Relative confidence.
Softmax expresses preference within a supplied set. A high value is not proof of correctness. Thresholds need empirical validation.
One context.
Every choice on its merits.
ZEV is a non-autoregressive decision model. Supply context, a question, and candidate choices; receive scores and a probability distribution over those candidates.
An explicit information boundary.
Inter-Choice Masked Blindness (ICMB) lets each candidate attend to the context, query, and its own tokens. It blocks attention to other candidates. The context and query cannot read candidates.
Segment-relative positions and shared scoring parameters preserve this structure when you reorder the list.
ICMB ATTENTIONDUAL-STREAM RESIDUALSROUTED EXPERTS
| Reader | Context / query | Own choice | Other choices |
|---|---|---|---|
| Context / query | Allowed | — | Blocked |
| Candidate | Allowed | Allowed | Blocked |
No candidate-to-candidate attention. Context is processed once per layer.
THE GUARANTEEReorder the choices.
Reorder the choices.
Reorder the scores.
For fixed weights and unchanged segment text, permuting candidates permutes the score and probability vectors in exact arithmetic. Floating-point execution introduces numerical tolerances.
Candidate text, labels, and index references matter. Rewriting a candidate changes the input and falls outside this guarantee.
When scores tie, a “first result wins” rule can depend on order. Use candidate identity to resolve ties when your application needs order-independent selection.
Adding candidates preserves existing raw scores under the same assumptions, but changes probabilities through softmax normalization.
INTEGRATIONA small, explicit
A small, explicit
application boundary.
ZEV serves over a high-throughput HTTP API compatible with the System One decision standard. The engine can also be embedded directly in your Rust or Python process with zero network overhead.
The API accepts 2–10 choices per decision question. It returns the winning choice, calibrated confidence, probability distribution over all candidates, and token accounting with zero output generation lag.
Download the HTTP API contract ↓// REQUEST: POST /v1/systemone
{
"state": "Customer: My card was swallowed by the ATM dispenser.",
"questions": {
"intent": {
"type": "choice",
"instructions": "Route to the appropriate support department:",
"criteria": {
"card_swallowed": "ATM card swallowed by machine",
"atm_support": "General ATM transaction support",
"declined_cash": "Card declined during cash withdrawal"
}
}
}
}
// RESPONSE (200 OK — zero generation lag):
{
"model": "ZEV",
"answers": {
"intent": {
"choice": "card_swallowed",
"confidence": 0.9412,
"probabilities": {
"card_swallowed": 0.9412,
"atm_support": 0.0486,
"declined_cash": 0.0102
}
}
},
"usage": { "input_tokens": 58, "output_tokens": 0 }
}System One API standard (/v1/systemone). Returns winning choice, calibrated confidence, full probability distribution, and usage accounting.
SCALE & HIERARCHYScaling beyond 10 choices:
Scaling beyond 10 choices:
Two-stage cascaded routing.
ZEV enforces a bounded attention budget of K ≤ 10 choices per forward pass to maintain strict mathematical order equivariance and microsecond latency.
For high-cardinality enterprise catalogs (50–500 support queues, API endpoints, or workflows), production systems deploy a two-stage cascaded architecture:
Stage 1 (Micro-Embedding Filter): An ultra-lightweight static embedding lookup (e.g. Model2Vec, running in ~14 µs on CPU) narrows high-cardinality choices to a Top-8 candidate set.
Stage 2 (ZEV Decision Verification): ZEV runs a single non-autoregressive forward pass with ICMB across the candidates, producing calibrated probabilities and contrastive confidence.
Stage 2 (ZEV Decision Verification): ZEV runs a single non-autoregressive forward pass with ICMB across the candidates, producing calibrated probabilities and contrastive confidence.
BOUNDARIESKnow what you’re
Know what you’re
putting into production.
Architectural symmetry and predictive quality answer different questions. Validate both as part of your integration.
Application control.
Scoring does not validate an action or enforce permissions. Keep authorization, execution, and escalation in your application.
Independent scoring.
Tasks requiring direct comparisons among candidates may need a different model or additional context. ICMB deliberately excludes those interactions.
YOUR WORKLOAD. YOUR ACCEPTANCE CRITERIA.
Evaluate ZEV Make the next decision on evidence.
Bring a routing workload. Find out where ZEV fits.