Primitive Reflex
A fast System-1 routing layer that selects relevant experts, skills, tools and context without granting execution authority.
Reflex is a provider-neutral routing contract under active development on the feature/primitive-reflex-v1 branch. It can choose an anchor expert, up to three constructive and three critical neighbours, relevant skills/tools/context and a recommended model route. It always returns executionPermitted=false; governed execution remains elsewhere.
Reflex is the receptionist who quickly decides who should look at a problem and which folders/tools are probably useful. The receptionist can route the work, but cannot sign the cheque, merge the code or press the deploy button.
Reflex operationalises the bounded expert-neighbour idea: one anchor plus up to ±3 related experts, with a hard maximum of seven. It compiles only authorised context, records why candidates were or were not selected, and escalates to a stronger reasoner when confidence, risk, novelty or authority conditions demand it. The neural adapter work is a separate experiment and cannot become authority by learning a classifier.
Implemented on a feature branch; not canonical main yet
The current Reflex implementation, tests, CLI and neural experiment exist on primitive-workspace branch feature/primitive-reflex-v1. Docs therefore marks it experimental until it is reviewed/merged into the canonical Workspace line.
Anchor + bounded neighbourhood
Default bounded selection
| Setting | Default |
|---|---|
| Confidence threshold | 0.78 |
| Minimum candidate score | 0.22 |
| Neighbour threshold | 0.30 |
| Positive neighbours | max 3 |
| Negative neighbours | max 3 |
| Skills | max 4 |
| Tools | max 6 |
| Context items | max 8 |
| Context characters | 16,000 |
| Provider timeout | 1,500 ms |
Fast routing abstains/escalates when it should
- Provider failure or no relevant anchor.
- Confidence below policy threshold.
- Medium/high risk or sensitive request.
- Authority decision is not allowed.
- Write/execute intent without governed execution admission.
- Novel request.
- No downstream model route available.
Every decision is explainable
The service stores routing evidence for experts, skills, tools and context, including scores, provider/version, selected status, risk, authority state, context reduction metrics and escalation reasons. It writes a private routing-trace artifact rather than treating the recommendation as an effect.
Learning may improve routing, not authority
The neural-reflex experiment tests compact specialist adapters and calibrated fallback. Its README explicitly claims no trained adapter, benchmark result or optimal topology. Training data requires consent/privacy review and human labels; predictions cannot become gold labels or authorise external tools.
Related pages
Canonical sources
These pages are explanatory read models. Where wording conflicts with implementation or a Control decision, the linked implementation/decision is authoritative and this page should be corrected.