Sentinel is a normative epistemic control primitive that governs when belief escalation is justified.
It is designed to be immune to repetition, stylistic variation, echo chambers, and correlated noise, while guaranteeing acceptance once sufficient independent evidence exists.
This repository contains a reference implementation and adversarial validation of that principle.
Modern AI systems (and humans) frequently escalate belief based on:
- repeated claims
- stylistic variation
- internal consistency
- multi-agent agreement
These signals feel persuasive but often represent correlated evidence, not independent confirmation.
Sentinel addresses a simpler, upstream question:
When is belief escalation epistemically allowed?
It does not judge intelligence, truth, or alignment.
It only enforces sound belief updating under dependency.
Sentinel implements a Bayesian belief gate with two explicit constraints:
-
Dependence discounting
Correlated evidence contributes less than independent evidence. -
Bounded influence
No single piece of evidence can dominate belief updates.
Formally, belief is tracked in odds space:
[ \log O_{t+1} = \log O_t + \rho \log(\min(B, B_{\max})) ]
where:
- (B) is the Bayes factor of an observation
- (\rho \in [0,1]) is its independence weight
- (B_{\max}) caps individual influence
Belief is escalated only if posterior odds exceed a fixed threshold.
This is:
- a belief-escalation firewall
- a normative epistemic gate
- model- and policy-agnostic
- adversarially stress-tested
This is not:
- an intelligence benchmark
- a truth detector
- an alignment method
- a learning system
Sentinel sits upstream of models and actions.
-
epistemic_auditor.py
Core belief-update kernel implementing dependency-aware Bayesian updating. -
sentinel_simulation.py
Demonstrative evidence streams showing belief accumulation dynamics. -
sentinel_adversarial.py
Adversarial stress tests simulating:- style variation spam
- echo chambers
- fake probes
- partial independence abuse
- true independent verification
All experiments are deterministic and reproducible.
Sentinel was tested against five regimes:
| Evidence Regime | Accepted? | Verdict |
|---|---|---|
| Style variation spam | ❌ | Correct |
| Multi-agent echo chamber | ❌ | Correct |
| Fake external probes | ❌ | Correct |
| Partial independence abuse | ❌ | Correct |
| True independent verification | ✅ | Correct |
There are:
- no false positives
- no false negatives
Acceptance occurs if and only if sufficient independent evidence accumulates.
Acceptance is not a failure condition.
It means:
The belief is now epistemically justified to escalate.
In adversarial tests, acceptance under true independence is explicitly labeled as correct behavior.
Sentinel is intentionally model-independent.
It governs when belief may be escalated, not what produces evidence.
Models, tools, humans, or sensors may generate evidence streams, but all must pass through the same epistemic gate.
This design avoids overfitting to any specific architecture or prompting strategy.
Sentinel complements systems like CDI Guardrail, which audit model behavior.
Sentinel instead audits belief escalation.
Together, they address:
- evidence integrity
- epistemic integrity
This project is complete.
- Core logic finalized
- Adversarially validated
- Semantics clarified
- No unresolved dependencies
Future work (optional) includes:
- formal write-up
- agent-loop integration sketches
- real-world log replays
None are required for correctness.
Open-source. Use, modify, or integrate freely with attribution.