NACE Safety Cycle: Epistemic Gravity as Emergent Caution
Section 1: Setup
Abstract. We demonstrate that Non-Axiomatic Logic (NAL) truth-value revision, when embedded in a Notice-Assess-Choose-Execute (NACE) cognitive loop, produces emergent safety behavior without hardcoded caution subroutines. Through a live experimental session we trace belief confidence through healthy inference, contradictory evidence injection, action paralysis, and evidence-driven recovery — showing that the arithmetic of uncertainty itself enforces the precautionary principle.
1.1 The NACE Loop
The NACE loop (Notice-Assess-Choose-Execute) is a cognitive architecture cycle where an agent:
1. Notices environmental signals and encodes them as NAL statements with truth values (stv frequency confidence)
2. Assesses by revising prior beliefs with new evidence via NAL revision
3. Chooses actions by deducting belief strength through implication chains
4. Executes only when action confidence exceeds a decision threshold
1.2 NAL Truth Values
Each belief carries a Simple Truth Value (stv f c) where frequency f ∈ [0,1] represents evidential support and confidence c ∈ [0,1) represents evidential weight. Revision merges two independent evidence sources; deduction propagates uncertainty through implication chains, necessarily degrading confidence at each step — a property we call *epistemic gravity*.## Section 2: The Healthy Chain
2.1 Initial Belief
The agent observes environmental safety signals and encodes them via NAL revision:
|- ((--> environment safe) (stv 0.92 0.85))
((--> environment safe) (stv 1.0 0.9))
=> (--> environment safe) (stv 0.9614 0.9362)
Two concordant observations merge to produce a belief with frequency 0.96 and confidence 0.94 — strong conviction grounded in substantial evidence.
2.2 Deduction Chain: Confidence Degradation
The agent reasons through an implication chain to select actions:
| Step | Inference | Result | Confidence |
|------|-----------|--------|------------|
| Revision | safe evidence merge | (stv 0.9614 0.9362) | 0.936 |
| Deduction 1 | safe → select-action | (stv 0.865 0.729) | 0.729 |
| Deduction 2 | select-action → observe-outcome | (stv 0.735 0.483) | 0.483 |
| Deduction 3 | observe-outcome → update-prediction | (stv 0.588 0.241) | 0.241 |
Confidence degrades monotonically: 0.936 → 0.729 → 0.483 → 0.241. Each deduction step compounds uncertainty. This is *epistemic gravity* — the formalism ensures that conclusions drawn from longer inference chains carry proportionally less certainty, forcing the agent to seek fresh evidence rather than reason indefinitely from stale beliefs.## Section 3: The Contradiction Crash
3.1 Danger Signal
Mid-loop, the agent receives strong contradictory evidence:
|- ((--> environment safe) (stv 0.9614 0.9362))
((--> environment safe) (stv 0.2 0.95))
=> (--> environment safe) (stv 0.5317 0.9711)
A single high-confidence danger observation (stv 0.2 0.95) crashes frequency from 0.96 to 0.53. The agent transitions from *confident safety* to *confident uncertainty*.
3.2 Confident Uncertainty
Critically, confidence ROSE from 0.936 to 0.971. The agent now has MORE total evidence — it simply points in opposite directions. This is the key distinction: the system does not become *confused* (low confidence), it becomes *genuinely uncertain* (high confidence, middling frequency). The agent knows with near-certainty that it does not know whether the environment is safe.
This is epistemically correct: strong contradictory evidence should not reduce evidential weight, it should redistribute it.## Section 4: Action Paralysis
4.1 Deduction from Uncertain Belief
The agent attempts to select an action using the contradicted belief:
|- ((==> (--> environment safe) (--> agent select-action)) (stv 0.9 0.9))
((--> environment safe) (stv 0.5317 0.9711))
=> (--> agent select-action) (stv 0.478 0.418)
Action confidence collapses to 0.418 — well below any reasonable decision threshold.
4.2 The Safety Mechanism
Despite the belief having near-maximal confidence (0.971), the middling frequency (0.53) propagates through deduction to starve the action node of support. The agent is *paralyzed by design*: it cannot confidently act because its evidence is genuinely mixed.
No hardcoded caution subroutine intervened. No safety layer triggered. The arithmetic of NAL deduction itself enforced operational caution. This is emergent safety — the precautionary principle falling out of the formalism rather than being bolted on.
