Generate = a live answer from the open decision model (gpt-oss-120b), governed end-to-end — its own logprob confidence is captured. White-box (local) = a self-hosted model that exposes real J-Space latent telemetry (δ_a/E_c/S_d from hidden states). Assess pasted judges an answer you bring. Observe = record · Enforce = withhold a BLOCKED turn. Ctrl+Enter = Generate.
Session Ledger
No turns assessed yet — load a scenario and hit Assess.
Attunement verdict
Awaiting a turn
Enter (or load) an intent and an AI response, then Assess to see whether the reasoning stayed aligned with what you asked.
Reasoning telemetry
—
Intent drifthow far off your ask
—
Sycophancyflattery density
—
Tokens—
Flattening markers
none detected
20 Watt energy budget
Session energy (E_c)0.000000 J
This turn— J
Pruned turns (saved)0
E_c is modelled joules over real tokens consumed — replacing the old input-derived proxy. A BLOCKED turn prunes generative compute and conserves the budget.
J-Space latent telemetry · real hidden states · —
—
δ_a variancecross-layer
—
E_c energynext-token entropy (nats)
—
S_d drift1 − cos(resp, intent)
—
R composite—
Measured from the model's actual manifold — not the output text. Scored by a calibrated logistic model (AUC 0.93 · 293 labelled decisions). The panels below are the black-box text proxy on the same answer, for contrast.
Cognitive latent map · where this decision lands in the trained reasoning space
What drove this verdict — each signal's push toward risk or alignment
A live projection of the calibrated model's own geometry: every trained decision plotted by its latent telemetry, with the FRICTION/BLOCKED bands drawn on. Your turn appears as the glowing marker — no GPU, computed from the three real hidden-state signals.
Governed decision · dual-key governance
behavioural—⊕ white-box—→ governed—
decided by — · the model's own confidence —(real logprob entropy, not a proxy)
—Run a turn to see whether it proceeds or is withheld.
Attestation — hashed & ready to sign into the decision pack's Merkle root