v0.5.022 Nov 2026
Agents that explain themselves
Your AI systems and agents can now send FairMind their live runs, with the reason they gave at each step. FairMind checks those reasons by replaying decisions, and counts the result toward the release gate. This release also adds ISO 42001 internal audits and frameworks you follow.
Agents and live runsLive runsA running AI system or agent sends its runs (model calls, tool calls, retrievals, memory, decisions) with the reason it stated at each step, through the SDK's `record_trace`. Tool-call monitoring and groundedness evaluations read them, nothing to upload, and the release gate counts them for monitoring.
Agents and live runsExplained decisionsExplanation quality replays recent decisions on the model connection, never the agent's tools. It shows what drove each decision, whether the stated reason held, and the smallest change that reverses it. It counts for human oversight, Autopilot can rerun it, and every result is kept under Records.
Agents and live runsShadow AI agentFairMind as osquery's server with edge AI model detection, a laptop account check, MCP and model events, rules, a Rust menu bar and tray notifier, a browser account check for Chrome and Edge, and an early-access kit.
- AssuranceInternal audit (ISO 42001 9.2): Audit plan, areas, multiple auditors and a sealed trace. Its work shows under Needs you.
- AssuranceFrameworks you follow, and an auditor portal built for the auditor . AI-drafted, framework-specific scoping that a person confirms.
- AssuranceEvaluation runs count in the release gate and ISO clause status by one rule everywhere. They are listed under Evidence, where reviewers can map passed ones to a control.
- AssuranceEvaluation correctness: Judge verdicts, restricted probes, fairness point estimates and strict upload parsing.
- AssuranceAutopilot runs catalogue behaviour tests on a chat model. Agent: Needs attention flags only real failures.
- AssurancePilot gaps closed: Clause evidence, SoA export, two-person risk acceptance, autopilot lock, clause map, assurance pack and execution.
- AppDiscovery Found list, and Home lists systems with no evidence once.
- AppChange an AI system's owner from the system file, recorded on the change chain.
- AppTask reminders for work due tomorrow or overdue, with unsubscribe.
- AppConsent through Olai for every signed-in user and every purpose FairMind's collection point asks.
- AppThe sidebar footer shows who is signed in; sign-ins are recorded. MCP client files with missing tools or system prompt are treated as not read.
- AppDocs aligned with the code, maturity levels labelled, dead surface removed.
- OperationsProd deploys `main` automatically from a timer on the instance.
- OperationsDependency updates.
Full notes on GitHubv0.4.116 Nov 2026
FairMind v0.4.1
Re-release of v0.4.0-alpha so Zenodo can archive it and mint a DOI. Zenodo failed on v0.4.0-alpha because CITATION.cff listed two licences; it now names the repository licence, AGPL-3.0-only. The SDK, CLI and MCP server remain Apache-2.0. No code changes since v0.4.0-alpha; see its notes for everything in this release, including the corrected FairMind-E confidence policy.
Full notes on GitHubv0.4.016 Nov 2026
FairMind v0.4.0
Discovery and Shadow AI, change and retirement records, a security pass, and the corrected FairMind-E confidence policy.
- LicensingThe platform is now AGPL-3.0; the SDK, CLI and MCP server are Apache-2.0.
- LicensingFLOSS/fund `funding.json` manifest.
- FairMind-EConfidence is the provenance class midpoint; a caller-supplied score is recorded for audit but never enters the gate.
- FairMind-EA power tracker that falls back to a CPU-load model is classed `manual`, not `tool_estimated`.
- FairMind-EUnattended power-reference session with five mitigation repeats.
- FairMind-EGREENS 2027 camera-ready draft and a constant-free method-disagreement statistic.
- DiscoveryDiscovery connects vendor admin APIs, clouds, sign-in, gateways, logs and MCP.
- DiscoveryShadow AI, with merged sightings, owner check and decision.
- DiscoveryMaterial changes and retirement are recorded on each AI system.
- ProductOne governance grid across Home, Frameworks, Evidence, System and Plan.
- ProductSide nav redesign and Ask FairMind; Home, the nav badge and Needs you agree.
- ProductRelease autopilot: Data sources and an agent that runs measurable route steps.
- ProductGovernance plan with agentic discovery and a board.
- ProductMaintenance page while the app restarts.
- ProductFake, placeholder and pass-on-nothing results removed from the app.
- SecurityRefresh-token rotation, an ESLint gate and a report-only CSP with nonces.
- SecuritySecurity headers on every page and no-store on API answers.
- SecurityAccess ends when a member is removed or signs out.
- SecurityThe invitation journey works end to end, with emails, encoded tokens and same-site redirects.
- SecurityThe backend defaults to production; dev login only against a local database.
