Open-source · Self-hosted

The evidence store for coding agents.

Capture coding-agent work on your infrastructure. Use the evidence to evaluate agents, reuse context, and build training datasets.

Quickstart GitHub
01

What Sediment does

  1. Evaluate agent work

    Compare models by recorded decisions, code retention, and check results. Measure agent work →

  2. Reuse evidence as context

    Give an agent selected messages, tool calls, and results from a previous session. Continue a task →

  3. Build training datasets

    Export fine-tuning, preference, and reinforcement learning examples with traceable labels. Choose a training export →

02

Capture the evidence once

Capture five kinds of evidence. Available signals depend on your capture integrations.

  1. Inference call

    Prompts, responses, and model details from gateway callbacks.

  2. Developer decision

    Accepts and rejects, with human choices distinguished from automatic actions.

  3. Edit

    How much of the agent's edit remains at session end.

  4. Commit and push

    Attribution through git notes or similarity matching, with the source recorded.

  5. CI outcome

    Recorded check results across workflow runs and retry attempts.

Run capture, storage, and exports on your infrastructure. Your agent's model endpoint determines where inference data goes. Read the network boundaries.

03

Export training data

Choose an export for your training objective. Each recipe defines which evidence qualifies a row.

  • SFT

    Supervised fine-tuning examples selected using human acceptance, edit retention, or clean CI passes. Includes diff-shaped exports.

  • DPO

    Preference pairs from matching prompts, labeled by explicit human decisions or clean CI outcomes.

  • RLVR

    Tasks and captured trajectories for reinforcement learning from verifiable rewards. Labels use recorded CI results; rerunning checks requires your environment and verification command.

Example RLVR task · excerpt
{"repo":"acme/payments","problem_statement":"Fix the retry backoff test","recipe_id":"rlvr_ci","reward_source":"resolved_ci_pass","attribution_source":"git_notes"}
04

Get started