Appearance
Research: what one SDS costs to curate
Question: What does it cost a human to admit one SDS to the Corpus — about an hour per document, or about a day? (D-32, the two-document spike ADR-0017 owes stage 2 sizing.)
Sources: primary only — two public manufacturer SDS obtained over the network and read end to end, plus a third consulted opportunistically. Text layers extracted with pypdf 6.16.2; every span quoted below was read against its own page.
Findings continue the R- sequence. The highest existing id was R-34; this note starts at R-35. No id here is reused or renumbered.
Read this before any number in this note
These documents are the proxy corpus under D-28. Public manufacturer SDS, used as scaffolding to measure Curation. They are not The Plant's binder and must never be presented as the Corpus (D-29). No document in this spike came from The Plant.
The timing figures are not measured human time. The spike was executed by an agent, not by a person with a PDF viewer. What was genuinely measured is agent elapsed time on born-digital PDFs for the mechanical stages — locating, extracting, structuring. That is a lower bound on machine work and it measures nothing at all about the human stages. Every figure below is labelled either measured (machine) or derived estimate (human), and the derived ones carry their assumptions inline. A number shaped like "eight minutes per document" would answer D-32 wrongly in the direction that hurts, so no such number is offered as the answer.
The short answer: roughly half a day to a full day per document, decisively not an hour, and review and reconciliation dominate — not parsing, not translation drafting. See R-35.
Coverage gate. This note declares R-35 through R-45. New R-nn ids turn the coverage gate RED until a later stage accounts for them. That is expected and correct; do not "fix" it here.
The documents
Doc A — the English-only case. Acetone, Fisher Scientific Company, revision date 18-Dec-2025, revision number 9, created 28-Apr-2009. 10 pages, born-digital. Obtained from https://www.fishersci.com/store/msds?partNumber=A18P4 (US, English).
Doc B — the Thai-published case. อะซีโตน : Acetone, APC (Asia Pacific Chemical, Bangkok), document code 02-001-0, validation date 03-Jan-2018, prepared by APC Laboratory. 11 pages, born-digital, bilingual Thai/English field labels with Thai-only field values. Obtained from https://www.apcbkk.com/pdf/products/Ketones/SDS/Acetone_TH.pdf.
Doc C — consulted, not part of the spike. ACETONE, Modern Chemical Co., Ltd., Bangkok. 8 pages, Thai. Pulled while searching for Doc B and read because it disagreed with Doc B. It is the source of R-39, R-41 and R-43 only. The spike itself is two documents, per D-32. Obtained from https://www.modernchemical.co.th/products/pdf/msds-th/Ketones/SDS_Acetone_MCC_Th.pdf.
Under D-18 these are three Chemical Records, not one chemical — three product names, three suppliers. Three SDS for Acetone confirming D-18's shape is worth one line; R-41 records where the key breaks.
Findings
R-35 One document costs roughly half a day to a full day, and review dominates
Measured (machine), per document:
| Stage | Doc A | Doc B | Note |
|---|---|---|---|
| Locating the document | untimed | ~1 min | Doc A's timer started after it was already in hand. Doc B was found on the first search and downloaded on the first URL. Neither was hard. |
| Extracting the 16 SDS Sections | < 1 min | < 1 min | text layer present, section headers regular |
| Structuring the fields | ~2 min | ~4 min | Doc B slower: field values are Thai, labels are bilingual |
| Total mechanical | ~3 min | ~5 min |
Derived estimate (human), per document:
| Stage | Doc A (English-only) | Doc B (Thai-published) |
|---|---|---|
| Locating, obtaining, recording Provenance | 10–20 min | 10–20 min |
| Reading all 16 SDS Sections and marking Source Spans | 45–90 min | 45–90 min |
| Curated Translation of the safety fields | 60–120 min | n/a — replaced by transcription verification, 45–90 min (R-40) |
| Review: verifying each field against its page, and reconciling restrictions across sections | 90–180 min | 90–180 min |
| Total | ~3.5–7 h | ~3–6.5 h |
Doc B's lower total is an artifact of band arithmetic, not a finding. R-40's point is that the two paths cost about the same by different routes — Doc A pays in Curated Translation review, Doc B pays in transcription verification. Do not read the table as evidence that a Thai-published SDS is cheaper.
Assumptions. One curator, competent in Thai and in reading an SDS, working from a born-digital PDF with a machine-assisted first pass at extraction. The review band is wide because it contains the reconciliation work R-38 and R-39 describe, which is judgement and does not compress. The Curated Translation band assumes drafting is assisted and the human cost is reviewing the draft, per D-11.
