Skip to content

The Corpus lifecycle

The question this answers: how does a supplier's document become an answer on a device with no signal?

Corpus, The Plant, Study Area, Not Stated and Corpus Unavailable are glossary terms, and CONTEXT.md holds the definitions the rulings use. The stages below are described in Specification §2, §4 and §5.

From binder to device

Sourcing: which documents, and how many

QuestionRuling
Whose documents the Corpus is made of, and whether public manufacturer sheets may stand in for themADR-0011 — The corpus is the case-study plant's own SDS, ingested as files
Whether a live integration with a plant chemical-management system is in prototype scopeADR-0011 (D-15)
What sizes the Corpus, now that a document count is no longer the unitADR-0034 — The Corpus is sized to one production area, not to a document count
Which artifact is requested from The Plant first, and why it is a smaller ask than the binderADR-0034 (D-80)
What each admitted document then costsADR-0028, summarised in Curation

An earlier sourcing ruling is kept and marked dead rather than deleted: ADR-0003 — superseded by ADR-0011; do not follow it.

The proxy set

The dotted lane in the diagram is the only path into Curation that is not The Plant's. What it is for, what it may never become, and the single document that distinguishes a real Corpus from a scaffolded one are settled in ADR-0016 — A proxy corpus develops the pipeline; it is never the delivered corpus.

That ADR carries an explicit abuse warning naming the phrase that would be used to ship the option ADR-0011 rejected. The proof test is restated in the spec's Declared Gaps, which is where a demonstration running on scaffolding is obliged to say so.

Getting the Corpus onto the device

QuestionRuling
Why the whole Corpus is on the device rather than fetched at the point of needADR-0005 — Safety content is offline-first; only the chatbot requires a network
Whether a cache of recently viewed chemicals would serve insteadADR-0005, rejected options
What the web platform costs this promise — storage eviction, installability, service workersADR-0020 (D-42)
Which premises of ADR-0005's rationale were retracted, and what it now rests onADR-0018 (D-34)
What a published version is, and whether its bytes may change afterwardsADR-0056
What a server with nothing published answers, and which version a device takes as its sentinelADR-0057
Why that refusal is never the same thing as Corpus UnavailableADR-0057, against ADR-0032
In what shape an SDS Section crosses the snapshot seam, so a Thai-first device can name itADR-0072

What syncs, in which direction, what is versioned, and why Corpus replacement is atomic are set out in Specification §2 and §4. The client and server split itself is in System architecture.

Between the "Corpus of record" and "Corpus snapshot" boxes in the diagram sits a publish, and it is a gate rather than an export: the parent repository's Seam 1 checker runs against the payload before a version exists at all. What a refusal from that gate carries is ADR-0062; what the payload is once it passes is ADR-0056.

When the device holds nothing verifiable

QuestionRuling
Whether a device holding no verifiable Corpus may answer the way a silent SDS doesADR-0032 — A missing Corpus is its own answer, and it is never Not Stated (D-72)
What "verifiably complete" is checked against, and whether a partial Corpus is readable at allADR-0032 (D-73)
What still works on a device holding nothingADR-0032 (D-74)
Why Escalation precedes procedure wherever it appearsADR-0006
What the recovery path costs, and what this ruling does not solveADR-0032, Consequences

Not Stated and Corpus Unavailable are separately defined in CONTEXT.md, and the distinction between a statement about a document and a statement about a device is the whole subject of ADR-0032. The rule that neither may be produced by generated prose is in The content split.

Where the lifecycle is checked

MechanismWhere
Assertions over the Corpus snapshot as pure data, before any interface existsSpecification, Seam 1
The assertion that no safety read touches the networkSpecification, Seam 4
The companion case: with the Corpus removed, every safety surface reports Corpus Unavailable and none reports Not StatedADR-0032, Consequences
The gaps this lifecycle openly carries — no automatic updates, no coverage guarantee outside the Study AreaSpecification, Declared Gaps

Where the rest of this is