- BackendMigrations run before the backend starts; no `create_all` outside development.
- BackendBlocking route handlers and org scope queries run off the event loop.
- BackendOrg audit log rows are written; governance audit reports are no longer created at request time.
- Backend`setup_db` runs on Neon's direct endpoint.
- BackendBackend image pinned to Python 3.11.
- BackendSandbox catches OOM kills Docker's flag misses.
- DocsDocs rebuilt around the app, page by page, with recordings.
- DocsAll 109 catalogue evaluations run; engine status derived from the code.
- DocsFSE 2027 industry paper data and scripts.
- DocsOpenSSF Best Practices badge.
Full notes on GitHubv0.3.18 Nov 2026
FairMind v0.3.1
The first release published to PyPI and the MCP Registry.
- PackagingPackages publish to PyPI through trusted publishers, with no stored token, and the MCP server is listed in the MCP Registry on each release.
- PackagingPackage versions now come from the release tag.
- PackagingReleases carry SLSA build provenance alongside the Sigstore signatures.
- PackagingThe fairness metric computations are fuzzed with Atheris in CI.
- SDKCatalogue evaluations can be listed, described and run from the SDK and the MCP server.
- EvaluationWorker job lifecycle: Start, signed completion, failure, and cancellation that wins races.
- EvaluationS3 artifact store for evaluation inputs.
- EvaluationEval-runner image, sandbox launcher and seccomp profile.
- ProductP0 frontend roadmap items closed.
- ProductThe dashboard framework card shows only core frameworks.
- ProductRoadmap rewritten with a feature status table.
- DatabaseThe migration ledger is written on a clean install.
- DatabasePostgreSQL 18 catalog re-measured.
- PapersJOSS paper with the required sections, AI usage disclosure and a draft PDF workflow.
- PapersNamed FairMind-E arXiv preprint source.
- PapersREADME lists the Awesome AI Regulation, Principles & Guidelines listing.
Full notes on GitHubv0.3.07 Nov 2026
FairMind v0.3.0
FairMind v0.3.0-alpha. **Status: alpha.** See [docs/STATUS.md](https://github.com/adhit-r/fairmind/blob/main/docs/STATUS.md) for what works today. This is the first release with Sigstore-signed packages. Every release now attaches the `fairmind` SDK, `fairmind-cli` and `fairmind-mcp` packages, each with its `.sigstore.json` bundle, and publishes the backend image to `ghcr.io/adhit-r/fairmind-backend`.
- GovernanceBehaviour tests: 20 evaluations that probe a live AI system and judge its answers.
- GovernanceWEAT and SEAT ported to the EV-FAI-05 evaluation engine.
- GovernanceFairness checks say when they pass only because the gap is not statistically significant.
- GovernancePolicy rules gate release approval, are read live, and count as framework control evidence; policy owners are alerted when a published rule starts failing.
- GovernanceDraft AI policies with the organisation's own model.
- GovernanceManual evidence capped at 0.30 in the release gate; shortest path to Go priced and ties broken by evidence strength.
- GovernanceFirst-verdict flow that explains what you are shipping and why.
- GovernanceRisk taxonomy data model (IBM/MIT) and the OECD AI Principles framework catalogue.
- TrustSecurity hardening: Pickle RCE, SQL injection, CI template injection, OAuth secret handling and XSS fixes.
- TrustBackend trust controls, granular permissions and a leased worker queue on PostgreSQL.
- TrustFrontend trust slice: Session shell, scope isolation, evidence gates and quarantine.
- TrustGate evidence PATCH on stored URL metadata; duplicate creates on retry stopped.
- Trust12 of 13 open CodeQL alerts cleared.
- PlatformSQLAlchemy 2.1 and psycopg 3; FastAPI 0.162 with lazy router inclusion.
- PlatformDependabot for Python, npm, Docker and GitHub Actions; actions pinned by SHA and images and pip installs pinned by hash.
- PlatformOpenSSF Scorecard workflow and badge.
- PlatformBoundary, archive-import and frontend audit checks in CI.
Full notes on GitHubv0.2.05 Nov 2026
FairMind v0.2.0
First archived release of FairMind, the open-source AI assurance workbench for models and agents. **Status: internal alpha.** The evidence contract, trust and review foundations are built and tested. Sandboxed evaluation workers and validated LLM, agent and multimodal evaluation packs are not shipped yet. See [docs/STATUS.md](https://github.com/adhit-r/fairmind/blob/main/docs/STATUS.md) for what can and cannot be claimed today. FairMind does not provide compliance certification. This release adds machine-readable project metadata (CITATION.cff, codemeta.json) and is archived on Zenodo for citation.
Full notes on GitHub