Loading…
Wrap any local model, Jev-style: it keeps working, but can now decide: a typed answer with measured confidence. Confident ones run instantly; the rest fall back or escalate. Copy the wrapper into your app. No server, no key.
Confidence 0.8821 clears 0.5; entropy 5.899 bits within 6. Code decided, not the model.
Synthetic sampleAda, backend developer shipping an LLM feature
Your own machine
--:--:--
Inference runs in this tab. Nothing is sent anywhere.
| Case | Top-1 | Margin | Verdict | Mass |
|---|---|---|---|---|
| 88.0% | Confident | |||
| 68.0% | Confident | |||
| 31.0% | Uncertain | |||
| 22.0% | Refused | |||
| 50.0% | Refused | |||
| 45.0% | Uncertain | |||
| 62.0% | Confident | |||
| 65.0% | Confident |
Filled mark: top token mass. Hollow mark: top two combined. Center line: 50% majority threshold. A narrow gap between marks indicates model hesitation.
Run your own prompt through the kernelPrompt
The capital of France is
Candidates
One token dominates with 0.74 bits of entropy. This is what certainty looks like.
score your own prompt on /livelib/kernel.ts
Runs on your device. No account, no API key needed. Every decision has a re-checkable receipt. Numbers re-derived by one command.
5 / 5 held on WEV-01
Every figure in those windows is recomputed in your browser by lib/kernel.ts from the cases committed to evidence/campaign-report.json. The same kernel pnpm claim:verify runs. No server is involved: the measured gate is on /proof.
Six on-device stages, from raw token logits to a tamper-evident receipt. Select any stage to inspect its phone screen and live kernel invariants.
Input
My card shows two $22 payments for one order
on this device
Model
SmolLM2 · 135M
cached · ~130 MB
"My card shows two $22 payments for one order"
One real held-out decision (T-002), end to end. Run it live on /live or stream items through /inbox.
Held-out decisions the gate got wrong stay on screen with the ones it escalated: counted, never hidden.
“The invoice has the wrong company name, can you fix it” → refund (gold: review)
Auto-handled at 89.9% confidence but the gold label is review. Counted on the gate report, never hidden.
“Do you ship to Taipei, and how long does it take” → refund (gold: review)
Auto-handled at 83.7% confidence but the gold label is review. Counted on the gate report, never hidden.
Executes SmolLM2-135M locally on CPU via transformers.js. Zero tokens leave the device.
lib/kernel.ts evaluates probability mass with plain arithmetic. The model proposes; code decides.
Closed-label confidence plus full-vocabulary entropy. Below the calibrated bar, the item escalates.
Strict verdicts: AUTO acts, FALLBACK generates marked unverified, ESCALATE queues a human. Malformed output fails closed.
Every committed run re-verified across INV-1 through INV-5 on every build (pnpm claim:verify).
Every verdict signs an immutable canonical digest. Any tampered byte exits non-zero.
Open a prompt, watch the probabilities, read the verdict. Everything runs in your browser. Nothing to install, nothing to sign up for.