Skip to content

ROADMAP_2026.md

Latest commit

 

History

History
76 lines (59 loc) · 4.99 KB

ROADMAP_2026.md

File metadata and controls

76 lines (59 loc) · 4.99 KB
Raw

Public Roadmap 2026

This roadmap describes the open-source ClaimBound foreground. It does not include private systems and does not change any private or already-submitted materials.

Position

ClaimBound is an open-source evidence-card toolkit for narrow public AI, ML, data and software claims. It should remain independently usable from this repository.

It is not a model leaderboard, certification authority, hosted review queue or raw-data archive.

Public Work Plan

This plan describes the current public implementation path: harden an existing prototype into reusable evidence-card infrastructure.

Phase Work package Deliverables Acceptance
1 Public baseline consolidation Release checklist, link/metadata audits, status wording, card/report/protocol consistency checks validate-all passes; artifact-only records are not presented as completed cards.
2 SourceProbe v1 Source URL metadata, final URL, content type, byte size, SHA-256, marker checks and source-boundary recommendations Probe supports public AI documentation and public-data source audits; probe output is not a positive result by itself.
3 Scaffold workflow hardening Request, protocol charter, playbook, checklist, operator declaration, draft card and registry-patch scaffolds One command can prepare a safe scaffold; drafts stay clearly gray/not evidence.
4 Manual and AI-assisted workflows Local run roots, deviation logs, source-rights checklist, raw payload/transcript policy and AI stop rules Operators can follow manual or AI-assisted paths without hidden steps; final status remains human-reviewed and validator-checked.
5 Public AI claim protocol Source-audit rules for public AI docs, model/API metadata fields, prompt/transcript hash policy and screenshot-only limits Source-audit cards do not imply model safety, runtime behavior or benchmark superiority.
6 Static registry MVP Generated read-only registry views, filters by status/source/domain/audience/reproduction and SVG links Registry remains inspectable without hosted accounts or a database; entries validate against card JSON.
7 Independent rerun workflow Reproduction request template, rerun checklist, reproduction status fields and source-byte drift handling Reruns can update evidence without changing the original claim.
8 Completed example set and tutorials Honest cards or scaffolds across public AI, public data, reproducibility and software examples Project credibility does not depend on only positive results.
9 Release, review and sustainability Tagged release, maintainer guide, dependency/security notes, AI-use disclosure summary and follow-on roadmap The repository remains independently usable after this work.

Work Boundary

The public work package should produce reusable open-source outputs:

  • validators and test coverage;
  • source-boundary tooling;
  • static registry generation;
  • manual and AI-assisted runbooks;
  • documented rerun paths;
  • examples, tutorials and release process.

It should not produce:

  • closed private systems;
  • private data or private transcripts;
  • hosted account systems;
  • authenticated write APIs;
  • commercial hosting;
  • legal certification services;
  • broad product or private performance claims.

This boundary keeps the work small enough for a single-maintainer open-source project and clear enough for external review.

Strong Public Examples To Keep Visible

Example class Current example Why it matters
Public AI source audit Anthropic, OpenAI, Google DeepMind and Grok source-audit cards Familiar AI transparency sources, safely narrowed to source boundaries.
European public data EEA source audit and EEA AQ manual blocked card European open-data source-boundary and blocked-source discipline.
Narrow positive empirical result NASA POWER D-103 Shows that fixed public-data gates can produce reusable evidence.
Honest negative result NOAA CO-OPS D-131 Shows a protocol can fail without being renamed as success.
External operator path Starter pack, rerun issues and boundary-review issues Shows how another reader can inspect, rerun or challenge a card.

Risk Register

Risk Control
Scope grows into a general verification platform Keep work to validators, source-boundary checks, registry, workflows and examples.
Source audits are misread as safety or runtime claims Repeat claim boundaries in cards, docs and issue templates.
Blocked/negative cards look like project failure Explain that blocked and negative outcomes are first-class evidence.
Artifact-only records are confused with validated cards Keep artifact-only records in a separate section and mark them as not evidence cards.
Single-maintainer bus factor Publish release process, maintainer boundary, governance notes and reproducible workflows.
External operators do not know where to start Maintain issue templates, rerun workflow and external operator starter pack.