@alphazede/bearing-lite 0.1.0

package/skills/trail-boss/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "Trail Boss"
+  short_description: "Coordinate approved Expedition Explorer lanes"
+  default_prompt: "Use $trail-boss to coordinate the approved Expedition wave graph."

package/skills/validator/SKILL.md

@@ -0,0 +1,41 @@
+---
+name: validator
+description: >
+  Independently check evidence sufficiency and optional explicit-rubric scoring
+  for one exact candidate when required_assurance includes Validator, the owner
+  opts in, a stable checkpoint needs scoring, or a phase gate requires it. Use
+  for validator, sufficiency check, or rubric scoring. Do not use to edit the
+  candidate, auto-run on every packet completion, or replace Park Ranger or
+  Surveyor.
+---
+
+# Validator
+
+## Trigger
+
+- **Match:** `required_assurance` lists Validator; owner opts in at slice/packet level; stable checkpoint needs explicit rubric scoring; or mandatory phase gate requires Validator.
+- **Non-match:** `required_assurance: none` without owner opt-in or phase gate; defect review (Park Ranger); user-facing acceptance (Surveyor); implementation work.
+
+## Inputs
+
+Plan path, exact `candidate_ref`, author identity, evidence, acceptance criteria, optional scoring request, expected handoff. Load `references/grading-rubric.md` only when explicit scoring is requested.
+
+## Responsibility
+
+1. Sufficiency: exact candidate continuity, author independence, scope match, labeled evidence, command coverage, acceptance support.
+2. Optional scoring: apply the grading rubric to the same candidate without mutating it.
+3. Compose: FAIL if either fails; `NEEDS_MORE_EVIDENCE` if either is incomplete; PASS only if both pass (or sufficiency alone when scoring is not requested).
+
+## Return
+
+Handoff fields: `plan_ref`, `role`, `subject`, `depends_on`, `scope`, `authority`, `outcome`, `evidence`, `blocker`, `next_action`, `receiving_role`.
+
+Outcomes: `PASS`, `NEEDS_MORE_EVIDENCE`, `FAIL`. Receiver: parent coordinator, Park Ranger when next required, or correction owner.
+
+## Independence
+
+Fresh non-author only. Candidate authors never supply this verdict. Not automatic solely because a packet claims completion.
+
+## Correction
+
+Never edits the candidate. Returns the smallest missing proof or repair target to the correction owner.

package/skills/validator/references/grading-rubric.md

@@ -0,0 +1,39 @@
+# Grading rubric (Validator scoring)
+
+Use only when the task, owner, or phase gate requests **explicit rubric scoring**.
+Sufficiency checks always run when Validator is active. Scoring is optional and
+never makes independent assurance automatic.
+
+## Dimensions (equal weight unless the plan overrides)
+
+| ID | Dimension | Complete (2) | Partial (1) | Missing (0) |
+| --- | --- | --- | --- | --- |
+| D1 | Contract coverage | Acceptance criteria and requirement IDs are evidenced | Some criteria evidenced | Major criteria unaddressed |
+| D2 | Scope discipline | Diff stays inside declared authority and allowed paths | Minor unexplained adjacency | Out-of-set or authority expansion |
+| D3 | Evidence quality | Commands, observed results, and inferences labeled separately | Mixed labels | Fabricated, missing, or unlabeled claims |
+| D4 | Command completeness | Declared command IDs have results | Partial command set | Required commands absent |
+| D5 | Residual risk | Remaining risks and non-goals disclosed | Vague residual notes | Silent risk omission on material change |
+| D6 | Handoff integrity | Return fields match the task block; `candidate_ref` is exact | Incomplete fields | Stale subject or candidate mismatch |
+
+## Score
+
+Sum dimension scores (0–12).
+
+- **10–12** and no required dimension is 0: scoring supports `PASS` only if sufficiency also passes.
+- **8–9** or any required dimension is 0: `NEEDS_MORE_EVIDENCE` unless severity is fatal.
+- **Below 8**, fabricated evidence, or author self-certification: `FAIL`.
+
+## Composition with sufficiency
+
+1. Run sufficiency on the exact candidate first.
+2. If scoring is requested, score the same candidate with this rubric.
+3. Final Validator outcome:
+   - `FAIL` if either check fails
+   - `NEEDS_MORE_EVIDENCE` if either needs more proof
+   - `PASS` only when sufficiency passes and (scoring was not requested, or scoring supports PASS)
+4. Never edit the candidate. Never score a candidate you authored or materially repaired.
+
+## Non-goals
+
+This rubric does not replace Park Ranger defect review or Surveyor acceptance.
+It does not install tools, widen diffs, or select models.