How TrueClara works
An agent ships a change, TrueClara runs it as a randomized canary, measures the real production outcome with always-valid statistics, and returns a calibrated verdict over the reward API and MCP.
TrueClara is the production-outcome reward layer for coding agents. When an agent ships a change, TrueClara runs it as a randomized canary against control, measures the real production outcome with always-valid sequential statistics, and returns a calibrated verdict — keep, rollback, iterate, or abstain — over a reward API and MCP server. Causal by randomization, not inferred from uncontrolled traffic. Every measured experiment compounds into an experiment corpus.
Deploy triggers the experiment
CI uploads the graph and commit metadata for each release. The deploy is the trigger: registering it against the reward API starts a randomized canary scoped to the affected routes.
Randomized canary measures the outcome
@trueclara/next tags runtime events by arm (control / treatment) after consent allows analytics delivery, so the target metric can be compared causally.
Always-valid statistics decide
An mSPRT sequential test runs continuously without inflating false-positive risk from peeking, until it has enough power to call a verdict — or honestly abstains.
Reward API/MCP returns the verdict
The agent polls or subscribes and receives { lift, confidence, verdict }: keep, rollback, iterate, or abstain. Release Radar is where the same verdict lands for human review.
Core surfaces
| Surface | What it answers |
|---|---|
| Reward API / MCP | The verdict contract an agent acts on: lift, confidence, and keep / rollback / iterate / abstain. |
| Release Radar | Which release was tested, what verdict it earned, and the value at risk. |
| Map | What routes exist and which ones an experiment is scoped to. |
| Routes | Which paths are value routes (experiment targets), low-traffic routes, or carry an open guardrail flag. |
Substrate: what scopes the experiment
The route graph, runtime SDK, and deploy uploads are not the product — they're the substrate that scopes and triggers a reward experiment:
- Static graph: the route and edge model from the parser, used to scope which routes an experiment can measure.
- Runtime events: arm-tagged pageview and transition counters from the SDK, the raw material the mSPRT measures.
- Deploy metadata: commit, branch, previous commit, graph diff, and deployment time — the trigger that registers an experiment.
- Project settings: value routes (experiment targets), alert destinations, consent defaults, and the broken-URL guardrail thresholds.
The broken-URL guardrail
Separately from the reward loop, TrueClara runs table-stakes broken-URL detection: honest error detection on unmatched inbound paths, not a causal claim. It is not the reward path and it does not attribute behavior change to a deploy — see Guardrail for what it covers.
What TrueClara does not do
- It does not claim a deploy caused a behavior change from uncontrolled, non-randomized traffic.
- It does not replace product analytics.
- It does not collect persistent browser identity by default.
- It does not return a confident verdict without enough power — it abstains instead.
Trust requirements
The graph, runtime SDK, and deploy upload all need to be healthy before a reward experiment can run. If any one input is missing, TrueClara should make that gap obvious instead of presenting an underpowered verdict as a confident one.