The distinction D-32 asked for lands on the "day" side — a working day per document, not an hour. Not because parsing is hard — it is minutes (R-36) — but because the review stage cannot be shortened without giving up the thing ADR-0015 exists to protect. The Thai-published shortcut D-11 hopes for saves less than expected (R-40): it removes translation review and substitutes transcription verification.
Source: this spike, Doc A and Doc B, walked end to end · Confidence: medium — the mechanical figures are measured, the human figures are reasoned from what the documents contain, not observed. What would raise it: one curator timed on one document with a stopwatch. R-45 is the standing caveat.
R-36 The documents did not fight the extractor; mechanical parsing is not the cost
All three documents are born-digital with a complete text layer. Zero scanned pages. All sixteen SDS Sections are present in all three, numbered 1–16 in order, with no gaps, no duplicated numbers and no vendor renumbering. The feared failure did not occur, and the absence is a result.
What did degrade, and none of it is expensive:
- Two-column layout interleaves. In Doc A the supplier address block is emitted in the middle of Section 2, between the classification line and the hazard categories. Reading order follows the page's physical layout, not its logical structure.
- Tables lose their grid. Doc A's Section 8 exposure-limit table (five occupational limit columns) extracts as an unlabelled run of values. Doc A's Section 9 loses most of its physical properties entirely —
Appearanceextracts with no value after it. - GHS pictograms are images. Doc B's Section 2 pictogram field extracts empty. A pictogram is not Safety-Critical Content by itself, but a field that silently extracts as empty is indistinguishable from a field that is genuinely Not Stated.
None of these touch the fields a Chemical Record needs, except through the review burden they add — which is R-35's point.
Source: Doc A pp.1–10, Doc B pp.1–11, Doc C pp.1–8 · Confidence: high
R-37 Section 8 contains no restriction spans in either spike document
SDS Section 8 Exposure controls / personal protection is the section ADR-0015's worked example is set in. Reading both spike documents' Section 8 in full, hand-verified:
- Doc A Section 8: zero restriction spans. A keyword scan produced one apparent hit,
to prevent skin exposure, which is a purpose clause, not a restriction. There is no "not compatible with", no "do not use", no exclusion, no negation anywhere in the section. - Doc B Section 8: zero restriction spans. Every field is a positive instruction — wear this respirator, wear these gloves, wear this eye protection.
Restrictions are common in the document — they are simply not in Section 8. An approximate keyword scan of Doc A puts them in Section 7 Handling and Storage (6), Section 10 Stability and reactivity (4), Section 2 (3), Section 11 (3), Section 5 (2), Section 1 (1). Doc B's Thai negation terms (ห้าม, หลีกเลี่ยง, อย่า, ไม่ควร) cluster the same way: Section 6 (9), Section 7 (8), Section 2 (5), Section 4 (4), Section 10 (1), Section 13 (1), Section 8 (0).
The per-section distribution is an approximate scan and should be read as a shape, not a count. The Section 8 = zero claim is hand-verified; both sections are short enough to read whole.
Source: Doc A p.4 Section 8; Doc B p.6 Section 8 · Confidence: high
R-38 The restriction that would falsify a Section 8 PPE span lives in Sections 7 and 10
This is the finding that matters most, and it does not fit the "restrictions are rare or common" question the spike set out to answer. Both halves are true at once.
D-26's premise is strengthened. ADR-0015 hypothesised a Section 8 span reading "not compatible with nitrile" being silently dropped. A materially identical case is present in a real document:
- Doc B Section 8 Hand Protection (p.6): หากต้องมีการสัมผัสกับสารเคมีควรสวมใส่ถุงมือชนิดที่ทนต่อ สารเคมีชนิดนั้นได้ดี เช่น ถุงมือไนไตร หรือ นีโอพรีน — wear gloves resistant to the chemical, for example nitrile or neoprene gloves.
- Doc B Section 7 Storage (p.5): ห้ามเก็บรวมกันกับยางธรรมชาติ, ยางบิวทิล, ยางไนไตร, ยางนีโอพรีน, รวมทั้งพลาสติกทั่วไปและอลูมิเนียม — do not store together with natural rubber, butyl rubber, nitrile rubber, neoprene rubber, general plastics or aluminium.
- Doc B Section 10 Conditions to Avoid (p.8): ยางธรรมชาติ, ยางสังเคราะห์ — natural rubber, synthetic rubber.
D-26's mechanism is insufficient. D-26 scopes the rule within a section: every restriction span in a section a selection draws from must be carried through. A PPE card built from Section 8 alone satisfies D-26 in full — there is nothing in Section 8 to carry (R-37) — and still shows a Handler "nitrile gloves", verbatim, correctly sourced, with no sign that the same document says nitrile rubber is incompatible with the chemical three pages earlier. The exact failure ADR-0015 was written to prevent passes its own rule.
This is a tension, not a settled contradiction, and that is the expensive part. A storage incompatibility for bulk containment is not automatically a prohibition on a splash glove; Doc C's Section 8 says nitrile is the long-duration choice and PVC or neoprene the splash choice, which is arguably the opposite ordering. A restriction detector can surface this. It cannot resolve it. Resolution is a Safety Officer's judgement, and it is inside the review stage R-35 says dominates.
Source: Doc B pp.5, 6, 8 (Sections 7, 8, 10); Doc C p.5 Section 8 · Confidence: high
R-39 Two Thai SDS for the same chemical disagree about the same glove material
Doc B's Section 7 forbids storing Acetone with nitrile rubber. Doc C's Section 8 (p.5) recommends ถุงมือยางไนไตรล์ — nitrile rubber gloves — for prolonged contact, reserving PVC or neoprene for accidental splash. Same chemical, same country, two suppliers, opposite advice on the same material.
Under D-18 these are two Chemical Records and both are legitimately in the Corpus. Nothing in the decisions so far says what the product does when two records the user could plausibly be looking at give conflicting Safety-Critical Content. Cross-document reconciliation is Curation work that no current decision accounts for, and it does not appear until the Corpus holds a second SDS for a chemical — which The Plant's binder, holding multiple suppliers of the same solvent, will.
Source: Doc B p.5 Section 7; Doc C p.5 Section 8 · Confidence: high
R-40 A Thai-published SDS does not remove the Curation cost; it changes its shape
ADR-0009's premise survives: Thai-published SDS exist and are obtainable. Two were found in one search. What does not survive is the implied cost saving.
Doc B's Thai text layer is systematically corrupted. The character ำ (SARA AM) never appears correctly anywhere in the document — 0 correct occurrences, 87 broken — decomposing into a space plus า. So จำหน่าย extracts as จ าหน่าย, น้ำ as น ้า, ทำ as ท า. Tone marks pick up stray spaces elsewhere (ทั่ วไป, รั่ วไหล). Doc C, from the same country, is almost clean: 643 correct against 12 broken.
Two consequences:
- Every Thai span must be verified character by character against its page before it can be treated as Extractive. Repairing
าback toำis a mechanical substitution, but a substitution applied to Safety-Critical Content without a human reading the page is a silent rewrite — the thing D-02 forbids, arriving through the text layer instead of through a model. Every Thai span quoted in this note is a reconstruction and would need that verification in a real Curation. - Corpus sizing cannot assume Thai documents are cheap. Cleanliness varied between 100% corrupt and 2% corrupt across two documents from the same country. Which one a given SDS is cannot be known before opening it.
So D-11's ordering (prefer Thai-published) stands. D-11's bottleneck claim needs widening: the bottleneck is not human translation review, it is human verification of the Thai text, whichever way that Thai arrived.
Source: Doc B, full text layer, character counts over the extracted text; Doc C, same · Confidence: high
R-41 One of three SDS carries no revision date, so D-18's identity triple cannot be formed
D-18 identifies a Chemical Record by product name, supplier and revision date. Doc A carries all three cleanly (revision date 18-Dec-2025, revision number 9). Doc B carries a Validation Date : 03-Jan-2018 and a document code, which is close enough that a curator would use it — but "validation date" is not "revision date", and treating them as the same field is a curator's judgement, not a read.
Doc C carries no date of any kind. There is no revision date, no issue date, no version number and no revision history anywhere in its eight pages. Its Section 16 is a liability disclaimer and nothing else. Its Provenance can record supplier and how the copy was obtained; it cannot record a revision date, so under D-18 the record has no complete identity and D-06's staleness signal has nothing to read.
Source: Doc A p.1; Doc B p.1; Doc C, all pages · Confidence: high
R-42 Section 8 PPE text can be too generic to satisfy the per-item PPE rule
The glossary requires PPE to be recommended per item with a stated reason — "nitrile gloves", never "gloves". Doc A's Section 8 Skin and body protection reads, in full:
Wear appropriate protective gloves and clothing to prevent skin exposure.
No material, no standard, no glove type. Its Eye/face protection field defers to a regulation (29 CFR 1910.133 or European Standard EN166) rather than naming an item. Only the respiratory field is specific (Type AX. Brown. conforming to EN371).
Doc B and Doc C both name glove materials. So this is not universal — but a Corpus assembled from whatever The Plant's suppliers issue will contain documents that cannot support the PPE promise extractively, and the product cannot fill the gap by generating the missing material name without breaking D-02 and D-16.
Source: Doc A p.4 Section 8; Doc B p.6 Section 8; Doc C p.5 Section 8 · Confidence: high
R-43 Not Stated occurs in real SDS text, written by the supplier
Doc C's Section 8 ends: มาตรการสุขอนามัย: ไม่มีข้อมูล — Hygiene measures: no information. Doc A's Section 5 carries Sensitivity to Mechanical Impact — No information available and Sensitivity to Static Discharge — No information available.
The Corpus will therefore contain fields whose Not Stated is asserted by the supplier, not inferred from an absent field. Both surface to the user as Not Stated, but only one of them is a fact about the document; the other could equally be an extraction failure (R-36's empty pictogram field). Curation is where the two are told apart, and that is review work.
Source: Doc C p.5 Section 8; Doc A p.3 Section 5 · Confidence: high
R-44 Spill Response is not always one ordered procedure
The glossary defines Spill Response as the ordered procedure for a release, whose order is semantic. Doc B's Section 6 is not shaped that way. It has a protective-measures preamble, then two alternative procedures selected by spill volume — หกหรือรั่วไหลเล็กน้อย (< 200 LT) and หกหรือรั่วไหลมาก (> 200 LT) — plus a notification note. Within each branch the steps are ordered; across the branches they are exclusive alternatives, and the large-spill branch carries a restriction the small-spill branch does not (ห้ามใช้น้ำสาดของเหลวที่ตกค้าง — do not hose down the residue with water).
Doc A's Section 6 is a flat three-field structure with no branching. Both shapes are legitimate GHS. A Chemical Record that models Spill Response as one ordered list can hold Doc A and loses Doc B's branch condition — and the branch condition is the thing that decides which restriction applies.
Source: Doc B p.5 Section 6; Doc A p.4 Section 6 · Confidence: high
R-45 The proxy corpus systematically understates what Curation will cost
Every figure in R-35 was produced against a favourable sample, and the favour is systematic, not accidental:
- All three documents are born-digital manufacturer PDFs published for download. The Plant's binder is a binder — photocopies, scans, and sheets that arrived with a delivery. R-36's "zero scanned pages" is a property of the proxy corpus, not a forecast.
- All three are Acetone, a Pure Substance with a CAS Number and a well-standardised SDS. D-28 names Acetone, IPA and Toluene precisely because they are the tractable cases. The chemical that motivates the whole Corpus decision, น้ำยาล้างชิ้นงาน, is a Formulated Product with no public SDS at all (D-29) — so its Curation cost cannot be measured here by construction.
- All three were located in under a minute. Locating a document inside The Plant's holdings is a different activity with a different bottleneck, and it is not this one.
R-35's band should be read as a floor for The Plant's binder, not an estimate of it.
Source: D-28, D-29, and the composition of this spike's sample · Confidence: high
One worked Chemical Record
A data-shape illustration only — the fields of one record with the Source Span each safety field is taken from. Not a schema, not a specification, not code. Built from Doc B, the Thai-published case.
Every Thai span shown is a reconstruction of a corrupted text layer (R-40). In a real Curation each one is verified against its PDF page before it is admitted; here they illustrate shape, and the corrupted form of the first one is shown alongside so the problem is visible rather than described.
Identity
| Field | Value |
|---|---|
| Product name | อะซีโตน : Acetone |
| Supplier | APC (Asia Pacific Chemical), Bangkok |
| Revision date | 03-Jan-2018 — recorded from a field labelled Validation Date; see R-41 |
| CAS Number | 67-64-1 (optional under D-18; present because Acetone is a Pure Substance) |
| Record type | Pure Substance |
Provenance
| Field | Value |
|---|---|
| Supplier | APC (Asia Pacific Chemical), Bangkok |
| Revision date | 03-Jan-2018 (Validation Date), document code 02-001-0, prepared by APC Laboratory |
| How the copy was obtained | Downloaded from the supplier's public website, 2026-09-02. Not from The Plant. Proxy corpus under D-28. |
| Published language | Thai (bilingual field labels, Thai-only field values) |
| Curated Translation | Not required for display. Transcription verification required — the text layer's SARA AM is 100% corrupt (R-40). |
Safety fields
Hazards — Source Span: Section 2 Hazards Identification, p.1
ของเหลวไวไฟ : ประเภท 2 · การระคายเคืองตา : ประเภท 2 · เป็นพิษต่ออวัยวะที่สัมผัสครั้งเดียวได้ : ประเภท 3 H225 ไอระเหย และของเหลวไวไฟสูง
Health effects — Source Span: Section 2 Health Hazard, p.1
ไอระเหยอาจทำให้เกิดอาการมึนงงและวิงเวียนศีรษะ ระคายเคืองต่อดวงตา ผิวหนังและระบบทางเดินหายใจ
As extracted, before verification: ไอระเหยอาจท าให้เกิดอาการมึนงงและวิงเวียนศีรษะ... — the missing ำ in ทำ is the whole of R-40 in one word.
PPE — hands — Source Span: Section 8 การป้องกันทางมือ / Hand Protection, p.6
หากต้องมีการสัมผัสกับสารเคมีควรสวมใส่ถุงมือชนิดที่ทนต่อสารเคมีชนิดนั้นได้ดี เช่น ถุงมือไนไตร หรือ นีโอพรีน
⚠ Restriction from another SDS Section, carried per R-38 (not per D-26, which does not reach it) — Source Span: Section 7 การเก็บรักษา / Storage, p.5 ห้ามเก็บรวมกันกับยางธรรมชาติ, ยางบิวทิล, ยางไนไตร, ยางนีโอพรีน, รวมทั้งพลาสติกทั่วไปและอลูมิเนียม Unresolved: the same document names nitrile as a glove material and nitrile rubber as an incompatible material. A Safety Officer resolves this at Curation time; the product must not resolve it by choosing one span.
PPE — eyes — Source Span: Section 8 การป้องกันทางตา / Eye Protection, p.6
สวมใส่แว่นครอบตาหรือหน้ากากป้องกันสารเคมี
PPE — respiratory — Source Span: Section 8 การป้องกันทางการหายใจ / Respiratory Protection, p.6
สวมหน้ากากกรองไอสารเคมีอินทรีย์ชนิด NPF 400 (Gas Only) หากอยู่ในที่ที่การระบายอากาศไม่ดีในที่อับหรือ ห้องทึบให้สวมเครื่องช่วยหายใจชนิดมีถังอากาศในตัวมาตรฐาน NPF 2000
Storage — Source Span: Section 7 การเก็บรักษา / Storage, p.5
เก็บในสถานที่ที่มีการถ่ายเทอากาศดี ห่างจากแสงแดด แหล่งกำเนิดประกายไฟและความร้อน ไม่ควรเก็บในที่ที่ อุณหภูมิสูงเกิน 30 องศาเซลเซียส จัดเก็บไว้ให้ห่างจากสารออกซิไดส์ซิ่ง ห้ามเก็บรวมกันกับยางธรรมชาติ, ยางบิวทิล, ยางไนไตร, ยางนีโอพรีน, รวมทั้งพลาสติกทั่วไปและอลูมิเนียม
First Aid — Source Span: Section 4 การปฐมพยาบาล / First Aid Measures, pp.3–4. Structured by exposure route, in the document's own order.
| Route | Span |
|---|---|
| Inhalation (การหายใจเข้าไป, p.3) | ถ้าหายใจเข้าไปให้เคลื่อนย้ายออกไปที่อากาศบริสุทธิ์ ถ้าผู้ป่วยหยุดหายใจให้ช่วยผายปอด ถ้าหายใจลำบากให้ออกซิเจน นำส่งแพทย์ |
| Eye (การสัมผัสทางตา, p.4) | ให้ฉีดล้างตาทันทีด้วยน้ำสะอาดปริมาณมาก ๆ อย่างน้อย 10 นาที พร้อมกระพริบตาถี่ ๆ ขณะทำการล้าง นำส่งแพทย์ |
| Skin (การสัมผัสทางผิวหนัง, p.4) | ให้ฉีดล้างผิวหนังทันทีด้วยน้ำสะอาดปริมาณมากๆ อย่างน้อย 15 นาที พร้อมถอดเสื้อผ้าและรองเท้าที่เปรอะเปื้อนสารเคมีออก |
| Ingestion (การกลืนกิน, p.4) | ให้ผู้ป่วยดื่มน้ำปริมาณมากๆ ทันที ห้ามทำให้อาเจียน ห้ามให้กินนมหรือน้ำมันที่ย่อยสลายได้ ทำให้ผู้ป่วยหายใจสะดวก นำส่งแพทย์ |
The ingestion route carries two restrictions inside the First Aid span itself. This is the case D-26 does reach, and it is real.
