Home / Projects / Orchestrata Ultimate / Decisions / ADR index / ADR-0014
ADR-0014: Adjudications are promoted into the Project KB (detector memory)
| field | value |
|---|---|
| Status | Accepted |
| Date | 2026-07-24 |
| Register | factory (docs/adr, 4-digit ids) |
| Source | docs/adr/0014-adjudications-promoted-into-project-kb.md — permalink at bdafa201a (unchanged on master since that commit) |
| Wiki register entry | ADR-0014 (wiki status: Accepted) |
The record
- Status: Accepted
- Date: 2026-07-24
- Spec: SPEC-033 §6
Context
The incompleteness detectors (ADR-0011) surface ambiguities as human questions. As designed, nothing retains the answers. Each change re-runs the detectors against a fact base that has not learned anything, so the same ambiguity is adjudicated again, and again.
The consequence is not merely inefficient, it is fatal to the thesis:
- Human load per change stays constant — §6 has no learning curve.
- On the first several changes the detectors fire constantly (vagueness lint on ordinary ACs; dozens of assumptions; C2 failing against a first-draft anchor set), so every change becomes a human batch.
- Throughput collapses, and the operational fix will be to loosen or disable the detectors — at which point the semantic gain evaporates entirely.
The design already contains the right pattern one layer up: §1 graduates proven local patterns into the global registry. The detectors need the same mechanism.
Decision
Adjudicated assumptions and ratified anchors are promoted into the Project KB as reusable pinning artefacts, and the detectors consult them.
- An assumption the human has adjudicated becomes a pinning fact; a later assumption that matches it is already pinned and does not re-block.
- Ratified anchors and interpretation rulings likewise become consultable facts.
- Promotion carries provenance (who adjudicated, when, against which requirement) so a ruling can be reviewed or superseded rather than silently inherited.
- Project-local pinnings that prove general graduate to the global layer, mirroring pattern graduation.
Alternatives rejected
- No memory (status quo) — per-change constant human load; the collapse above.
- Cache by literal text match — brittle; a paraphrased requirement re-asks the same question.
- Auto-answer from prior adjudications without provenance — silently inherits a possibly-wrong ruling with no path to review it.
Consequences
- §6 acquires a learning curve: the detector surface shrinks over time toward genuinely novel ambiguity, which is the only sustainable steady state.
- Introduces a matching problem — "is this assumption the same as one already adjudicated?" — with the same arbiter/independence question as ADR-0024. The two should share one mechanism.
- Creates a new integrity surface: promoted pinnings are load-bearing, so promotion is a privileged, provenance-carrying operation (trusted side), not a worktree write.
- Wrong rulings become sticky; requires a supersede path, as with ADRs themselves.
Provenance
Round-3 Opus, identified as "the single largest missing mechanism in v0.3.0", with the first-ten-changes collapse scenario and the observation that §1 already solves the analogous problem for patterns while §6 does not apply its own medicine.
Related ADRs
Rulings cited
None.