Bounded research guidance
Research Copilot and literature discovery can prepare clearly scoped proposals from recorded context, while reviewers retain the decision to create or advance work.
What's New
Research decisions now retain their exact source and lineage.
v0.5.0 hardens immutable literature, grounded Evidence citations, patch execution, approved implementation diffs, bounded execution, and research memory without treating fixture tests as live validation or automatically accepting governance evidence.
Reviewable experiment lineage
ParcelDigits now shows one baseline root, parent-linked variations, current-best comparisons, and visible failed or inconclusive outcomes.
Trusted runner boundary
The bounded ParcelDigits path can run through the trusted runner with protected evaluation and recorded constraints; other projects remain non-executable.
Focused Article 50 readiness
Customer-facing assistants now receive role- and use-specific transparency actions and a compact Transparency Pack preview, separate from Article 4.
Public changelog
Product updates
v0.5.0
Patch-execution integrity and truthful campaign readiness
July 30, 2026
- Research Setup now separates configured OpenAlex, Copilot and trusted-runner capabilities from the selected campaign’s objective, Evidence, lineage, envelope, package, patch and approval gates.
- Resolvable Evidence citations, immutable source provenance, current task packages, valid patch review and protected-path validation remain inspectable without exposing raw payloads as the default workflow.
- Research Memory and inspectors now preserve decision-first scanning at mobile widths, and OpenAlex results show transparent literal title and abstract match indicators without re-ranking provider results.
v0.4.0
Research integrity and decision-first product surfaces
July 29, 2026
- Literature-to-Evidence provenance now retains the exact immutable source version, extraction and provider records, grounded candidate, and reviewer acceptance.
- Paid extraction actions require explicit preflight confirmation, cited proposals retain their structured context snapshot, and selected experiments carry one canonical parent.
- Implementation diffs must be visible and approved before the bounded ParcelDigits runner can execute; parent and root comparisons remain reviewable before a finding decision.
- Product-wide density, responsive inspector, modal accessibility, and historical/demo provenance treatments now support a calmer release-validation path.
v0.3.1
Live research closeout
July 29, 2026
- Literature extraction now distinguishes live provider, queue, timeout, authentication, structured-output, and grounding outcomes instead of showing a failed request as zero candidates.
- Exact supporting excerpts are located in the immutable source version by Epok; imported literature remains awaiting extraction until a reviewer accepts grounded evidence.
v0.3.0
Bounded research and focused transparency readiness
July 28, 2026
- Research Copilot, literature context, proposal batches, experiment lineage, and reviewer-led findings now support a bounded research path before governance evidence review.
- ParcelDigits is the only supported trusted-runner demonstration, with protected evaluation and explicit boundaries around what the platform does not execute or advance automatically.
- Focused Article 50 readiness gives customer-facing assistants a role- and use-specific transparency path while keeping Article 4, draft evidence, and reviewer decisions distinct.
v0.2.0
Research Workspace and repository bridge
July 12, 2026
- Organisation Research Library and project Research Workspace added for Mosaic Vision research memory, task planning, and reviewer state.
- Five seeded Mosaic Vision projects show how datasets, repositories, experiments, findings, risks, and decisions connect before governance evidence is accepted.
- Deterministic Research Task Packages and repository/result import flows support a real ParcelDigits MNIST experiment without live repository checkout or code execution inside Epok.
v0.1.2
Trust and private-beta readiness
July 11, 2026
- Public Trust and Data Handling page for beta data boundaries, source-mapped drafts, Logger token handling, and human review expectations.
- Homepage, Product Flow, Pricing, AI Act basis, Trust, and What's New copy aligned to the two-system demo and paid evidence sprint.
- Release CI expanded across frontend, backend, logger, Docker Compose config, lightweight secret scanning, and CodeQL.
v0.1.1
Evidence Run and project review stabilization
June 17, 2026
- Evidence Runs validate runnable CSV/XLSX Dataset Cards before queueing and fail early with clearer diagnostics.
- Completed Evidence Run actions stay inside the selected AI System workspace.
- Project Overview, Evidence Run, Dataset Card, and Evidence Pack views emphasize status, blockers, next actions, and destinations.
Release posture
Built for evidence review, not legal sign-off.
Epok helps teams collect, connect, and review AI System evidence. v0.3.0 keeps research proposals, findings, draft evidence, and reviewer decisions separate from legal advice, certification, regulatory approval, and guaranteed compliance claims.