nWave-ai/nWave / nWave/skills/nw-at-completeness-check
nw-at-completeness-check
VerifiedVerify that one minimal oracle falsifies every declared delivery obligation without checklist ceremony or duplicate tests.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 9.0 KB |
SKILL.md
View on GitHub---
name: nw-at-completeness-check
description: Verify that one minimal oracle falsifies every declared delivery obligation without checklist ceremony or duplicate tests.
user-invocable: false
disable-model-invocation: true
---
# Acceptance completeness: obligation-to-observation closure
**Kind**: KNOWLEDGE. Apply to acceptance completeness during DISTILL; the
caller orchestrates, ATD authors the oracle, and its existing reviewer judges.
When the crafter reviewer applies it at AT-density review, its closure checks are
criteria only: that role's typed outcome governs, a declared behavior absent from
the oracle is a rejection owned by `oracle`, and INDETERMINATE stays reserved for
absent evidence.
Review one candidate oracle against runner-derived facts from durable authority.
Completeness is a total
relation, not a score:
```text
every declared obligation -> one or more falsifiable observations
every oracle observation -> exactly one declared obligation or boundary law
```
PASS only when both directions hold. Never invent coverage from a fixed
checklist, scenario count, percentage, test pyramid or framework convention.
Each obligation of the DISTILL manifest's `obligation_citations` cites the
`outcome-N`/`scenario-N` public promise ids it checks, and the citations
collectively cover every public promise of the contract.
## Before-code alignment
First compare the original agreed value and its authorized exclusions/amendments
with the selected criteria, retaining provenance. A human-authorized exclusion
is not a missing requirement; a producer-only scope reduction is a defect.
Preserve every promised outcome, including already satisfied behavior that
requires no new production. Record a concrete selection/authority finding
before ATD or oracle agreement; this is adequacy, not a new requirement.
When the original value, selected-contract identity, and prior adequate
alignment are unchanged with no new finding, reuse that alignment. Do not repeat
the existing ATD review. Human scope authority is unchanged.
Only for new, changed, or unresolved semantic input, supply the same selected
revision, original value to preserve, and inherited criteria to ATD. The
independent examiner fixes expectations once per feature, before the first
acceptance designer. Keep those expectations independent of oracle authorship.
Include the ATD output and that fixed set in the existing acceptance-designer review. Preserve
behavioral specificity; summarizing criteria away or weakening them to make the
oracle pass is not alignment.
| Participant | Evidence and responsibility |
|---|---|
| ATD | Derive executable tests/properties from the original value preserved by the selected contract and its declared public observations. |
| Examiner, preimplementation | Derive public expectations independently from the user's point of view before the first acceptance designer; source, tests, implementation choices and runtime proof are not inputs. |
| Existing ATD reviewer | First check value-to-criteria adequacy, then compare selected obligations, executable tests/properties and examiner expectations for coverage, feasibility and consistency. This is the existing oracle review, not an additional review role or approval ceremony. |
The host supplies this preimplementation reviewer packet explicitly: original
Request and authorized scope amendments, selected behavioral contract and
revision, architectural facts/native substrate needed by ATD, proposed oracle/
supports and the examiner's fixed expectations. A postimplementation candidate
review packet is not this preimplementation packet. The examiner works from
the public contract and observations; code, tests, native selectors and producer
classifications are not needed to judge the user's expected experience.
Check timing and public constructibility: required ATD helpers/fixtures must be
available before craft, while future production may be absent. A constructible
future capability is not impossible solely because it is unimplemented. Batch
known consumer questions/counterexamples for the same upstream owner; do not
create an architect-plus-ATD review phase for every layer.
Before-code approval judges executable coverage and constructibility, not a
completed implementation. `Tested` names a test locator; it does not assert
that the test has passed. Require semantic RED for the missing behavior, not
GREEN or completed live delivery before craft. Run promised live observations
on the implemented candidate before completion; do not drop their obligations.
Once this basis is sound, start craft without another readiness round.
Resolve each mismatch in that review:
- **Grounded expectation missing from tests**: ATD strengthens the executable
oracle for the selected criterion before craft.
