Noul
Probability of yes on a binary question, with [P(no), P(yes)].
0
1
v0.3.0 · batch-native · beta
Decision infrastructure
without the chat layer.
Specter turns application state into bounded, typed numeric decisions — Noul, Choice, and Score — with canonical snapshots, batch-native backends, versioned calibration and explicit failure states.
Python 3.11+0 runtime depsBatchBackendOpenAPI 3.1
{
"id": "route",
"type": "choice",
"status": "ok",
"value": "engineering",
"probabilities": [0.05, 0.85, 0.10],
"confidence": 0.81
}bounded outputno fabricated fallback
01 / primitivesThree outputs.
Three outputs.
No ambiguous shape.
The framework keeps the decision surface deliberately small. Your backend supplies finite logits; Specter validates, calibrates, projects and returns a predictable contract.
Choice
One option from an explicit candidate set. Deterministic argmax projection.
Score
A continuous expected level over an ordered rubric you define.
Independent
Questions do not contaminate siblings.
Each evaluation sees the same immutable state snapshot, not preceding answers.
Explicit failure
Errors never become invented decisions.
Timeouts, invalid outputs and missing calibration are typed failure states.
Calibrate what you deploy
Confidence is an artifact, not a vibe.
Temperature, isotonic calibration and versioned domain artifacts keep uncertainty explicit.
02 / batch-native engineEncode once.
Encode once.
Decide many.
0.3.0 moves shared-state work into the kernel. A BatchBackend can prefill the state once for N question heads, while signature deduplication collapses repeated work before calibration and projection.
shared-state costCprefill + Q · Cheadbatch path amortizes state encoding; fan-out remains compatible
1×state prefill for a batch-capable backend
2.17×measured gain on 64 repeated signatures
231k/smeasured kernel decisions/s at batch 128
boundedLRU caches and deadline-aware admission
Engineering measurements use the repository benchmark on Linux x86_64, 1 vCPU, 4 GiB RAM and CPython 3.12.3. They measure kernel mechanics with cheap local backends — not model intelligence or universal production latency.
03 / measured engineeringMeasure the kernel.
Measure the kernel.
Not the myth.
The published benchmark isolates snapshot, validation, scheduling, softmax and projection. It deliberately does not claim semantic quality or trained-model speed.
batch 128 · p500.550 ms
231,293 typed decisions/s on the recorded 1-vCPU run.
signature dedup2.17×64 repeated questions
process RSS max25.4 MiBCPython included
questionsp50decisions/s
10.029 ms31,235
80.071 ms107,934
320.171 ms179,963
1280.550 ms231,293
04 / installFrom zero to a verified
From zero to a verified
framework install.
The pinned 0.3.0 wheel is served with this page. Create an isolated environment, install the exact artifact, then inspect capabilities. No Git clone required.
SHA256 pinned
- 1Create a virtual environment
Keep Specter isolated from your system Python.
- 2Install the 0.3.0 wheel
The command below resolves to this site's checksum-pinned build of the tagged source.
- 3Verify capabilities
Confirm version and contract, then optionally benchmark your own hardware.
terminal
Important: Specter 0.3.0 is a beta framework build. The bundled System-One local backend is an auditable lexical baseline, not a trained semantic model. A successful install or benchmark does not certify domain accuracy, calibration quality or production readiness.
05 / machine handoffSend one link.
Send one link.
Let the LLM do the careful part.
The site publishes a concise /llms.txt entry point plus a pinned install brief. An agent gets the exact artifact, SHA-256, environment rules, capability verification and the limits it must not silently cross.
Shareable endpoint
06 / architectureA narrow waist
A narrow waist
for decision systems.
Python applications can call the SDK directly. Specter Core can use the BaseHTTPRequestHandler adapter, while other runtimes integrate through the authenticated HTTP contract and OpenAPI schema.
POST
/v1/decision/evaluateGET
/v1/decision/capabilitiesGET
/v1/decision/openapi.json