needs no new scrape. Distinct-but-related to the both-held invisible-faculty failure in #1890.

[quntao-z]

Second-opinion correction (holdfix worker, Development): the dept-mcdb-mark-mooseker FACULTY_RESEARCH_AREA previously logged here as a source-empty instance is NOT source-empty. Its research prose exists on the professor's secondary official Yale profile (medicine.yale.edu/profile/mark-mooseker/), which the description harvest never checked because only the bare mcdb.yale.edu/profile/... stub was on-record. Harvested it, wrote grounded description observations, and the entity is now student_ready. Tracking the real root cause (secondary-profile harvest gap) in #1937 rather than here.

[quntao-z]

Root-cause detail from a HoldFix second-opinion pass (Development), for the humanities slice of this source-empty class.

For James Berger - Research (FACULTY_RESEARCH_AREA, _id 6a058ceeba66f3c14bd8501b, source americanstudies.yale.edu/people/james-berger) the description was empty (profile_fallback_only + missing_card_description). Direct cause: the Yale department profile pages return HTTP 403 to our fetcher. Both americanstudies.yale.edu and english.yale.edu 403 even to a browser User-Agent (Akamai/WAF block), so the scraper never grounds a description.

Key finding: for published faculty this class is NOT truly source-less. An authoritative, citable description was recoverable entirely from OUTSIDE the Yale profile - the book publisher (NYU Press author bio + book description) and OpenLibrary work subjects/descriptions - which also independently confirm the Yale American Studies/English identity.

Suggested follow-up:

• Add a description-backfill fallback for humanities FRAs that mines publisher/library metadata (OpenLibrary works API, publisher author-bio pages) keyed off the faculty member authored works, when the Yale profile is unreachable.
• Separately track the Yale-profile 403/WAF access wall as the upstream blocker for this whole class (it also blocks archive.org in this env).

Sources used for the Dev repair (each field written as an evidence-first observation with its sourceUrl): nyupress.org author bio + description for The Disarticulate; openlibrary.org works OL1964757W (After the End) and OL21072042W (The Disarticulate).

[quntao-z]

Redacted 2026-09-13: this comment paired person-bearing identifiers with claims about those people, and this repository is public. Editing cannot remove published text, because GitHub serves every prior revision to anyone without an account, so the original was deleted and this rewrite posted in its place. 1 line(s) that identified rows by identifier were removed; the analysis is otherwise unchanged. Originally posted . See docs/person-identifier-convention.md (#2624).

---

Additional instance - fixer shard t2434.

1 line(s) omitted here: they identified rows by person-bearing identifier. See the predicate and counts in this comment.

• Only source is a bare CV scaffold (campuspress.yale.edu/stevenfraade) with menu pages (Personal Data / Education / Professional / Publications) and contact info but no research-summary prose; ORCID has no biography/keywords. Zero observations exist for the entity.
• No truthful, source-backed description can be synthesized without falling into the data-quality(descriptions): humanities faculty CV/bio dumps served as descriptions (#1456/#1507 residual) #1533 CV/bio-as-description anti-pattern, so held. (Separately cleaned a namesake physics researchAreas graft on this row; tracked on data-quality(dedupe residual): namesake researchAreas grafts still live on student_ready (prior purges slug/entity-scoped) #1407.)

[quntao-z]

Redacted 2026-09-13: this comment paired person-bearing identifiers with claims about those people, and this repository is public. Editing cannot remove published text, because GitHub serves every prior revision to anyone without an account, so the original was deleted and this rewrite posted in its place. 3 line(s) that identified rows by identifier were removed; the analysis is otherwise unchanged. Originally posted . See docs/person-identifier-convention.md (#2624).

---

Another instance from the shard sweep (thread t2495), notable because the profile page here DID carry harvestable research prose the FRA seed never captured, which directly supports this issue's "confirm the source carries a blurb and capture it" direction.

3 line(s) omitted here: they identified rows by person-bearing identifier. See the predicate and counts in this comment.

• Data repair (Development, done): derived a source-backed fullDescription/shortDescription from that profile, set sourceUrls + fieldProvenance to the profile URL, recomputed to student_ready, Meili synced. Emeritus but present on the live directory as graduate + undergraduate faculty and activeAtYaleCache: true, so a promote (not a suppression) per the emeritus guidance.

Scraper follow-up direction (deferred here, beyond a safe one-shot): the Music (yalemusic.yale.edu) faculty-directory scraper should crawl each faculty profile page and capture the research/bio blurb into a fullDescription observation, same pattern the issue proposes for medicine.yale.edu. The profile markup clearly carries the prose, so this is a capture gap, not a source-empty case.

