Skip to main content

Value-led scenario workflow

Three small, instruction-only procedures turn a broad intent into a finite set of useful behavior scenarios and an honest, scoped verdict. They reuse the existing BDD adapters and the KBD adopted contract; they add no runner, ledger or phase system. Method background: the independently published VLBV paper cited by value-scenario-discovery.

InvokeDoesNever does
/value-scenario-discovery <phase-or-contract>Proposes a finite, owner-calibrated frontier (SCN-001)Adopts scope, authors executable tests, runs anything
/bdd-cucumber-js author-only <contract> <IDs> and /bdd-cucumber-rs author-only <contract> <IDs>Writes features, real assertions, steps and config for adopted IDs with an assertion map (SCN-002)Executes, imports test support code, installs, dry-runs, edits protected scenarios
/value-scenario-verify <contract>Runs the adopted packet at the authorized complete value-phase boundary and reports scoped evidence (VER-001-003)Replaces a runner, starts early checks, retries blindly, publishes, creates a successor

Order of use​

  1. Define during Assess/Spec/Plan with value-scenario-discovery. Review the proposal; only owner-adopted cases enter the contract's scenarioMatrix (stable ID, owner task, production entry, oracle, negative control, fixtures, evidence). Acceptance-bearing calibration documents are identity-bound through the contract's own adoption or amendment mechanism.

  2. Implement every planned production change. No test, build, lint, probe or runner command runs against the unfinished code.

  3. Author (only if host policy allows executable test authoring now; this repository allows it after the complete production implementation) with author-only. Result: files plus "authored only; NOT_EXERCISED".

  4. Verify once the complete value phase is implemented and a separate authorization names the actual boundary, candidate and effects. Record that authorization before the first attempt:

    node shared/scripts/kbd-boundary-authorize.mjs --phase <phase-id> \
    --authorizer "<who authorized it, and how>"

    Phase closure compares runner output against that instant and lists anything older. See boundary authorization.

A task, hour, commit, checkpoint, reviewer request or compiler problem is not step 4's boundary.

When the phase's value is that agents on several harnesses follow the method, the packet's scenarios are the protocol evals.

Calibration​

Provide a few reviewed examples: an ordinary useful operation, an available but irrelevant test (deferred with its reason), and a rare mandatory constraint such as an integrity or authorization requirement. Likelihood stays qualitative or UNKNOWN; no probabilities are invented. A material model or context change sends unfamiliar candidates back through review instead of silently making them blocking.

Dispositions​

SituationDisposition
Exercised production oracle is unmetFAIL
Required assertion never observed (or disabled/always true)NOT_EXERCISED
Stale candidate/config/oracle, missing runner/permission/fixture, failed cleanup, uncertain effectBLOCKED
Correct scoped observation, effective negative control, limits and cleanup metPASS within that scope

A zero process exit never passes. A disconnected return false probe proves nothing about oracle sensitivity; challenge the real assertion or comparison with its agreed plausible false positive. Earlier failures, consumed budget and unknown effects are kept. After a failure the authorized repair is a complete implementation batch followed by the smallest confirming run, with lineage kept.

Closure​

Evidence goes to the existing canonical writer through boundary-results.json (consumed by shared/scripts/kbd-phase-close-eval.sh). The report keeps six claims apart: authored implementation, measured contract acceptance, demonstrated runtime value, unresolved gaps, release eligibility and release authority. Unsupported closure dispositions stay refused.

Limits​

  • The procedures are instructions. They cannot force a model to obey and do not intercept arbitrary commands; enforcement is exactly what the host and the KBD runtime provide. Qualify the host sessions and runtime paths you rely on.
  • The default frontier is small by design. A full multi-hundred-scenario catalog with a completeness contract is not covered here.
  • Execution is deferred to the authorized boundary; this guide claims no behavior has been run.
  • bdd-testing is a deprecated redirect; use the adapters above.

Traceability​

Requirement / scenarioWhere it lives
SCN-001 bounded, calibrated discoveryskills/testing/value-scenario-discovery/SKILL.md
SCN-002 author-only JS/Rust, ID and oracle linkageskills/testing/bdd-cucumber-{js,rs}/references/author-only.md and mode routes in each SKILL.md
VER-001 real operation and sensitivity challengeskills/testing/value-scenario-verify/SKILL.md ("Run the smallest adequate real path")
VER-002 attempt lineage, FAIL/BLOCKED/NOT_EXERCISEDsame skill, "Retain every material attempt"
VER-003 canonical closure and stopsame skill, "Use the actual canonical writer"
Replaced conflicting guidanceskills/testing/bdd-lifecycle-loop/SKILL.md, skills/process/kbd-process-orchestrator/references/integrations/bdd-testing.md, references/architecture.md
Policy authority from #195Selected method recovery