Spill Response — Source Span: Section 6 ข้อปฏิบัติเมื่อเกิดการรั่วไหล / Accidental Release Measures, pp.4–5. Branches on spill volume (R-44); the branch condition is part of the content, not presentation.
Preamble, all spills: ระวังอย่าสัมผัสกับสารที่หกหรือระเหยออกมา ให้ถอดเสื้อผ้าที่ปนเปื้อนสารออกทันที · หยุดการรั่วไหลของสารเคมีโดยเร็วหากสามารถทำได้โดยปลอดภัย · นำสิ่งของหรืออุปกรณ์ที่อาจติดไฟได้ทั้งหมด ออกจากบริเวณ · ป้องกันการแพร่กระจายโดยการใช้ดินหรือทรายสร้างเป็นเขื่อนกั้น
Small spill (< 200 LT): ให้ถ่ายเทของเหลวด้วยวิธีกลไกเข้าสู่ภาชนะบรรจุที่ติดป้ายและปิดผนึกอย่างดี · ปล่อยของเหลวที่ตกค้างทิ้งไว้ให้ระเหยไปเอง หรือใช้วัสดุดูดซับ
Large spill (> 200 LT): ใช้รถบรรทุกสูบของเหลวจากถังที่หก · ห้ามใช้น้ำสาดของเหลวที่ตกค้าง แต่ให้ ปล่อยของเหลวที่ตกค้างทิ้งไว้ให้ระเหยไปเอง หรือใช้วัสดุดูดซับ · ขุดดินที่ปนเปื้อนสารเคมีออกและนำไปกำจัด
Escalation precedes both branches per the glossary — and has no Source Span. What the SDS supplies here is a notification obligation to the authorities (ควรแจ้งให้หน่วยงานราชการที่รับผิดชอบทราบ), which is not the same act. Directing a Handler to call 1669, evacuate or seek medical attention is product behaviour sourced from no SDS Section at all. Curation cannot record Provenance for it, because there is none to record.
Not Stated — Section 5's extinguishing-media detail and Section 9's several physical properties survive in the PDF but not in the text layer (R-36). Under Curation these are read from the page, not recorded as Not Stated. A field genuinely absent from the document is Not Stated; a field lost by extraction is a defect, and telling them apart is review work.
Open questions
Not findings. Each needs an owner decision or a wayfinder ticket. No R- ids.
- Does D-26's within-section scope need widening to the whole document? R-38 shows a Section 8 PPE card that satisfies D-26 completely and is still unsafe. Widening the scope makes "carry every restriction" much more expensive and may reintroduce the long-and-unreadable problem ADR-0015 explicitly rejected. This is ADR-0015's territory and is not decided here.
- What does the product do when two Chemical Records for the same chemical disagree? (R-39.) Show both with their suppliers? Prefer the record matching the container's supplier? Refuse to answer? Nothing decided so far covers it, and The Plant's binder will contain the case.
- Can a Chemical Record be admitted without a revision date? (R-41.) D-18 makes it part of the identity and D-06's staleness signal reads it. Refusing the document shrinks the Corpus; substituting the download date invents Provenance.
- What happens when Section 8 will not support the per-item PPE rule? (R-42.) Doc A says "appropriate protective gloves". Display the generic span as-is and break the glossary's promise, or mark the field Not Stated and lose real guidance?
- Does the Chemical Record shape need to carry a branch condition on Spill Response? (R-44.) The volume threshold decides which restriction applies.
- Who is the curator, and does one exist? R-35's band assumes a person competent in Thai and in reading an SDS and able to make the R-38 judgement. That is a Safety Officer's skill set, and Safety Officer time is the scarcest thing in the project. Twenty to forty documents at R-35's band is roughly two to five person-weeks.
- What does Curation cost on a photocopied binder page rather than a born-digital PDF? (R-45.) Every figure here is a floor, and the gap between floor and reality is not bounded by anything measured.
Coverage
Research is the head of the pipeline; stage 0 has no upstream artifact, so this table is empty by construction. The findings above are themselves the upstream input that stage 1 must account for.
| Upstream | Landed in | Evidence | Note |
|---|