[quntao-z]

Redacted 2026-09-13: this comment paired person-bearing identifiers with claims about those people, and this repository is public. Editing cannot remove published text, because GitHub serves every prior revision to anyone without an account, so the original was deleted and this rewrite posted in its place. 3 line(s) that identified rows by identifier were removed; the analysis is otherwise unchanged. Originally posted . See docs/person-identifier-convention.md (#2624).

---

Another instance of this source-empty class (smart-fixer shard 5, t2349).

3 line(s) omitted here: they identified rows by person-bearing identifier. See the predicate and counts in this comment.

• The only fullDescription/shortDescription ever captured came from the department-wide aggregator page /undergraduate/undergraduate-research-opportunities (sourceName lab-microsite-undergrad-llm): a generic MCDB department fields-list, and its sibling contact quote on the same scrape named a different person ("Dr. Jacob Stagray"). Correctly superseded; served value is empty.
• Host: mcdb.yale.edu (departmental long-tail in the Research-home descriptions: 511 student-ready-blocked entities have no stored prose to synthesise from (scraper gap) #1878 breakdown).

Decision: HELD in operator_review — correctly blank rather than wrong. No promotion without genuine source-backed prose, which this source does not provide. Filing here rather than opening a duplicate.

[quntao-z]

Redacted 2026-09-13: this comment paired person-bearing identifiers with claims about those people, and this repository is public. Editing cannot remove published text, because GitHub serves every prior revision to anyone without an account, so the original was deleted and this rewrite posted in its place. 3 line(s) that identified rows by identifier were removed; the analysis is otherwise unchanged. Originally posted . See docs/person-identifier-convention.md (#2624).

---

Another instance of this scraper gap, from the dept-faculty-roster (not medicine.yale.edu) host, with a concrete and already-verified fix path.

Was operator_review, blocked only by thin_description + missing_card_description.

Mechanism (distinct sub-case worth noting): this is NOT source-empty.

1. The dept-faculty-roster scraper seeded the entity from the roster INDEX page (/people?page=1), whose only description content is the thin "Fields of Interest" list, materialized as the Studies <research-area chips> echo.
2. The per-person profile URL was never recorded into sourceUrls (it only survived in fieldProvenance/studentDecisionExplanation), so data-quality(descriptions): student_ready FACULTY_RESEARCH_AREA seeded from department faculty-index pages serve directory-thin area-echoes; linked personal/lab prose unharvested #1716's automated re-derive (candidateDescriptionLabsFromDocs, which reads doc.sourceUrls) could never reach it.

Verified fix path: the repo's own extractOfficialResearchDescription cleanly grounds the real bio from that per-person profile page (isDescriptionGroundedInSource: true). So the durable scraper fix is: dept-faculty-roster should follow the roster index to each per-person profile page, record that profile URL into sourceUrls, and harvest its bio into a fullDescription observation (same as the medicine.yale.edu investigation proposed here). That would let the existing #1716 machinery cover this cohort automatically instead of leaving it on operator_review.

1 line(s) omitted here: they identified rows by person-bearing identifier. See the predicate and counts in this comment.

[quntao-z]

Redacted 2026-09-13: this comment paired person-bearing identifiers with claims about those people, and this repository is public. Editing cannot remove published text, because GitHub serves every prior revision to anyone without an account, so the original was deleted and this rewrite posted in its place. 2 line(s) that identified rows by identifier were removed; the analysis is otherwise unchanged. Originally posted . See docs/person-identifier-convention.md (#2624).

---

Held source-empty instance with a poisoned namesake observation (t2402 shard audit):

Kemp Professor of Chemistry), activeAtYale, researchAreas ["Chemistry"]. Served short/full are EMPTY and sourceUrls is EMPTY, so it is a member of the source-empty held cohort - not promotable (no homepage, no NIH/NSF grant to ground on).

Extra wrinkle worth flagging for the repair pass: its only description observation (lab-microsite-description-llm, non-superseded) is a WRONG-PERSON graft - it describes "Maria Johnson, a Research Associate in the office of the director at CORE ... School of Medicine ... Robert Wood Johnson Clinical Scholars Program ... staff writer", a different person matched only on the surname "Johnson". It is not currently served (entity full is empty, so the materializer did not adopt it), but a future full re-materialize could pull this wrong-person text onto the card. This is the namesake description source-fusion class (closed #1537; related to the namesake grafts in #1407).

[quntao-z]

Redacted 2026-09-13: this comment paired person-bearing identifiers with claims about those people, and this repository is public. Editing cannot remove published text, because GitHub serves every prior revision to anyone without an account, so the original was deleted and this rewrite posted in its place. 2 line(s) that identified rows by identifier were removed; the analysis is otherwise unchanged. Originally posted . See docs/person-identifier-convention.md (#2624).

---

Another instance, and a concrete scraper-gap data point on a long-tail departmental host (economics.yale.edu, FAS):

2 line(s) omitted here: they identified rows by person-bearing identifier. See the predicate and counts in this comment.

• The synthesizer never captured any of it: the only description signal was descriptionSource='PI_PROFILE_SYNTHESIS' with profileSynthesisDescription set to the CV/education block ("Ph.D., Economics, University of Minnesota, 2003 ..."), which is not research-focused prose, so the card was correctly held.

Takeaway for the deeper dive: Yale department Drupal profile pages expose a labeled 'Fields of Interest' block (and often a related-publications list) that the profile-synthesis lane is not extracting; it falls through to the degrees/CV block. Capturing the 'Fields of Interest' section into a source-backed fullDescription/researchAreas observation would recover a large share of the FAS departmental long tail, not just economics.

Data repaired row-by-row (Development): wrote a source-backed short + full derived from the Fields of Interest and Related Research, recomputed to student_ready, Meili synced. Recording here so the scraper-side fix (not attempted in this row-by-row pass) has the instance.

[quntao-z]

Redacted 2026-09-13: this comment paired person-bearing identifiers with claims about those people, and this repository is public. Editing cannot remove published text, because GitHub serves every prior revision to anyone without an account, so the original was deleted and this rewrite posted in its place. 3 line(s) that identified rows by identifier were removed; the analysis is otherwise unchanged. Originally posted . See docs/person-identifier-convention.md (#2624).

---

Second-opinion hold-fix data point (Development), and a distinct actionable sub-class.

Held at operator_review on thin_description + missing_card_description (the description-family hard blockers). Lead resolved correctly (Bhullar, PI); not a duplicate (only sibling is archived).

So this was not "the source page has no blurb" - it was a fan-out gap: dept-faculty-roster minted the LAB from earth.yale.edu/faculty and stored the lab websiteUrl/sourceUrls, but the lab-microsite description extractor never followed the external lab-site URL to harvest a fullDescription/shortDescription observation. The entity therefore sat blank/held despite carrying a link to a source that clearly has the prose.

Suggested refinement to the follow-up: in addition to the medicine.yale.edu blurb-capture work, add a lane that fans lab-microsite description extraction out to the external lab-site URLs already captured in sourceUrls/websiteUrl (e.g. *.org lab sites), not just Yale-hosted lab index pages. These are high-yield because the URL is already known and verified as the lead's site.

Secondary defect on the same row (unprovenanced area graft): researchAreas was ["Planetary Science"] with NO fieldProvenance.researchAreas - a coarse department-level area derived from the "Earth and Planetary Sciences" department, not his actual research (vertebrate paleontology / evo-devo). This is the department-name-as-area family (cf #1407 / #1800) surfacing on a dept-faculty-roster LAB.

Fully sourced, no quality decrease. Development only.

[quntao-z]

Split of this cohort by cause, measured on Development

Read through the real gate planner (planStudentVisibilityGate, dry run) and the real description representation (buildResearchEntityPublicDescriptionRepresentation), so every class below is judged on the served surface rather than on stored fields. Non-archived rows only, 2026-09-22.

Instrument caveat first. Another session writes Development concurrently, so the cohort total read 1001, 990, 953 and 925 across four runs inside ten minutes. Only a per-row diff over one fixed row set compares reliably; do not read a delta of tens as an effect.

Cohort: rows the gate holds out of student_ready carrying missing_card_description, thin_description or missing_description. 925 before the fix below landed, 824 after. The triage verdict's 1,061 and the body's 442 are the same population measured at other moments.

"Releasable" counts rows carrying no hard blocker outside the description family, where the family includes the two derived reasons blank_public_description and public_description_invariant_failed. Counting those two as independent blockers reports the releasable set as zero, which is what an earlier pass concluded.

Sub-population	Rows	Releasable	What it needs
A. No stored prose at all	511	110	Acquisition. Neither fullDescription nor profileSynthesisDescription. This is the population this issue was opened about.
B. Body fails the quality bar	127	92	A body rewrite, not a card. Dominant flag is area-echo-fallback at 47, meaning the stored body is the research-area echo template; then blank 40 after sanitize, source-news-fragment 8, appointment-only 6.
C. Body under 120 characters	76	58	Acquisition or synthesis from a longer source.
D. Card derivation only	197	135	Body prose is useful and the card is not. The slice landed below took this from 197 to 117 and its releasable set from 135 to 77.
E. Profile-synthesis text only	7	7	Acquisition.
F. Card complete and still held	2	0	Held by something outside this family.

