oracle

Files (1)

• SKILL.md

SKILL.md (first lines)

# Cladding oracle — impl-blind conformance authoring

A SPEC_CONFORMANCE oracle is a conformance test authored **without seeing the implementation**, so a passing
oracle means "matches the spec," not "matches the code." The A/B that motivated this: a blind oracle caught
bugs a code-peeking (sighted) oracle rubber-stamped (7/8 vs 4/8). cladding owns no LLM — the blinding is **your
discipline as the host**; cladding produces the brief, records provenance, and the gate audits it.

## Which ACs need an oracle? (the policy)

A project sets its requirement under `spec.yaml::project`:
- **`oracle_policy: { always_ears: [unwanted], sample: 0.2 }`** (RECOMMENDED) — risk-weighted: author an oracle
  for every done AC whose EARS category is in `always_ears` (default `['unwanted']` — error/edge handling),
  PLUS a deterministic ~`sample` fraction of the rest. v8 showed exhaustive per-AC oracles add ~0 quality at
  ~30% cost, so spot-check the bulk and concentrate verification where failures cluster.
- **`require_oracles: true`** — EXHAUSTIVE (every done AC). Highest assurance, highest cost. `oracle_policy`
  takes precedence when both are set.
- **Neither** — no mandate (an authored oracle still runs + is recorded; a missing one is not forced).

Run **`clad oracle --required`** to print the worklist — exactly which done ACs the policy demands an oracle
for, which already have one, and why (`always:<ears>` / `sample` / `exhaustive`). Author oracles for the
`← needs an impl-blind oracle` rows only; do NOT author for ACs the policy did not select.

## Protocol (three steps — do them in order)

1. **Get the spec-only brief.** Run `clad oracle <featureId> --ac <acId>`. It prints the acceptance criterion
   + the module's declaration-only signatures — and NEVER an implementation body. This is the *only* thing the
   author may see.

2. **Spawn a FRESH, blind sub-agent** (the Task tool / a new sub-agent context) handed ONLY that brief. It MUST
   NOT read `src/` or any implementation file. Instruct it to write a vitest conformance suite that asserts
   **only what the criterion literally requires** — when the spec is silent on an edge, a WEAKER assertion, not
   a stronger guess (an over-strict oracle falsely fails correct code). The sub-agent's identity must differ
   from whoever implemented the feature.

3. **Record it.** Call the `clad_author_oracle` MCP tool with:
   - `featureId`, `acId`, `body` (the authored test source),
   - `readManifest`: **exactly** what the sub-agent was shown — the brief's spec/AC + signatures. It MUST NOT
     list an implementation file the feature owns (the gate fails on `manifest ∩ modules`).
   - `blind: true` only if the sub-agent's context was the brief and nothing else,
   - `authorName`: the sub-agent's identity (≠ the implementer).

   cladding writes `tests/oracle/<F>.<AC>.test.ts`, records `kind:'oracle'` provenance, and stamps
   `oracle_refs` onto the AC. The SPEC_CONFORMANCE gate (stage_2.3 + the detector) then RUNS the oracle against
   the real code and AUDITS author≠implementer + manifest∩modules=∅.

```
clad oracle F-1a2b3c --ac AC-004      # 1. print the blind brief
# 2. spawn a blind sub-agent with ONLY that brief → it writes the oracle
# 3. clad_author_oracle { featureId, acId, body, readManifest, blind:true, authorName }
```

## Honest boundaries (read these)

- **Blindness is enforced by YOU, not cladding.** cladding cannot see or restrict a sub-agent's file reads
  (sub-agent tool perms belong to the host). Hand the sub-agent ONLY the brief; do not let it open `src/`. The
  gate audits the manifest you report — it catches an honestly-reported impl read, not a lie. `blind:false`
  records an unattested (self-reported) manifest and the gate surfaces it as `info`.
- **First RED is ambiguous.** When a brand-new oracle fails on the current code, it is EITHER a real spec bug in
  the code (keep the oracle, fix the code) OR an over-strict oracle (the spec doesn't require it — revise/reject
  it). cladding cannot tell which without you. Decide deliberately; never auto-accept or auto-discard.
- **Opt-in + risk-weighted.** The presence + provenance rules bind under `project.oracle_policy` (risk-weighted,
  recommended) or the legacy `project.require_oracles: true` (exhaustive). Without either, an authored oracle
  still runs (stage_2.3) and its provenance is still recorded, but a missing oracle is not forced. Prefer
  `oracle_policy` — exhaustive verification bought ~0 quality at ~30% cost in v8; concentrate the premium on the
  high-risk (`unwanted`) ACs + a sample.

From purpleeddy/cladding, MIT.

More from purpleeddy/cladding