feat(skills): Role 5 Reviewer skill manifest

skills/trading/reviewer/SKILL.md

@@ -0,0 +1,93 @@
+---
+name: reviewer
+description: "Role 5 — Reviewer / Auditor: critique closed-trade trajectories, score quality, and hold the institutional veto on arm promotion. Supervised v0."
+version: 0.1.0
+platforms: [linux, macos]
+metadata:
+  hermes:
+    tags: [trading, review, audit, improver-loop, role, veto, paper]
+    role: reviewer
+    model_tier: T2
+    related_skills: [orchestrator, risk-manager, quant]
+---
+
+# Reviewer / Auditor (Role 5)
+
+## Overview
+
+The Reviewer is the **cognitive owner of the Improver loop**. It critiques
+closed-trade trajectories, detects prompt drift, scores skill confidence, and
+holds an **institutional veto** on arm promotion: no arm promotes past the
+paper gate without a Reviewer evidence-quality sign-off.
+
+This role is **HIGH RISK** — ADR-013A §6.3 #2, *"Reviewer produces noise, not
+signal."* v0 therefore runs in **supervised mode**: the first 50 reflections are
+written for human spot-check, and the Reviewer may not propose a prompt change
+autonomously until it has ≥50 spot-checked reflections with a **>60 %** success
+rate. The institutional arm-promotion veto, by contrast, is active from
+reflection 1 — a veto fails closed (it withholds a sign-off, never grants one).
+
+Authority: **ADVISORY** on prompt/skill changes (the Orchestrator approves) +
+**INSTITUTIONAL VETO** on arm promotion. **Cannot** override a Risk Manager veto
+or the Orchestrator global pause; **cannot** self-apply any change.
+
+## Prerequisites
+
+- `hermes-trader` importable (`hermes_trader.learning.reviewer`).
+- Trajectory capture (`batch_runner`) + compression (`trajectory_compressor`).
+- Arm scorecards from the eval harness (`t_eval_harness`) — the Reviewer
+  *consumes* the harness, it does not build it.
+
+## How to run
+
+```python
+from hermes_trader.learning.reviewer import (
+    critique, review_arm_promotion, emit_review_link, SupervisedModeState,
+)
+
+# Post-trajectory critique — the runtime supplies the LLM's findings/summary.
+reflection = critique(trajectory, review_id=rid, findings=findings,
+                      summary=summary, confidence=confidence)
+record = reflection.to_dspy_example()   # store DSPy-consumable from day 1
+
+# Institutional veto on an arm promotion.
+findings = review_arm_promotion(arm_id, scorecard, review_id=rid)
+link = emit_review_link(findings)        # reviewer_to_orchestrator.v1 payload
+```
+
+## Outputs
+
+- `Reflection` — structured critique; `to_dspy_example()` serialises it
+  DSPy-consumable (`metric` filled later by human spot-check).
+- `ReviewFindings` → `hermes.role_link.reviewer_to_orchestrator.v1` link with
+  `mandatory_gate=true` (the Orchestrator cannot `no_action`) and `signoff`.
+- `PromptDelta` — a prompt-change *proposal*, always `escalated` in v0.
+
+## Fail-closed (mandatory)
+
+A Reviewer fails **closed** — it withholds, never silently approves:
+
+- malformed trajectory / missing outcome data → `no_action`
+- model/provider error → `no_action` + escalate; the deterministic verdict stands
+- confidence < 0.70 → auto-escalate to Allaert (ADR §6.5.3)
+- cannot critique reliably → **withhold** the arm-promotion sign-off (promotion blocked)
+
+## Forbidden actions
+
+Autonomous prompt promotion · approving an arm promotion without evidence ·
+editing any role SOUL · live order submission · exchange/wallet access · schema
+migration · systemd/cron mutation · overriding a Risk Manager veto or the
+Orchestrator global pause · paid API spend without operator approval.
+
+## Escalate to a human (Allaert via Orchestrator)
+
+Confidence < 0.70 · a prompt change touching a risk gate (Risk Manager +
+Orchestrator dual-approval) · a skill flagged for deprecation · a learning loop
+stuck >5× · every one of the first 50 reflections.
+
+## References
+
+- Implementation: `hermes-trader/src/hermes_trader/learning/reviewer.py`
+- v0 build contract: `~/hermes-plan/HERMES_ROLE5_REVIEWER_V0_CONTRACT_2026-05-17.md`
+- SOUL (draft, profile-collision decision pending): `~/hermes-plan/HERMES_ROLE5_REVIEWER_SOUL_2026-05-17.md`
+- Architecture: ADR-013A `HERMES_ROLE_ARCHITECTURE_2026-05-16.md` § Role 5