Verified training data
Rubric-passing traces harvested from production, with rejection sampling on grader confidence.
Recorded with the runTrace and label lineage
Knowledge & learning
Context RL
Reinforcement learning as a service.
The loop that converts captured traces, rubric grades, and expert feedback into measurable improvement: better routing, distilled models you own, and releases that ship only when the evals say they should.
Context RL
Example workflow
Verified traces
Train candidate
Held-out eval
Release gate
Held-out evaluation · example Gate passed
Policy adherence82 → 94
Evidence quality78 → 91
Exception handling74 → 88
Release comparison
Baseline
78
overall score
Candidate
91
overall score
All configured thresholds cleared
The release record keeps the dataset, grader, model, and rubric versions together.
How it works
From request to a reviewable result.
Each run keeps the request, actions, evidence, and final decision together.
01 · Curate
Build from verified traces.
Context selects rubric-passing runs, removes low-confidence labels, and keeps a held-out set for release decisions.
Output · Versioned dataset
02 · Improve
Train or optimize the candidate.
Run distillation, routing, or policy optimization against the approved training split and grader.
Output · Candidate version
03 · Gate
Compare before release.
The candidate is evaluated against the held-out suite and promoted only when it clears the configured thresholds.
Output · Release decision
What you can inspect
Capability is only useful when it leaves evidence.
Grader alignment first
Judges are validated against expert consensus before any optimization runs against them.
Recorded with the runAgreement by rubric
Distillation to models you own
Open-weight students trained on your verified work, deployed behind Context Inference.
Recorded with the runDataset and model version
Eval-gated releases
Candidates ship only after clearing the held-out thresholds your team defines, with regressions visible by rubric and cohort.
Recorded with the runBaseline comparison
Deployment boundaries
Control is part of the product.
Context RL can run as a standalone surface, while its traces, evals, and policy decisions remain part of the same Context system.
A gate reflects its suite
Held-out evals expose measured regressions; they do not guarantee performance on cases the suite does not represent.
Experts define success
Context automates the loop around rubrics, but domain owners still approve graders, thresholds, and releases.