Technical docs
Verdict.
A model router that can say no.
Architecture, routing policy, context handling, delivery workflow, and evidence for Verdict, a fail-closed control plane for AI coding tools.
Start here
The three-minute brief.
Reviewed against the public repository on 2026-09-24 · ff18aa5f2be0ea9b8ab6e32b56e713801b2d8b83.
- What is it?
- A control plane that decides which AI models may run a coding task before any ranking happens, and records why every other candidate was dropped.
- What problem does it solve?
- Routers that score models can still pick a stale, unqualified, or policy-excluded model, and they leave no record of why. Verdict separates admission from preference. When nothing qualifies, it blocks instead of silently falling back to an expensive model.
- What did Nicholas build?
- The architecture, the decision records, and the Python reference implementation: policy gates, context packing, receipts and replay, the CLI and HTTP API, gateway and coding-tool adapters, and the evidence tooling.
- What runs today?
- Eligibility gates, fail-closed routing, receipts, replay, and a credential-free demo are in the public repository and covered by CI. Live-gateway routing and coding-tool adapters are experimental.
- What supports the claims?
- A claims ledger that ties each public claim to source files and tests, a reproducible quickstart that needs no credentials, and dated receipts for live observations. Numbers without a reproducible artifact are not claimed.
Go deeper
Articles.
Architecture, trade-offs, tests, and decision records for engineers who want the detail.
- 01
Start here
Verdict overview
What Verdict is, the model-routing problem it addresses, and what is shipped today versus experimental or planned.
Status: ShippedStatus: ExperimentalStatus: Roadmap - 02
Architecture
Architecture: harness, Core, and gateway
How Verdict separates coding-tool harnesses, the Core decision plane, and an optional execution gateway, and how decision-only, completion, and worker paths differ.
Status: ShippedStatus: Experimental - 03
Routing
Dynamic model selection
How Verdict decides which models may run a task: eligibility gates, health qualification, free and paid policy, explicit worker selection, and fail-closed outcomes.
Status: ShippedStatus: Experimental - 04
Context
Context Intelligence
How Verdict plans and hydrates context for a task, keeps provenance on every unit, and enforces a token budget with disclosed omissions.
Status: ShippedStatus: Experimental - 05
Workflow
Autonomous development workflow
How a ticket moves from Linear through context hydration, a routed worker, layered verification, and a proof-gated pull request, and which parts are shipped versus planned.
Status: ShippedStatus: Experimental - 06
Observability
Routing receipts and observability
How Verdict records and explains each routing decision with privacy-safe receipts, an explain endpoint, and a dependency-free verifier.
Status: ShippedStatus: ExperimentalStatus: Roadmap - 07
Evidence
Benchmarks and evidence
What Verdict's public evidence supports, how to reproduce it, and which claims are deliberately not made.
Status: ShippedStatus: Experimental - 08
Decisions
Engineering decisions
A curated set of Verdict's architecture decision records: the choices that shape how it routes, proves, and integrates, and the trade-offs each one accepts.
Status: ShippedStatus: Experimental
Reading these docs
What the labels mean.
- Status: Shipped
- Implemented in the public repository and covered by automated tests or a reproducible, credential-free check.
- Status: Experimental
- Implemented, but depends on a live gateway, provider, or local runtime. Evidence is limited to dated observations or partial certification.
- Status: Roadmap
- Planned or accepted in a decision record, but not implemented or not yet verified.