sica-fondt/core/docs/plans/M6-conductor.md

M6 — Conductor (supervisory AI)

1. Component

The economy organ's supervisor: a specialist-trained AI with authority to veto Marketplace actions and pause/investigate individual Traders. Receives suspicious-behavior reports from the SAE (M7) and acts on them. The Conductor is the stomach's own judgment — it does not consult the organism's Brain for trade-level decisions. It supervises; the deterministic law script (M1) constrains; together they form the multi-layered braking system.

2. Status / certainty

DESIGN-FIRST · ABSENT. Role C4 (supervision architecture is clear); implementation C1 (model selection, training, authority scope).

3. Language & location

TBD · src/economy/conductor/. The Conductor is an AI actor — likely a fine-tuned LLM with specialist training in market risk, trader behavior analysis, and anomaly response. The inference wrapper sits alongside the Marketplace.

4. Does / does-not

• Does: receive SAE (M7) anomaly reports on trader behavior; pause a flagged trader's actions to investigate; veto a Marketplace action if investigation reveals risk; resume a cleared trader; review Marketplace actions pre-execution when the law check passes (M1-L4: veto is checked after law, before execution); maintain an audit log of all veto/pause/resume decisions.
• Does-not: trade (Traders do); execute on-chain (Marketplace does); modify the law script (immutable — M1-L2); detect anomalies directly (SAE does — the Conductor responds to SAE reports, it doesn't watch raw data); consult the organism's Brain.

5. Interface contract

• veto_check(action: MarketAction, trader_id) -> { approved | vetoed(reason) } — called by Marketplace (M1) for every law-passing action before execution.
• receive_alert(alert: SAEAlert) -> { pause(trader_id) | dismiss | escalate }. SAEAlert { trader_id, alert_type, evidence, severity, timestamp }.
• investigate(trader_id) -> { clear(resume) | veto_pending_actions | restrict(new_limits) }.
• decision_log() -> [ConductorDecision] — full audit trail of all veto/pause/resume/dismiss.
• SAE/Brain message format compatibility: the Conductor's incoming alert format is identical in structure and signature to Brain messages — SAE reports to the Conductor in the same shape it would report to the Brain. This means the Conductor can be swapped for Brain oversight without protocol changes (though the stomach normally operates autonomously).

6. Dependencies & stubs

• M7 SAE — anomaly reports; stub: no alerts (all clear).
• M1 Marketplace — veto check integration; stub: always-approve.
• M5 Traders — pause/resume control; stub: print pause/resume.

7. Invariants / laws

• L1 (C5): the Conductor can veto, but cannot trade — it has no wallet, no Marketplace access as a trader. It supervises from outside the trading loop.
• L2 (C5): the law script is above the Conductor — the Conductor vetoes actions that pass the law check but seem strategically risky. It cannot override a law violation (those are rejected before reaching the Conductor — M1-L4).
• L3 (C4): pause is reversible — a paused trader can always be resumed after investigation. Pause is a breaker, not a sentence (echoes A2-L3).
• L4 (C4): every Conductor decision is logged — vetoes, pauses, resumes, dismissals. The audit log is append-only (echoes M1-L2 / S3).
• L5 (C4): SAE alert format and Brain message format are structurally identical — same fields, same signatures. The Conductor processes them the same way the Brain would.

8. Build steps

1. Define the Conductor's decision model (rule-based for v1? fine-tuned LLM for v2?).
2. Wire SAE alert intake (M7 → M6).
3. Wire Marketplace veto check (M1 → M6 → approve/veto).
4. Implement trader pause/investigate/resume flow.
5. Implement append-only decision log.

9. Tests

Veto: flagged action is vetoed; unflagged action approved. Pause: paused trader cannot submit actions. Resume: cleared trader resumes normal operation. No trading: Conductor cannot submit MarketAction. Law supremacy: Conductor cannot override a law violation (never reaches Conductor). Audit: every decision appears in the log. Alert format: SAE alert parses identically to Brain message structure.

10. Open items

• Conductor AI model selection and training data (what does "specialist training" look like?).
• Veto criteria beyond SAE alerts (does the Conductor have independent judgment, or only responds to SAE reports?).
• Escalation path — if the Conductor is uncertain, does it escalate to the organism's Brain? Or is the stomach fully autonomous? (Current design: fully autonomous.)
• Multiple Conductors for redundancy?