Synth Nova Manifest

CriticalityPolicy

7 min read

Criticality Policy

Scope: All pipeline questions and agent decisions within Synth Nova modules (M1 Niche Evaluation, M2 Team Implementation, M3 Deliberation Chamber, future modules). Defines three criticality levels that determine whether a question is resolved by an agent autonomously, by an auto-quorum Chamber session (no founder), or by a full-quorum Chamber session with founder participation.

Governance: Operates under Constitution. Primary Laws: 1 (Founder Interests First), 2 (Sufficient Quality at Minimum Cost), 5 (Human Veto on Irreversible Decisions), 6 (No Important Decision Without Trace), 8 (Tokens are Capital). Integrates with EscalationPolicy (triggers 2, 3, 5, 7) and DecisionRights.

Why three levels

A binary “agent decides / founder decides” model either wastes founder attention on trivial questions or lets agents autonomously handle decisions that deserve multi-model scrutiny. Three levels match resolution cost to decision stakes:

  • Level 1 — cheap, deterministic, low-impact → agent resolves alone
  • Level 2 — uncertain but reversible → Chamber runs without founder (auto-quorum)
  • Level 3 — high-stakes, irreversible, or strategic → full Chamber + founder participation

Criticality is evaluated by the agent encountering the question, at the point of decision. Ambiguity resolves upward (when in doubt, escalate to a higher level per Law 9).

Level 1 — Agent Auto-Resolve

No Chamber. No founder. Agent decides and logs per ObservabilityContract.

ALL of the following conditions must be true:

  • Agent confidence ≥ 0.8
  • No conflicting evidence (i.e., EscalationPolicy Trigger 3 has NOT fired)
  • Judge score ≥ 6/10 on the relevant output
  • Budget impact of the decision ≤ $5 — this is the impact of the decision being wrong (e.g. “if we pick the wrong competitor name, what does the downstream cost look like”), NOT the cost of the LLM calls made to produce the decision
  • Question has a factual or deterministic answer — TAM number from a cited source, legal fact, competitor name, price lookup, regulatory rule, definition

If any condition fails → escalate to Level 2 or Level 3 per the criteria below.

Trace: standard audit log entry per ObservabilityContract. No special artifact required.

Rationale: Level 1 covers the vast majority of pipeline operations. Invoking Chamber on factual lookups is a Law 8 violation (token capital waste) and a Law 2 violation (expensive instrument where cheap one suffices).

Level 2 — Auto-Quorum (Chamber without founder)

Chamber runs automatically. Founder is NOT consulted during the session. Founder can review post-factum via the mandatory trace artifact.

ANY of the following triggers Level 2 (when Level 3 criteria below are NOT met):

  • Agent confidence between 0.5 and 0.8
  • Two agents give different but not mutually exclusive conclusions (e.g. Scout and Researcher estimate different but overlapping TAM ranges)
  • Question involves estimation or forecast — market sizing, growth rate, risk scoring, demand projection — where reasonable models can legitimately disagree

Rules

  • Chamber runs automatically with the 3-provider panel per ADR-0014-m3-deliberation-chamber
  • Founder is NOT consulted during the session
  • If arbiter confidence ≥ 8/10: result auto-accepted, pipeline continues
  • If arbiter confidence < 8/10: auto-escalate to Level 3 (the 8/10 threshold itself is the trigger — this is not “no consensus,” it is a deterministic escalation rule)
  • Mandatory trace (Law 6): every auto-quorum session writes the full session artifact per MultiLLMDeliberationPolicy §Audit Trail — panelist responses, divergence analysis, arbiter reasoning, arbiter confidence, final synthesis. Founder can review any Level 2 session post-factum.
  • Budget: Level 2 sessions count against the same $100/month Chamber budget and 10/day session limit defined in MultiLLMDeliberationPolicy §Rate Limits.

Constitutional justification

Law 5 (Human Veto) applies to irreversible, expensive, public, or reputationally significant decisions. Level 2 decisions are — by definition — reversible: estimates, forecasts, and sizing values can be recomputed, refined, or overridden when better data arrives. No irreversible commitment is made at Level 2. Therefore Level 2 does not trigger Law 5.