Other hard blockers co-occurring across the cohort, which is why the releasable column is far below the rows column: citations_identify_no_person 302, missing_lead 297, duplicate_risk 166, exact_url_duplicate_risk 149, grant_only_no_current_yale_source 19.

What just landed, and against which class

Class D was a card-derivation defect on rows already carrying useful prose, exactly as the 2026-09-21 decision predicts. Its root cause is a card-length bar with two owners, filed as #2959 and fixed in #2967 because it is not this issue's scraper gap. It also corrected served copy on rows that were never held at all.

• 61 rows moved operator_review to student_ready, verified by re-reading all 61 through getResearchGroupDetail after applying the gate on Development: 61 of 61 served, 0 blank cards.
• 430 of the 462 student_ready rows whose stored card line sits past the 200-character box now serve their own prose through the detail route, against 0 before. Re-read on the merged head; the band total moves because concurrent scraper and repair sessions keep rewriting shortDescription.

Named leads for whoever picks this up next

1. Class B is the best remaining code-shaped target at 92 releasable rows. 47 of them store the research-area echo template as their body. That template's output is rejected by design, so the question is whether a body can be synthesized from the same source, not whether the bar should be relaxed.
2. Class A at 110 releasable rows is the acquisition target this issue names. The triage verdict's host ranking stands: ysph.yale.edu leads, then medicine.yale.edu. Check the donor-page finding in data-quality(scrapers): all 517 dept-ysph rows cite a donor fundraising page and a directory index as their only evidence #2460 first, because all dept-ysph rows citing a donor page would explain the first figure as one cause rather than 331.
3. Class D's residue is 117 rows, 77 releasable. The deterministic derivation produces nothing from prose the body bar calls useful. The grounded LLM card lane already exists (resolveGroundedCardDescription with synthesize, driven by research-homes:backfill-descriptions) and is the instrument for it. It needs a cost and yield measurement before a corpus-wide run.
4. The producer is uncorrected. The card-synthesis prompt bounds a card by words, not characters, which is what puts a line in the 201-280 band in the first place. Correcting it changes CARD_SYNTHESIS_PROMPT_HASH and re-synthesizes gated rows on the next sweep, so it needs its own measurement rather than a quiet edit.
5. 58 rows now serve the first sentence of a multi-sentence stored card line, and 23 of those open with "This ". A handful are ungrammatical revoicer residue of the form "This primary research interests are focused on ...". That belongs to the revoicer family rather than to length, and it is now visible on a card instead of hidden behind a chip echo.

Do not re-derive any of this from stored fields. shortDescriptionQuality on a stored short answers a different question from the one the gate asks, because four sanitizers and a chrome strip run before the gate sees the value.

[quntao-z]

The four named leads, measured. Three have a deterministic ceiling of zero.

Development, 2026-09-22, read through planStudentVisibilityGate and buildResearchEntityPublicDescriptionRepresentation rather than any stored column. Non-archived rows only. Cohort is the rows the gate holds carrying missing_card_description, thin_description or missing_description.

Same instrument caveat as the previous pass: another session writes Development concurrently. The cohort read 787 here against 824 at the end of the previous pass. Only a per-row diff over one fixed row set compares reliably.

Updated sub-population counts

Class boundaries are the previous pass's, with the length test applied before the quality test so the two tables line up.

Sub-population	Rows	Releasable	Previous pass
A. No stored prose at all	495	108	511 / 110
B. Body fails the quality bar	104	73	127 / 92
C. Body under 120 characters	68	54	76 / 58
D. Card derivation only	113	74	117 / 77
E. Profile-synthesis text only	7	3	7 / 7
F. Card complete and still held	0	0	2 / 0

Co-occurring hard blockers outside the description family, which is why the releasable column sits far below the rows column: citations_identify_no_person 287, missing_lead 264, duplicate_risk 115, exact_url_duplicate_risk 104, grant_only_no_current_yale_source 19, profile_fallback_only 6.

Lead 1, Class B. Ceiling zero. The body is a chip echo we minted.

area-echo-fallback is the dominant flag on 76 releasable rows across B and C, not 47. Read directly, those bodies are the row's own researchAreas chips restated as a sentence. fieldProvenance.fullDescription names the writer: lab-microsite-undergrad-llm 35, lab-microsite-description-llm 24, dept-faculty-roster 12, ysm-faculty-directory 2, department-undergrad-research 1, absent 2. Median body length 119 characters.

