---

type: adr id: 0030-unified-artifact-set status: accepted created: 2026-06-02 updated: 2026-06-02 supersedes: superseded_by:

ADR-0030: The unified artifact set

Context

The pre-kernel model defined four core document types — research, spec, audit, bug-report — with extended types orbiting them (ADR 0001). That set named the upstream sources of truth well, but it had no place for the artifacts the rest of the pipeline depends on: there was no first-class trace carrying implementation claims against obligation IDs, no finding as the unit of durable evidence, no memory recall layer, and no home for a verdict (the judged outcome of a review). Without these, completion claims, drift, and durable knowledge had nowhere structured to live, and a verdict risked being modelled as a standalone file. The recast in §29 records the decision to close these gaps while keeping every source document's epistemic stance intact.

Decision

The kernel defines one unified artifact set centred on the obligation graph (§20, §29). It comprises the seven Tier-1 pipeline-core artifacts — spec.md, task.md, trace.md, review.md, finding.md, adr.md, memory/INDEX.md — plus the Tier-3 stdlib source-doc templates (audit.md, research.md, bug-report.md, and the conditional prd.md/rfc.md). It adds the four artifacts the four-document model lacked: the trace, the VERDICT (a SOL block that lives inside review.md — there is no verdict.md, §20.2.3), the finding, and memory (§29.3). Each source document preserves its epistemic stance on promotion: observation, diagnosis, inquiry, intent, and proposal remain non-authoritative until an author pass turns them into spec obligations, exactly as the §29.1 stance table fixes. The full set, contracts, and templates are specified in §20–§21; the recast that maps the earlier four onto it is §29.

A conformant repository MUST ship the contract and a copyable template for each Tier-1 artifact, MUST ship the Tier-3 source-doc templates, and MUST NOT introduce per-artifact .* names for the working artifacts (audit/research/bug-report/finding/adr are plain .md); the obligation source is the single human-authored compiler-visible artifact spec.md (§29.2). Extended types (constitution, migration-plan, benchmark/cleanup reports) remain specializations that reuse a parent artifact's template, frontmatter, and stance, and MUST NOT acquire their own block types, lint codes, or IR node kinds (§29.4). Concretely, the kernel ships no standalone constitution template: a project constitution is authored as a spec.md (or a rank-1 adr.md) specialization, and its supreme-guideline authority comes from the source-authority rank — not from a separate document type. This is intentional, not a gap.

Alternatives considered

Alternative	Why rejected
Keep only the earlier four document types	They name upstream sources but leave trace, finding, memory, and the verdict with no structured home; completion claims and durable knowledge have nowhere to live (§29, §29.3).
Model the verdict as a standalone verdict.md file	A verdict is the output of a review pass, so it is a SOL block inside the review.md container, never a free-standing artifact (§20.2.3, §29.3).
Give each extended type its own block types / lint codes / IR kinds	Extended types differ only in conventional content; duplicating the language machinery would fork the obligation model for no semantic gain (§29.4).
Give audit/research/bug-report/finding their own .* names	The spec.md naming marks the one human-authored compiler-visible source; working artifacts are plain .md and parse under artifact contracts, not the SOL grammar (§29.2, §20.1).
Let an audit author its own obligation blocks (collapse stances)	Smuggling intent into an observation-only artifact is a distillation error; observed risk acquires obligation force only by promoting into a spec (§29.1, §29.5).

Consequences

Positive

• Implementation claims, judged verdicts, durable facts, and recall each have a single, contract-defined home, so provenance and drift can be tracked end to end.
• Every artifact's epistemic stance is preserved on promotion, so a lower-stance source cannot masquerade as approved intent (§29.1).
• One verdict location (review.md) and one obligation source (spec.md) remove the ambiguity of a free-standing verdict file or competing source names.

Negative

• The artifact catalogue is larger than the earlier four; adopters must learn the trace/finding/memory roles and the stance discipline.
• The "no verdict.md" rule is a constraint a future conformant tool MUST honour by convention; nothing in the markdown layer prevents someone authoring a stray file.

Neutral / tradeoffs

• Forbidden compositions are enforced by the distillation-loss budget (§24) and source authority (§22), not by a gatekeeper skill (§29.5) — a deliberate relocation of the same prohibition into the language/reference layer.
• Extended types stay available as named variants but inherit their parent's machinery, trading bespoke modelling for a smaller, stable type system (§29.4).

Status

Accepted (v0.1).

Affected obligations / constraints

• Adds: the trace, the VERDICT block (inside review.md), the finding, and the memory recall layer as kernel artifacts (§29.3).
• Adds: the conformance requirement to ship a contract + template for each Tier-1 artifact and the Tier-3 source-doc templates (§20.3, §20.4).
• Modifies: the four-document model of ADR 0001 — recast onto the unified artifact set while keeping each source document's epistemic stance (§29; ADR 0001 kept verbatim).
• Supersedes: none.

Ledger note (2026-06-11): refined by ADR-0059, ADR-0060, ADR-0061, ADR-0068 (artifact set re-tiered).