Law 6 (Trace) is satisfied by the mandatory full-session artifact. Law 8 (Tokens are Capital) is satisfied by the per-session cap, daily limit, and monthly budget. Law 1 (Founder Interests First) is satisfied because founder retains post-factum review and can override any Level 2 result.

Level 3 — Full Quorum (Chamber + Founder as participant)

Chamber runs with the founder as an active participant. Session blocks on founder input.

ANY of the following triggers Level 3:

  • Agent confidence < 0.5 on any key pipeline stage
  • Conflicting evidence (EscalationPolicy Trigger 3) on a strategic question
  • Go/No-Go decision on a niche — ALWAYS Level 3, no exceptions, even if all agents agree unanimously with confidence 0.99. This is a direct consequence of Law 1 (Founder Interests First) and Law 5 (Human Veto): committing Synth Nova to pursue or reject a niche is an irreversible strategic act.
  • Budget impact of the decision > $10 (impact, not LLM call cost)
  • Judge gave FAIL on a key pipeline stage
  • Question touches reputation, legal exposure, or financial commitment
  • Arbiter from Level 2 returned confidence < 8/10 (auto-escalation per Level 2 rules)
  • Founder explicitly requested participation

Founder roles in Level 3

The founder operates in one of four distinct modes during a Level 3 session. The role is chosen by the founder at session time based on the question and available evidence.

  • Approver — the quorum has produced a clear answer and the founder agrees. Founder responds “Accept” or “Reject” with brief reasoning. Session closes.
  • Challenger — the data is unconvincing to the founder. Founder responds “Go re-check [specific aspect X].” Agents fetch new data on X, then the quorum re-runs with updated context. Up to two challenge rounds before Level 3 must resolve.
  • Contributor — founder has insider knowledge the panel lacks (industry contact, prior engagement, non-public signal). Founder adds structured input — stance, reasoning, confidence — in the same schema as LLM panelists. Arbiter treats founder input as a fourth panelist voice per MultiLLMDeliberationPolicy §Founder Participation Rules.
  • Verifier — founder has an intuition or hypothesis but wants it checked before it influences the decision. Founder states the claim; agents fact-check it via sources; the verified (or falsified) claim feeds back into the quorum as evidence.

Trace: full session artifact per MultiLLMDeliberationPolicy + founder_action field with one of: accepted / revised / rejected / challenged / contributed / verified.

Go/No-Go outcomes (always Level 3)

Every Go/No-Go decision on a niche is Level 3. There are exactly three valid outcomes:

  • GO — niche recommended for entry. Pipeline continues to deeper analysis (financial model refinement, team-fit assessment, execution plan).
  • NO-GO — niche rejected. Pipeline stops. Final report finalized with rejection reasoning and preserved in the audit trail.
  • NEED MORE DATA — insufficient information for a responsible decision. Agents receive a specific data-gathering assignment (what to obtain, from where, by when), then the quorum re-runs with the new data.

Example (production): the BADs Russia run produced a -302% ROI at $100K budget — a number so poor it invited an immediate NO-GO, but the underlying cost inputs were estimated rather than quoted. The correct Level 3 outcome was NEED MORE DATA — fetch real quotes from contract manufacturers, then re-run. This prevented both a false NO-GO (on bad cost data) and a false GO (on uncorrected estimates).

NEED MORE DATA is not a failure state. It is a valid decision that protects the founder from acting on insufficient evidence (Law 9).

Criticality Assessment Procedure

When an agent encounters a question, before acting:

  1. Evaluate Level 1 conditions (all five). If all pass → resolve autonomously.
  2. If Level 1 fails, evaluate Level 3 triggers. If any fire → invoke full-quorum Chamber (EscalationPolicy Trigger 7, Level 3 path).
  3. Otherwise → Level 2 auto-quorum (EscalationPolicy Trigger 7, Level 2 path).
  4. Log the chosen level and the specific criterion that determined it: audit.criticality_assessment {level, criterion, question_summary}.

When in doubt between levels, escalate upward (Law 9 — default to safety under uncertainty).

Cross-References

Graph View

Backlinks