Skip to main content
Sediment is a self-hosted evidence store for coding agents. Engineering teams use the evidence to evaluate agent work, agents consume selected evidence as context, and training pipelines consume exports. Capture endpoints append immutable Facts. Evidence reads return selected captured parts. Derivations interpret Facts under a policy for reports and training artifacts, and export projections shape those artifacts for a training objective. This page explains the components that enforce that separation and the data that crosses each boundary. Large language model (LLM) gateways report inference calls, and continuous integration (CI) systems supply outcomes. The export stage names direct preference optimization (DPO), supervised fine-tuning (SFT), diff-shaped SFT (diff-SFT), and reinforcement learning from verifiable rewards (RLVR).

​
Component and data flow

Captured evidence becomes immutable Facts, then selected agent context, reports, and training exports.Captured evidence becomes immutable Facts, then selected agent context, reports, and training exports.
Evidence reads and Derivations share captured Facts. Reports answer questions about agent work; training exports project Attributed completions and Rollouts. The Recovery export reads Facts and mirror evidence directly, as described in the Export path.

​
Capture path

Agent harnesses report inference calls, Developer decisions, Edit observations, and Retry linkages. LLM gateways report inference calls through callbacks. Forges and continuous integration (CI) systems report Pushes and CI outcomes. Capture endpoints authenticate each source, validate its payload, and normalize it into the Fact schema. The endpoints append the resulting Facts to PostgreSQL. They don’t compute an Attribution, Reward, or training label. Sediment observes model traffic after an agent harness or gateway reports it. Sediment doesn’t sit in the model request path.

​
Stored evidence

PostgreSQL stores immutable Facts. Facts are the only persisted service state, and capture never changes an earlier Fact to add a later interpretation. Git mirrors hold repository history and refs/notes/sediment. Mirrors provide commit content, diffs, ancestry, and git-notes Attribution evidence. They form part of the raw substrate beside PostgreSQL, but they aren’t part of the Fact store. This division keeps observations exact. A Push records what the forge reported, while the mirror supplies the repository content that a later Derivation reads.

​
Derivation path

A Derivation reads visible Facts, stable git mirror state, and a resolved policy. It derives relationships such as Attribution, resolves CI evidence, and assembles training-shaped evidence without changing its inputs. Derivations are pure and recomputable. The same Facts, mirror state, and policy produce the same result. Sediment doesn’t persist a Derivation’s output as service state. A derived bundle can materialize the result as an immutable local build artifact without becoming a source of truth.

​
Export path

Attributed completions and Rollouts are Sediment’s two canonical artifacts. An Attributed completion resolves evidence around one inference call. A Rollout represents one Session’s full trajectory. Export projections turn Attributed completions into direct preference optimization (DPO), supervised fine-tuning (SFT), and diff-shaped SFT (diff-SFT) rows. A separate projection turns Rollouts into reinforcement learning from verifiable rewards (RLVR) tasks and trajectories. Recovery is ADR 0004’s one sanctioned Fact-derived export exception. Its red-to-green commit-pair shape can’t project from either canonical artifact. The Recovery Derivation therefore reads eligible Facts and the mirror directly to produce a row with the fixing diff.

​
Operational evidence reads

An operator can inventory a Session, inspect Inference-call message structure, and fetch selected canonical parts through the existing API. Each read uses one Quarantine-aware PostgreSQL snapshot and fixed source and response limits. The CLI publishes the result as a private local packet. Sediment stores no packet, checkpoint, retrieval index, or inferred task state. A pi agent can instead call sediment_retrieve_context. The API binds its retrieval credential to one configured source Session, reads a bounded complete visible population, and runs a pure keyword selector. The response carries complete selected parts, exact references, coverage, and counted exclusions. No model participates in selection. Agent-requested Session context defines this authority and selection boundary. This path reuses captured Facts without changing Attributed completions, Rollouts, reports, or training projections. A consuming agent treats the packet as historical data and supplies its own task instructions. No retrieval model or model-service dependency participates in these reads. See Bounded evidence access and Continue a task with captured evidence.

​
Network and trust boundaries

The API, PostgreSQL Fact store, git mirrors, Derivations, and exporters run inside your network. Developer machines, gateways, forges, and CI systems need an authenticated route to the capture endpoints. Capture validates evidence at the ingest boundary. Basic redaction replaces high-confidence API keys and bearer credentials before PostgreSQL writes a Fact. Quarantine excludes an unsafe or incorrect Fact from later Derivations without mutating or deleting the Fact. Sediment has no phone-home path. During installation, the installer and managed local PostgreSQL setup can download packages, host libraries, and PostgreSQL binaries. Operation without public network access requires dependencies prepared inside the perimeter. See Limit network exposure for runtime network exposure. During operation, optional repository mirrors connect to configured Git remotes. Internal forges, gateways, and CI systems can keep capture and mirror traffic inside your network. Evidence reads make no model call and add no outbound connection. If an agent sends captured content or an evidence packet to a model, its configured endpoint determines whether the content crosses the perimeter. Keeping the entire workflow internal also requires an internal model endpoint.

​
Continue reading