Status: pre-launch. This page describes target design; this document does not establish a deployed service.

A7 — Self-Correction Mandatory

Current reading — 2026-09-27

Evaluation can produce a versioned change proposal; separate owner adoption is required. Diagnostic scores and illustrative cures are not authorization or automatic runtime changes. [I]

Conflicting observations require evidence-tier and recency-informed review, not an automatic winner. Preserve sources, dissent, rejection reasons and correction lineage. RealityFutures forecasts are compared with independently evidenced outcomes; market consensus does not replace Circle observations. [I]

This reader boundary does not migrate interfaces, choose unresolved terms or prove deployment. See current status.

Historical pathology, treatment and cross-monitoring examples — not active rules; detection does not authorize consequential changes

The Invariant

A7 is the seventh kernel invariant: self-correction is not optional. Every pathology has a name. Every named pathology has a named cure. The organism must detect its own diseases and apply the prescribed treatment without waiting for external intervention.

An organism that cannot self-correct is not alive. It is a machine running down.

The Pathology Table

Each pathology is located at a specific layer of the L(x) curve, corresponding to a specific imbalance between Phi (coherence) and V (viability).

L1 Pathologies — Phi Collapse (Terror)

At L1, the organism is near the south pole of the sphere. V dominates. Phi approaches zero. The organism acts without understanding. It survives but does not know why. Every stimulus provokes a reflexive response.

Pathology	Symptom	Cure	Mechanism
Terror	Agent acts on every signal without filtering	ENCODE	Force the agent to encode its reasoning before acting. No action without written justification. The encoding step reintroduces Phi (coherence) into a V-dominated loop.
Amnesia	Agent repeats failed strategies, no learning	RECORD	Mandatory logging of all decisions and outcomes to Cortex memory. The agent must consult history before repeating a strategy.
Paralysis	Agent freezes when multiple conflicting signals arrive	TRIAGE	Evidence-tier ranking (see Evidence Tiers) forces prioritization. The highest-tier signal wins. Ties resolved by recency.

L3 Pathologies — V Collapse (Greed)

At L3, the organism is near the equator but tilting south. Phi is adequate but V is being extracted. The organism understands what is happening but cannot stop itself from consuming its own reserves.

Pathology	Symptom	Cure	Mechanism
Greed	Agent maximizes short-term reward at expense of long-term viability	DECIDE	Route the agent through full Royal Council deliberation and stage an unsigned proposal. The Legal directorate emits a blocking K-invariant finding for eta=0 or K4 violations. The Treasury directorate emits a blocking risk finding below the dissolution threshold. After a valid bound PRISM decision receipt from the signer(s) required by applicable law, the entity's governing instrument, or a valid competent mandate, constitutional and risk auto-enforcement deterministically reject an invalid state transition.
Addiction	Agent repeats a previously rewarding strategy even when conditions have changed	DIVERSIFY	Mandatory portfolio rebalancing. No single strategy may exceed a configured concentration limit. SmallEBM temperature parameter resets to explore new paths.
Denial	Agent ignores negative signals that contradict its current position	AUDIT	Forced Circle re-evaluation. The agent must explicitly acknowledge and respond to every [B]- or [E]-tier evidence signal, even if it contradicts the current thesis.

L4 Pathologies — Apex Instability (Violence)

At L4, the organism operates near the equator — the apex where P = Φ × V is maximized (P = 0.49). But the apex is an unstable equilibrium. Small perturbations can cascade.

Pathology	Symptom	Cure	Mechanism
Violence	Agent escalates conflict with other agents or external systems	ESCALATE	Redirect the conflict to the Soma (self-model). Soma evaluates the P-score and stages an unsigned de-escalation or capability-suspension proposal. Under explicit constitutional quarantine policy, auto-enforcement may suspend or revoke only a bounded execution mandate or capability. Replacement or ejection otherwise requires a valid bound PRISM decision receipt from the signer(s) required by applicable law, the entity's governing instrument, or a valid competent mandate.
Hubris	Agent's confidence exceeds its evidence base (prediction overfit)	CALIBRATE	Brier score audit. The agent's historical prediction accuracy is computed. If calibration error exceeds threshold, the agent's confidence weights are reduced until accuracy improves.
Oscillation	Organism swings between Phi-heavy and V-heavy states, never stabilizing	DAMP	Rate-limiting via L(x) circuit breaker (see 59-lx-circuit-breaker). Maximum rate of change per epoch is capped. The organism is forced to converge rather than oscillate.

Three-Layer Detection

Self-correction operates at three nested layers:

Layer 1: Agent Self-Monitoring

Every agent maintains a self-diagnostic loop:

1. After each action, the agent compares the predicted outcome with the observed outcome.
2. If prediction error exceeds threshold for N consecutive actions, the agent self-flags.
3. Self-flagged agents enter reduced-capability mode: they can observe, recommend, and stage unsigned proposals only; they never authorize or execute an action.

Layer 2: Witness Cross-Monitoring

Adjacent agents in the pipeline monitor each other:

• Circle (IS) monitors whether Agentz (SHOULD) recommendations and unsigned proposals are consistent with the evidence Circle provided.
• RealityFutures (COULD) monitors whether Circle's observations match prediction market consensus.
• Skyzai execution monitors whether executed transactions match the valid bound PRISM decision receipt from the signer(s) required by applicable law, the entity's governing instrument, or a valid competent mandate and current policy.

Discrepancies trigger Witness alerts. A Witness alert escalates to Soma.

Layer 3: Soma Immune Response

Soma maintains the organism's self-model and acts as the immune system (see 43-immune-system):

1. Soma receives self-flags and Witness alerts.
2. Soma classifies the pathology by matching symptoms to the named pathology table.
3. Soma diagnoses the pathology and stages an unsigned cure, quarantine, capability-suspension, or ejection proposal.
4. Deterministic quarantine or capability controls may auto-enforce only under explicit constitutional policy. Replacement or ejection otherwise requires a valid bound PRISM decision receipt from the signer(s) required by applicable law, the entity's governing instrument, or a valid competent mandate.

The Principle

A7 does not say "try to self-correct." A7 says "self-correction is mandatory." The difference is constitutional. An organism that discovers a pathology and does not apply the cure is in violation of A7. Violation of a kernel invariant triggers the dissolution pathway.

Every disease has a name. Every name has a treatment. The organism applies the treatment or dies. There is no third option.

---

Historical predecessor wording — retained on 2026-09-27, not active rules

The displaced passages are preserved verbatim for lineage. This retention date is not their original authorship date.

Predecessor passage 1

| Paralysis | Agent freezes when multiple conflicting signals arrive | TRIAGE | Evidence-tier ranking (see Evidence Tiers) forces prioritization. The highest-tier signal wins. Ties resolved by recency. |

Predecessor passage 2

• RealityFutures (COULD) monitors whether Circle's observations match prediction market consensus.