Zero of the 76 carry a non-served fullDescription observation that passes the bar. There is no better value in the ledger for a confidence change, a re-resolution, or a rematerialize to find. The lead asked whether a body could be synthesized from the same source; the source is the chip list, so it cannot.

Worth recording separately: runResearchDescriptionBackfill takes its source text from the stored body, so on these rows it would ask an LLM to ground a research description in our own synthetic echo, which makes its grounding check vacuous there.

Lead 2, Class A. Deterministic ceiling about 7 of 108, and refused by the 2026-08-29 decision.

Of the 108 releasable rows, 12 cite no URL at all and 7 fail to fetch. Fetching the rest serially with a browser user agent and running the repo's own extractOfficialResearchDescription over them yields a body that passes the real serve invariant on 31.

But 20 of those 31 are flagged by the repo's own hygiene detectors as a career biography, a high-confidence person bio, or person-centric prose. Hand-reading the remaining 11 finds a bibliography entry, a leadership-programme marketing blurb, a pull quote, and a truncated question stem. So the genuine deterministic yield is about 7 of 108, and a lane built on it would put a CV on the other two dozen cards, which is exactly what the 2026-09-22 profile JSON-LD entry refuses.

This reproduces the 2026-08-29 decision at cohort scale and across kinds, where it was established on a probe of 27 FACULTY_RESEARCH_AREA pages. The pages do carry the prose: 26 of the 31 extractions came from medicine.yale.edu, the host this issue named in 2026-08. What cannot be done is copy it.

Host ranking over the releasable Class A rows, which corrects the triage verdict's: medicine.yale.edu 60, music.yale.edu 11, orcid.org 9, reporter.nih.gov 4, then a long tail of departmental hosts with 1 to 3 each. ysph.yale.edu does not appear, so the #2460 donor-page finding does not explain this set. By type: 76 FACULTY_RESEARCH_AREA, 29 LAB, and one each of CENTER, INITIATIVE, INSTITUTE. The 32 non-FRA rows are outside research-entity:fra-profile-synthesis by construction.

Lead 3, Class D residue. Ceiling 6 rows, and #3046 takes them.

The standing-answer entry measured deterministic card derivation at 12 of 100 and recorded the rule that a deriver returning text overstated the gate's verdict fourfold. Re-measured after #2967 landed, the ceiling is 6 rows corpus-wide of 4,743 non-archived, none of them already student_ready, 4 of them releasable.

#3046 takes those 6. It substitutes a derived line only where the served card fails the gate's own kind-aware card bar and the derived line clears it, so it can never return empty for a non-empty line and never replaces a line the gate would have accepted. Before and after over one fixed row set of 4,756 planned rows: 6 cards changed, 3 rows operator_review to student_ready, 0 the other way, 0 losing the serve invariant, description-held 788 to 782.

Also in Class D and not addressable this way: 17 of the 74 releasable rows serve a blank card while their body passes the bar, because they carry zero researchAreas so there is no chip fallback and the deriver returns nothing. Hand-reading them, roughly half carry a career-biography body where a blank card is the correct refusal, so this is a body defect rather than a card defect on those rows.

Lead 4, the producer. Unmeasured, unchanged.

The card-synthesis prompt still bounds a card by words rather than characters. Separately, the two LLM description lanes above mint a chip restatement as a fullDescription, a value the gate can never accept. Neither releases a row on its own, and the prompt change moves CARD_SYNTHESIS_PROMPT_HASH and re-synthesizes gated rows on the next sweep, so each needs its own measurement.

What the next bounded experiment should be

There is no deterministic slice left above 6 rows. The remaining cohort is acquisition and synthesis cost, so the next step is a cost-and-yield measurement rather than another lane:

1. Run research-entity:fra-profile-synthesis against a sample of the releasable Class A FACULTY_RESEARCH_AREA rows whose cited page fetched successfully, and record tokens spent per row promoted. The 2026-08-29 probe suggests the page carries the prose on nearly all of them, so the unknown is conversion and cost, not availability.
2. Run the grounded card lane (resolveGroundedCardDescription with synthesize) against a sample of the 74 releasable Class D rows, same measurement.
3. Only then decide a corpus-wide run. Nothing here should be extended to the 32 non-FRA Class A rows without first naming which lane owns an organizational row with no prose, because the FRA lane rejects them by construction.

Recorded in docs/decisions.md so the refusals are easier to find than the temptation.