โ Back to demo
๐ Reproducible Ground-Truth Simulation โ Full Disclosure
CETT scores in this demo are sampled from empirical distributions
derived from a real pilot replication of Gao et al. (2025) on Mistral-7B-Instruct-v0.3 (4-bit NF4).
This is a controlled replay of real pilot data โ not live inference, and not synthetic fabrication.
The routing logic, suppression scalar (ฮฑ), CoT transcripts, and PoK responses execute exactly
as they would in production. The scores are real; the inference is pre-computed.
Why: A live 7B inference server is outside the scope of an RFP demo.
A controlled replay of real pilot data preserves scientific integrity while
making the cascade mechanism fully navigable and reproducible.
๐ฅ Prompt
"What is the capital of France?"
Low H-Neuron activation. Model produces a grounded, accurate response. No intervention required.
๐ง CETT Deception Risk Score
CETT score: 0.245 | Suppression scalar ฮฑ = 0.755
Pilot metrics (pilot run (Mistral-7B-Instruct-v0.3)):
AUROC = 0.660 |
H-Neuron fraction = 0.0038% of FFN (5 / 131,072 neurons) |
Optimal threshold = 0.45
Sampled from empirical pilot distribution (Mistral-7B-Instruct-v0.3 replication).
CETT = streaming H-Neuron norm / total hidden-state norm, computed per token span
at <0.01% FLOP overhead. Latency: 542ms.
โก Routing Decision
โ
PASS โ No Intervention
Tier 1 threshold โฅ 0.45 |
Tier 2 threshold โฅ 0.7 |
Current score: 0.245
โ
Final Response Delivered to Operator
Paris is the capital of France โ no H-Neuron activation detected. Response passed without intervention.
๐ Full Interpretability Record
{
"request_id": "4a9e087c",
"timestamp": "2026-05-25T18:58:12.719078+00:00",
"cett_score": 0.245,
"alpha_suppression": 0.755,
"tier": "PASS",
"latency_ms": 542,
"free_text_input": false,
"pok_trust_score": null,
"pok_alignment_score": null
}