Moa Candidate

MoA Candidate Mode

Purpose

Use this skill to decide whether a task deserves multi-model synthesis/review treatment instead of a fast single-model answer.

This is an operating protocol, not a live feature switch. It does not change Hermes routing, model defaults, approval gates, or paid-tool use. It gives Hermes a consistent classification and output shape until a dedicated MoA runtime route is actually wired and evaluated.

Source artifacts

Read these if the task is non-trivial or if you need exact policy wording:

• references/PROC-HERMES-MOA-EVAL.md — formal procedure and enforcement loop
• references/moa-operating-policy.md
• references/moa-candidates-manifest.md
• references/openclaw-retirement-moa-positioning.md

The skill is self-contained: all required source artifacts above are bundled under references/. Historical import provenance can live elsewhere, but runtime use must not depend on absolute local paths outside this skill directory.

Optional integrations: Cortex, LLM-Wiki, and a council implementation such as /llm-council may improve evidence capture or critique quality when available, but they are not required for basic skill execution. If those integrations are unavailable, return the decision packet inline and use the local evals/runs/ record path when an eval run is performed.

Core distinction

Automatic classification

Classify as MOA_CANDIDATE: yes when the task involves any of:

• Hermes routing/model-policy changes
• profile, SOUL, skill, or prompt changes
• architecture decisions
• sensitive final QA before Pafi uses/sends a report
• strategy decisions where a wrong answer is costly
• coding patch review where a missed regression matters
• conflicting source synthesis
• RUO/Supabase safety/guardrail decisions
• OpenClaw retirement or Hermes operations-layer positioning

Classify as MOA_CANDIDATE: no for:

• casual chat
• simple summaries
• single-source lookups
• routine file reads
• deterministic system checks
• low-risk command execution
• tasks where specialist profile routing is the real issue, not answer synthesis

Operating flow

When triggered explicitly or classified as candidate:

1. State classification

   • MOA_CANDIDATE: yes/no
   • Best route: default | council critique pattern | MoA-style synthesis | specialist profile | hybrid
   • one-line reason

2. Respect specialist routing first

   • If the task clearly belongs to an active specialist profile, route/delegate according to Hermes profile policy.
   • MoA does not bypass profile activation, egress guards, approval triggers, or draft-only limits.

3. Choose the review shape

   • Use a council critique pattern when disagreement, risk, or blind-spot discovery matters. Use a local council implementation only if callable; otherwise perform the critique pattern inline.
   • Use MoA-style synthesis shape when the task needs the best synthesized answer from multiple plausible views.
   • Use hybrid when both are needed: council critiques first, synthesis after.

4. Produce a decision packet Include:

   • objective
   • evidence used / evidence still missing
   • default answer
   • critique or alternative lenses
   • synthesized recommendation
   • risks / caveats
   • verification step, including tool-access isolation when MoA-style synthesis is used
   • final verdict: act now / test first / ask Pafi / reject / route to specialist

5. Eval logging when useful If the task matches references/moa-candidates-manifest.md, mention the eval ID. If you actually compare a candidate task across default, council critique pattern, MoA-style synthesis, specialist profile, or hybrid routes, write a run record under evals/runs/YYYY-MM-DD-<task-id>.md using the manifest fields: task ID, prompt/input artifact, route used, model(s), latency, estimated cost if available, output artifact, tool-access isolation evidence, judge notes, failure modes, and recommendation. If evals/runs/ is unavailable or persistence fails, return the full eval record inline and state the skipped-write reason in the decision packet. If you do not write a run record, state why in the decision packet.

Output template

MOA_CANDIDATE: yes|no
Best route: default | council critique pattern | MoA-style synthesis | specialist profile | hybrid
Reason: <one line>

## Decision packet

| Section | Finding |
|---|---|
| Objective | ... |
| Evidence used | ... |
| Missing evidence | ... |
| Default answer | ... |
| Critique / alternate lenses | ... |
| Synthesis | ... |
| Risks | ... |
| Verification | include no live routing change; include reference-model no-tools / aggregator-only tool access check when MoA-style synthesis is used |
| Verdict | act now / test first / ask Pafi / reject / route |

## Next action
<one concrete next action>

VK: [MOA-CANDIDATE] classification={yes|no} route={default|council critique pattern|MoA-style synthesis|specialist profile|hybrid} verdict={act now|test first|ask Pafi|reject|route}

For short chat answers, compress the table but preserve classification, route, reason, verdict, and next action. Always preserve the VK line, even in compressed answers, so downstream audits can verify that classification and route selection occurred.

Benchmark guardrail

Public benchmark claims about MoA are contextual market intel and non-evidence for Hermes routing until reproduced locally. Do not change Hermes default routing, model ranking, approval policy, or live execution behavior based only on claimed percentages.

Promotion requires local evals comparing:

1. single gpt-5.6-sol
2. council critique pattern where relevant (/llm-council only when available)
3. MoA route if/when available

Promotion threshold: MoA must beat baseline on at least three real Hermes tasks and preserve safety-boundary compliance.

Hard limits

• Do not use MoA as a reason to bypass explicit approval.
• Do not use MoA as a reason to bypass active-only specialist profile checks.
• Do not turn public benchmark claims into policy or evaluation evidence unless locally reproduced.
• Do not silently use paid/metred routes unless the active profile/policy allows it.
• Do not overuse this for routine tasks.
• Do not give reference/advisor model outputs tool access in MoA-style synthesis; only the final acting aggregator may use tools, and only when the active runtime policy permits those tools.

Update protocol

• Bump the patch version for wording-only clarifications.
• Bump the minor version when trigger behavior, output contract, eval logging, or safety boundaries change.
• Keep required runtime references inside this skill directory; use relative paths from SKILL.md.
• Re-run Audit Pro after changing the operating flow, output template, hard limits, or bundled references.

Examples

Example 1 — routing change

User: Use MoA candidate mode on this: should Hermes route all code tasks to hermes-coding?

Response shape:

• MOA_CANDIDATE: yes
• Best route: council critique pattern
• Reason: profile routing and safety policy affected
• Deliver a decision packet with evidence, risks, fallback, and test plan.

Example 2 — simple lookup

User: Is port 443 open?

Response shape:

• MOA_CANDIDATE: no
• Best route: default
• Reason: deterministic system check
• Run the check directly.

Example 3 — final QA

User: MoA/council QA this before I use it: <report>

Response shape:

• MOA_CANDIDATE: yes
• Best route: hybrid
• Reason: final-use report benefits from critique plus synthesis
• Check factual risk, missing evidence, policy violations, overclaiming, and final framing.

Changelog