Skip to content

Curation

The question this answers: where is every irreversible judgement in this product made, and what does it cost?

Curation, Reconciliation, Restriction, Provenance and Curated Translation are all glossary terms, and their definitions in CONTEXT.md are the ones the rulings use. The spec gathers the mechanics in Specification §5 — The Corpus, and Curation as the place judgement happens.

The ordered steps

The steps are grouped by colour. Which of them is a person's judgement, which is a tool that only prepares one, and where each is permitted to run are settled in ADR-0022 and ADR-0027 rather than here.

Why the judgement sits here

QuestionRuling
When Selection happens, and whether the runtime performs any of itADR-0022 (D-50)
How far a Restriction is searched for before a span may be displayed, and who resolves itADR-0022 (D-44)
Whether the cross-section detector may resolve what it flags, or run at display timeADR-0022 (D-49)
Why the earlier within-Section form of that rule did not holdADR-0015 — superseded; premise upheld, mechanism replaced
Whether extracted text may become a Source Span without being checked against the pageADR-0027
Who may curate, from where, and where the page image is renderedADR-0048 (D-110, D-111)
What happens to an uploaded file when its admission is refusedADR-0048 (D-112)
Whether Thai may be produced at the point of read instead of at CurationADR-0009
Whether a curator may supplement a span from outside the SDSADR-0025
Why a runtime that decides nothing is what makes offline operation reachableADR-0005
What the verification gate checks a document's verified pages against, and whether a verification may be withdrawnADR-0074, superseding ADR-0059
Whether a refusal from a curation write says which rule fired, or only that something was refusedADR-0062
What a successful selection carries onto the next oneADR-0067
What the Reconciliation gaps read reports when it returns no gaps, and which of the three reasons a curator is toldADR-0068, amended by ADR-0069 and ADR-0071
Whether a term that already carries a Curated Explanation may be explained againADR-0073, superseding ADR-0061
Whether a curated selection may be changed after it is made, and what the act isADR-0084 (D-169 … D-177), amended by ADR-0085
Why a Spill Response branch is corrected by an act of its own, and how many states its condition hasADR-0086, amended by ADR-0087
What a curator may see of a record they are correcting, and what a corrected row may never showADR-0088 (D-184, D-185)
Which language the curator's own surfaces show a Section heading in, and where those headings are heldADR-0060

The line these steps sit either side of is drawn in The content split. How choosing spans relates to writing them — what may be selected and laid out, and what may never be composed — turns on ADR-0002 and ADR-0012.

Correcting what Curation already recorded

Until 2026-09-20 the Corpus was append-only with no way back: a selection made wrongly stayed wrong, and ADR-0073 recorded that hole in its own text rather than closing it. ADR-0084 closes it for selections. Nothing is edited and nothing is deleted — a correction is a second act, with its own named Curator and its own stated reason, standing beside the judgement it replaces.

There are two acts and they are deliberately not one:

  • Supersession replaces a row, so the field goes on saying something. The replacement takes the position its predecessor held, and a passage that did not change is carried forward rather than restated — which is what keeps a correction from re-opening the whole document's Reconciliation.
  • Retraction removes a row, so the field says less. It is the only act in the product that takes content out of a Chemical Record, and where it leaves a curated field empty the Curator must assert that the supplier's document is silent — because an empty field renders as Not Stated, which is a claim about a document.

The safety property is that a correction cannot be quiet. A Source Span is never rewritten, so a Restriction standing over one stays live; the replacement is a new span nothing has judged, which re-opens the pair and holds publication until a named Curator resolves it. Trimming a prohibition out of a displayed passage therefore costs a recorded judgement rather than a silent edit.

What a Curator sees while correcting is itself ruled (ADR-0088): the record surface serves the live rows with the ids a correcting act names, and reports superseded and retracted ones as two counts. It never shows what a corrected row said — a retracted passage rendered beside a live one is two answers to the same safety question on a screen shaped to be believed.

What it costs

The measurement, and the reason the number that came out of it is treated as a floor rather than an estimate, are in ADR-0028 — Curation costs roughly a day per document, and the measured figure is a floor. What decides how many documents that per-document figure is multiplied by is a separate ruling: ADR-0034 — The Corpus is sized to one production area, not to a document count, which amends ADR-0028's sizing consequence and leaves its per-document finding standing. ADR-0034 states which part of each earlier ruling it touches.

Two further pointers rather than restatements:

  • Whether preferring a Thai-published SDS removes Curation cost or reshapes it is settled in ADR-0027 (D-57), with the character counts that forced the correction quoted in its Context.
  • The spec's Sizing section and its Declared Gaps carry the same two rulings in the form the report will read them, including the Study Area's chemical list as an outstanding request.

What a curator produces

Every item below is a Corpus artifact, in the sense CONTEXT.md gives that phrase. The ruling beside each one is where its shape and its standing are fixed.

Once per document

ArtifactRuling that fixes it
Provenance for the documentADR-0011
An extraction verified against the rendered page imageADR-0027
The stored Selection of Source Spans a Chemical Record showsADR-0022, ADR-0012
Reconciled content, including any caveat attached to a spanADR-0022
A Curated Translation, with its original-language span reachableADR-0009
A printable QR label for the recordADR-0030

Once for the Corpus, not once per document

ArtifactRuling that fixes it
The Routing Threshold below which Routing returns Not StatedADR-0031 (D-67)
The Escalation Trigger phrase listADR-0031 (D-70)
The Routing Fixture of exact per-case outcomesADR-0031 (D-71)

Whether these three sit inside the per-document figure or on top of it is recorded in the spec's Declared Gaps. How they are used is in Identification and Routing.

Where the work is checked

MechanismWhere
The assertion that every Restriction anywhere in a document was reconciled against every stored selectionSpecification, Seam 1
Assertions that Provenance is complete, that a Curated Translation has a reachable original, and that no record carries a date its document does not stateSpecification, Seam 1
What that assertion can and cannot prove about a human judgementADR-0022 (D-49)
Why no case in the Routing Fixture states a percentage or a threshold targetADR-0031 (D-71), on ADR-0017 (D-30)
The refusal codes, the Section headings and the gaps read, each pinned by a golden fixture both repositories read rather than each retypingADR-0062, ADR-0060, ADR-0071

Irreversibility is the property this page is about, and one of its edges is open. A curator who writes the wrong Curated Explanation for a term has no move inside the product today. What is and is not possible there, what it costs and why the alternative was refused are all stated in ADR-0073 rather than decided here. It is carried as an open hole in Current state.

Where the rest of this is

  • What happens to a document after it is admitted — sourcing, sync and the device — is in The Corpus lifecycle.
  • What the curated record is made of is in Data model.
  • What is still blocked on The Plant is in Current state.