- **Overrestrictive expectation**: refine it to the selected criterion; retain
every required behavior and permit equivalent valid observations.
- **Ambiguous contract or new requirement**: return the concrete question to
the owning upstream authority explicitly; neither side silently expands scope.
The caller supplies the crafter one coherent selected revision after these
dispositions. The acceptance designer supplies complete acceptance facts,
including each stimulus, expected observation and coverage disposition.
Alignment adds no parallel document, ledger or handoff. Missing
observation evidence calls for evidence collection, not a production edit merely
to create a new candidate. A genuine oracle or authority defect returns to its
owner before craft proceeds against a revised contract.
Final EXAMINE remains a separate independent user-view assessment using the
original value, the same selected contract, and actual public observations.
Earlier alignment neither pre-approves implementation nor limits counterexample
discovery. An observed loss of original value rejects even when tests and
selected criteria agree. Missing observation is `INDETERMINATE` for the evidence
capture owner, not a product defect. Distinguish that state from a suspected
oracle defect, an observed contract violation, and a new expectation; the
examiner reports the semantic gap, while the existing oracle owner/reviewer
establishes a test defect from test evidence.
## Required closure
For every applicable obligation, verify that the oracle observes:
- the promised outcome and every materially distinguishable result;
- declared state, composition and preservation laws, using PBT for broad
domains and examples only for genuinely finite or singular observations;
- declared failure and recovery modes across domain, application/port,
adapter/integration and infrastructure boundaries;
- the real boundary type or protocol when a lookalike could pass falsely;
- one assembled installed journey when the user consumes that surface;
- the exact semantic checkpoint, so a dependency, fixture, import or setup
failure cannot masquerade as intentional RED.
An iterative or empty case is required only when the declared law induces it.
An additional test is required only when it adds a distinct observation.
Collapse equivalent examples, parameterize finite variants and keep one
property per independent universal law.
## Environment closure
The oracle's runner and dependencies must be reproducible from the repository's
actual manifest/build/test stanzas and runtime, not an ambient interpreter.
Follow native discovery in `nw-test-design-mandates`, including Dune/opam where
present. No closed manifest list, Python bridge or nWave adapter is required.
Missing runtime/dependency or build/setup evidence is BROKEN, not semantic RED.
## Boundary and reuse closure
Drive the nearest honest port that preserves the promised observation. A
cheaper seam is valid only with an explicit preservation map to the real
surface. Reuse existing helpers and oracles when they already own the same
law; do not duplicate them. New seams require an observation through the real
entry point, and the oracle must not create an architectural dependency that
the permanent design forbids.
For every broad-input/state/failure law projected below the real port, require
an explicit preservation map to the same promised observation. Without it,
return `EVIDENCE_GAP`; never downgrade the law to example-only coverage.
## Verdict
- `APPROVE` before craft: both relation directions close, the oracle is
constructible, and the missing behavior has an observed semantic RED.
Future implementation GREEN and live delivery are completion obligations.
- `NEEDS_REVISION`: a declared obligation lacks an executable falsifier, an
observation lacks authority, or the required RED is ambiguous with BROKEN.
- `INDETERMINATE`: evidence needed for the current phase is unavailable.
Re-review only the changed obligations and open blocking findings with the same
reviewer; retain unaffected judgments and evidence. An advisory or a later-phase
observation does not veto craft. An actual contract or oracle defect still does.
Route specification ambiguity to its owning upstream authority. Correct a
delivery-scope oracle gap inside DISTILL. Reviewer findings are ephemeral;
never create a checklist, gap ledger, receipt or parallel progress artifact.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| finishing-a-development-branchobra/superpowersUse when implementation is complete, all tests pass, and you need to decide how to integrate the work | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| systematic-debuggingobra/superpowersUse when encountering any bug, test failure, or unexpected behavior, before proposing fixes | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| test-driven-developmentobra/superpowersUse when implementing any feature or bugfix, before writing implementation code | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| skill-creatoranthropics/skillsCreate new skills, modify and improve existing skills, and measure skill performance. Use when users want to create a skill from scratch, edit, or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description for better triggering accuracy. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |