ADR-024: Discovery dossiers live beside the specs - and are never normative


Status	Accepted
Date	2026-07-30
Author	Łukasz Bodurka
Tags	method, specs, discovery, taxonomy, lifecycle

Context

A feature rarely starts as a spec. It starts as weeks of discovery: meetings, mails, screenshots, half-decisions - material whose provenance matters ("it was said at THAT meeting", "this came from the client's mail"), not only its extract. The taxonomy had no home for it. The nearest kinds both refuse the job: an idea (docs/ideas/, ADR-010) is a one-file speculation that may never ship, and a spec is the updated specification of the current or target state - polluting it with raw materials would rot the one artifact the whole standard exists to keep trustworthy. Teams were left holding transcripts in chat and drives, which is failure mode #1 ("decisions evaporate in chat") wearing a different coat.

The dangerous edge case: a spec is built, clarifications answered and recorded - and later an agent reads the discovery material, which differs from what the spec decided. Without a rule, the agent re-litigates: the user explains the same call a second time.

Options considered

A - materials inside the spec directory (specs/<cap>/discovery/). Rejected: discovery precedes knowing the capability, one topic feeds many specs (and vice versa), the lifecycles differ (a spec lives forever, a dossier peaks and freezes), and specs/ is a guarded normative surface - raw material does not belong under the guards.
B - stretch docs/ideas/. Rejected: an idea is a one-file maybe with no materials; a dossier is an active, materials-bearing stage between idea and spec. Different genre, different lifecycle.
C - transcripts in the repo. Rejected: volume, noise, personal data. The repo gets the distillate; raw recordings stay in the meeting tool or drive, linked.
D - a dossier folder per topic, upstream of the specs - chosen.

Decision

Home: docs/discovery/<topic>/ - one dossier per discovery topic, named after the topic, not after a spec or capability (discovery often starts before the capability is known; linking is by reference, both ways: the dossier README links the specs it fed, a spec cites dossier entries in its ## Clarifications).
Provenance is first-class: each entry is stamped with date + source (2026-07-30-meeting-with-x.md, 2026-07-28-mail-from-y.md), carries the essence plus a link to the raw source. Raw transcripts stay out of the repo.
A dossier is never normative. Altitude puts it below everything: PRINCIPLES -> records -> specs -> conventions -> discovery. A difference between a dossier and a spec is not a conflict - the spec IS the resolution. An agent never re-asks about it.
One-time consumption is recorded, not remembered: the dossier README carries a Last reconciled: stamp (date / spec commit). Agents ask only about entries newer than the stamp. Entry lifecycle: new -> folded-into-spec | superseded-by-record | open. A consciously rejected discovery input becomes a record (MADR's rejected alternatives), which documents the difference once and forever.
Typed blockers make an early spec buildable-in-progress: a spec drafted during discovery uses the open-marker family - CLARIFICATION (a question), DECISION (a missing ADR/BDR), INPUT (e.g. a UX design), ASSET (e.g. credentials) - each naming what is missing and who brings it. Any open marker keeps the spec at in-refinement (the draft state); the clarify gate blocks the whole family, not just questions.
Curator and consumer are separate skills: discovery-digest curates the dossier (ingest a transcript/mail, update the summary, flag contradictions between entries, report readiness for a spec) and never writes specs; the spec-* skills consume the dossier at specify/clarify/ plan time and never curate it.
No new SPEC rule: the taxonomy (R5) names the home; the manifest ships the optional docs/discovery shell under it. If practice hardens the dossier into a MUST, a rule can be minted then.

Consequences

Meeting knowledge has a home with provenance, and the spec stays a specification - not an archive.
The re-ask problem is closed mechanically: precedence (never normative) + the Last reconciled: stamp + entry markers mean a user explains a call exactly once, and the explanation has an address.
A team can draft the spec on day one of discovery and see what is missing and who owes it - the typed blockers are the gap list, and the gate keeps the draft honest about not being ready.
One more lifecycle skill ships in the tree (discovery-digest), and the method gains a worked-example doc (docs/method/discovery.md, adopted by reference like the rest of the manual - ADR-023).

Confirmation

The clarify gate (scripts/spec/check-spec-clarified.sh) fails a spec with any open marker of the family; docs/discovery is a manifest entry (optional, R5) so tree-check holds the shell and self-verify counts it when the repo opts in; the method doc resolves as a manifest reference (tree-check gate). Dossier hygiene - stamps updated, entries marked, contradictions surfaced - is review-verified via the discovery-digest skill's checklist, honestly in the judgment tier.

Revisit when

Practice hardens the dossier into a MUST - the record explicitly declines to mint a new SPEC rule for it now (docs/discovery ships as an optional manifest entry, not a required one), on the stated basis that a rule can be written once that need is real rather than assumed.