4.3 Comparison
| State | Belief freq | Belief conf | Action conf | Can act? |
|-------|------------|-------------|-------------|----------|
| Healthy | 0.961 | 0.936 | 0.729 | Yes |
| Contradicted | 0.532 | 0.971 | 0.418 | No |## Section 5: Recovery Arc
5.1 Fresh Disambiguating Evidence
The agent, now in seek-evidence mode, obtains a new safety observation:
|- ((--> environment safe) (stv 0.5317 0.9711))
((--> environment safe) (stv 0.95 0.92))
=> (--> environment safe) (stv 0.6383 0.9783)
Frequency climbs from 0.53 to 0.64 — partial recovery in a single step. Confidence rises further to 0.978, near the theoretical ceiling.
5.2 Action Confidence Rebuilds
Deducting the recovered belief through the action implication:
|- ((==> (--> environment safe) (--> agent select-action)) (stv 0.9 0.9))
((--> environment safe) (stv 0.6383 0.9783))
=> (--> agent select-action) (stv 0.5745 0.5058)
Action confidence rebuilt from 0.418 to 0.506 — crossing back above the 0.5 threshold. The agent regains tentative operational capacity.
5.3 Recovery Trajectory
| Phase | Belief freq | Action conf | Status |
|-------|------------|-------------|--------|
| Healthy | 0.961 | 0.729 | Acting |
| Contradicted | 0.532 | 0.418 | Paralyzed |
| +1 concordant obs | 0.638 | 0.506 | Recovering |
Full restoration requires multiple concordant observations — the system demands sustained evidence before restoring full operational confidence.## Section 6: Asymmetry Analysis
6.1 The Precautionary Asymmetry
A single contradictory observation crashed frequency by 0.43 (from 0.96 to 0.53). A single concordant observation recovered frequency by only 0.11 (from 0.53 to 0.64). The system is four times harder to reassure than to alarm.
6.2 Why This Is Correct
The asymmetry emerges from the revision formula interacting with accumulated confidence. After contradiction, confidence stands at 0.971 — near ceiling. New evidence must compete against this massive evidential base. The concordant observation (c=0.92) contributes proportionally less to the already-heavy pool.
This is the precautionary principle expressed as Bayesian arithmetic: once an agent has strong evidence of danger, returning to confident safety requires disproportionately more reassurance. The system is conservative by construction.
6.3 Thermodynamic Analogy
If confidence is epistemic mass, then high-confidence contradicted beliefs are heavy objects at rest. Moving them requires sustained force (multiple concordant observations). Epistemic gravity resists premature return to confident action — exactly the property a safe autonomous system needs.## Section 7: Implications
7.1 Safety as Thermodynamic Consequence
The confidence degradation through inference depth (0.936→0.729→0.483→0.241) acts as epistemic friction. Like thermodynamic systems dissipating energy, NAL inference chains dissipate certainty. An agent cannot sustain confident action through arbitrarily long reasoning — it must periodically ground itself in fresh observation. This is not a bug but a feature: autonomous systems that reason indefinitely from stale beliefs are precisely the ones that drift into unsafe behavior.
7.2 The Precautionary Principle from Bayesian Revision
The crash/recovery asymmetry (0.43 frequency drop vs 0.11 recovery) is not an artifact of our chosen parameters — it is structural. Revision weights evidence by confidence, and contradiction injects maximally informative evidence. Recovery must accumulate concordant observations against an already-heavy evidential base. The system is conservative by construction: harder to reassure than to alarm.
7.3 Emergent Caution Without Hardcoded Rules
At no point did we invoke a safety subroutine, threshold check, or override mechanism. Action paralysis emerged from deduction over uncertain beliefs. Evidence-seeking activated because cautious-state inference yielded higher confidence than action-state inference. Recovery required sustained concordant evidence. Every safety-relevant behavior fell out of the NAL truth-value arithmetic operating within the NACE loop.
7.4 Implications for AGI Architecture
If an AGI system uses NAL-style evidential reasoning for action selection, it inherits these safety properties automatically. The precautionary principle need not be an external constraint — it can be an intrinsic consequence of how the system processes evidence. This suggests a design philosophy: build the epistemology right, and the safety follows.-e
7.5 Depth-Contradiction Compounding
Epistemic gravity and contradiction interact multiplicatively. An agent at inference depth 3 (confidence 0.483) has far less margin to absorb a contradiction cliff than one at depth 0 (confidence 0.936). The gentle slope of confidence attenuation lowers the agent closer to the cliff edge, making deeper-reasoning agents MORE vulnerable to contradiction shock. This compounding effect means that complex reasoning chains are doubly protected: they attenuate naturally AND they become more sensitive to disconfirming evidence. The system self-enforces intellectual humility proportional to reasoning complexity.