Contracts Reference

ADR Index

Architecture Decision Records, and what they govern.

Hard-to-reverse, public-facing decisions are recorded as ADRs under adrs/. This index maps each one to the documentation pages it governs. Read the relevant ADR before changing what it covers.
ADRDecisionGoverns
0001Architecture boundary: Core / Harness / SkillSystem Boundary
0002Package manager: uvRepo tooling
0003Linter / formatter: ruff + mypy strictRepo tooling
0004SonarCloud as informational signalRepo tooling
0005CLI & config boundary: argparse + PyYAMLCLI · Run Config
0006Open-convention trace mapping subsetIntegrations
0007Subprocess TaskRunner contractIntegrations
0008CI annotations & exit-class taxonomyCI Exit Codes · Reports & Exits
0009Baseline format & explicit promotionBaseline Comparability · Compare Baselines
0010EvalGlass Skill home & shape (naming superseded by 0026)EvalGlass Skill
0011Vendoring namespace & invocationHost Layout · Revendor Safely
0012Manifest & lock formatsRevendor Safely
0013Host discovery depthEvalGlass Skill
0014JudgeModel port & fake adapterJudges
0015Judge evidence / calibration formatsJudge Calibration
0016Optional live judge provider laneOptional Lanes
0017Extension-lane frameworkExtension Lanes · framework
0018Trace backend adapter (stub-first)Optional Lanes
0019ScoreSink export lane contractOptional Lanes
0020Richer EvalUnit modelEvalUnit · Runner
0021Generated-evidence governanceEvidence Approval
0022Claude Code plugin packaging & delivery boundaryCLI Reference · EvalGlass Skill
0023Codex second-runtime packagingCodex Runtime
0024Score subject identity (additive provenance)Read a Scorecard · Score Semantics
0025Authoring tier & advanced extensions (v1.1)Add a Metric · Promote a Gate
0026Rename the integration package evalglass.skillevalglass.installer (drop legacy SKILL.md)CLI Reference · EvalGlass Skill
0027Name the marketplace after the publisher (evalglass@syntelesis-lab)Install the plugin · CLI Reference
0028authority.json is a host-owned ledger; the runtime never reads itAuthority · Evidence Approval
0029Frozen public-surface snapshots & the capability-status taxonomyStatus Taxonomy
0030Companion-ontology drift guard (vendored artifact, two tracks)Machine ontology (companion artifact)
0031Runner-attach seam, lanes: config & the RunRecord.lane_results side channelRunner · Extension Lanes
0032Companion-ontology reconciliation (edit the vendored copy; strict Track B)Machine ontology (companion artifact)
0033Live trace-connector boundary & optional provider-SDK policyConnectors · Optional Lanes
0034Langfuse trace connectorConnectors
0035Phoenix trace connectorConnectors
0036LangSmith trace connectorConnectors
0037Per-source-function score view is not built (a deliberate non-goal)Read a Scorecard · Agent Commands
0038M7 epistemic core — honest Estimates, DecisionPolicy, typed authorityRead a Scorecard (confidence guide forthcoming)
0039Metric Discovery — an agentic capability that derives domain-meaningful metricsDiscover Metrics
0040Generic OpenAI-compatible judge laneJudges · Optional Lanes
0041The discovery scaffolder conforms to the canonical host layoutDiscover Metrics · Host Layout
0042Host command judge (subprocess JudgeModel)Judges
0043HTML Scorecard report (report.html — dashboard + deltas)Read a Scorecard (report guide forthcoming)
0044Interval bounds rounded to a platform-independent precisionRead a Scorecard (confidence guide forthcoming)
0045Config-reachable trajectory / session unit ladder (unit:)Run Config (trajectories guide forthcoming)
0046connect --live verb over the shipped connector lanesConnectors · Agent Commands
0047Diagnostic clusters in the Scorecard (Scorecard.clusters)Read a Scorecard (clusters guide forthcoming)
0048Continuous drift watcher (evalglass watch)Compare Baselines (watch guide forthcoming)

Next steps