v0.8.0

0.8.0 — Offline Agent Skill validation

• Bundle the exact zero-dependency evidence contract and validator inside the
  Doubt Agent Skill so clients can produce real receipts without npm registry
  or network access.
• Fail closed in the skill instructions: a file hash, node count, JSON parse,
  or unverified assertion can no longer be reported as successful validation.
• Publish a controlled Codex CLI before/after portability result: direct and
  implicit prompts move from fail to pass while the negative control stays
  pass, with sanitized raw sessions and independently verified artifacts.

Install

Agent Skill for Codex, Claude Code, Copilot, Cursor, Gemini CLI, and other
compatible clients:

npx skills add alsoleg89/doubt --skill doubt

CLI:

npx doubt-ai@0.8.0 demo --out doubt-demo.html

Inspect the Agent Skill on skills.sh
or try the zero-install playground.

Published artifact

doubt-ai-0.8.0.tgz is the exact npm-published tarball.

sha256:cf12ea09ce718542bd5668bebae439af5d4bd8f2d3c9d698f8353d8f3804a3f8

v0.7.1

0.7.1 — Focused public surface

• Remove retired experiments, generated decision artifacts, and obsolete site
  links so the repository and npm package expose one coherent product.
• Focus the public gallery and five-topic reader study on current evidence-map
  examples, including the Agent Skills vs MCP decision map.
• Add a fail-closed release preflight that aligns the version, tag, public
  metadata, Action pins, npm tarball inventory, and GitHub release notes.

v0.7.0

Full Changelog: v0.6.0...v0.7.0

v0.6.0 — Evidence snapshots and graph integrity

Closes the core validation gaps found in adversarial review.\n\n- rejects malformed and future source dates\n- requires retrievedAt and hashes recorded excerpt snapshots\n- enforces bounded locators and rejects repeated filler\n- rejects duplicate edges, directed cycles, and disconnected reasoning subgraphs\n- fixes claim and contradiction metrics\n- expands the deterministic adversarial benchmark to 20/20 cases\n- makes skill archive digests reproducible across the Node 18/20/22 CI matrix\n- lets the Action exclude intentionally failing benchmark fixtures\n\nAll CI, Evidence contract, and Pages checks passed on the release commit.

Doubt v0.5.0 — Agent Skills portability benchmark

Agent Skills portability: from documentation to evidence

Doubt v0.5.0 adds a six-source evidence map and a contribution-ready benchmark
for testing one unmodified Agent Skill across Claude Code, Codex, GitHub
Copilot, Cursor, and Gemini CLI.

What is new

• Interactive Agent Skills portability map
  with 5 claims, 6 evidence items, 6 official sources, 1 contradiction, and 1
  explicit unknown.
• Frozen synthetic fixture plus direct, implicit, and negative prompt classes.
• Machine-readable protocol, JSON Schema, result template, and zero-dependency
  result validator.
• Exact client/configuration metadata, raw-output paths, artifact paths, and
  SHA-256 receipts are required.
• Honest pass, fail, and blocked runs are accepted.
• Single-client submissions are forbidden from claiming cross-client
  behavioral equivalence.

The committed baseline is deliberately 0 submitted clients. Contribute the
first reproducible result in issue #9.

Install from npm

npx doubt-ai demo --out doubt-demo.html

Package: doubt-ai@0.5.0

Verification

• 34/34 automated tests
• 12/12 adversarial evidence-contract cases
• 5 topics / 25 keyed reader-study tasks
• Clean install and CLI smoke test from the public npm Registry
• CI on Node.js 18, 20, and 22
• GitHub Pages and the dogfooded Evidence Contract Action are green

Install the Agent Skill:

gh skill install alsoleg89/doubt doubt

Gate evidence maps in GitHub Actions:

- uses: alsoleg89/doubt@v0.5.0

Release asset SHA-256:

2ad56697cd9f80a0853b9ff7f4766afc7b70c33ae784f000e1bc0cfb5ce99c0a

Doubt v0.4.1 — Portable evidence share links

Doubt maps can now travel without an upload. Open a valid map in the zero-install playground, choose Copy share link, and a fresh reader gets the same validated map and SHA-256 receipt from the URL fragment. Browsers do not send the fragment in the HTTP request; if clipboard access is blocked, the link remains ready in the address bar.

Also fixed release accuracy: doubt --version now reads package metadata and is regression-tested.

Try the playground: https://alsoleg89.github.io/doubt/playground/

Clean Agent Skill install with GitHub CLI 2.90+:

gh skill install alsoleg89/doubt doubt

Release asset SHA-256: 6292e6aaeac1ee5dc5899d23787b5e0cf7ca547bd5df7402eaee2af8faf82d1b

Full changelog: v0.4.0...v0.4.1

Doubt v0.4.0 — Zero-install evidence playground

Paste evidence. Inspect the edges.

v0.4.0 adds a zero-install, local-only path from evidence-map JSON to a
validated interactive artifact:

Open the Doubt playground →

What it does

• Paste, open, or drop a .doubt.json file.
• See exact fail-closed contract findings.
• Generate the same canonical SHA-256 receipt as the Node CLI.
• Preview the existing interactive map renderer.
• Download portable JSON or one self-contained HTML file.

Map content stays in the tab. There is no account, analytics, telemetry, or
network submission. Rendered user content runs in a sandboxed iframe without
same-origin access.

One contract, not a web fork

The evidence contract is now a browser-safe module shared by the CLI, GitHub
Action, adversarial benchmark, tests, and playground. The Node wrapper adds
filesystem loading and synchronous SHA-256; invariants, findings, and metrics
come from the same implementation.

The public Agent Skills vs MCP map produces receipt
0444afe44c523678f2ad8eb7267e0d7c7a392709921abf16474e68d2ef5a3991
in both browser and Node paths.

Distribution and launch surface

• Replaced the internal dogfood headline with a product-first 1280×640 card.

• Added Open Graph metadata and a direct GitHub star CTA to the main Pages
  gallery and playground.

• Kept the verified GitHub CLI install path:

  gh skill install alsoleg89/doubt doubt

For CI validation of every *.doubt.json map:

- uses: alsoleg89/doubt@v0.4.0

Verification: 25 automated tests, Node 18/20/22 CI, 12/12 adversarial
evidence-contract cases, and production browser validation of both a valid map
and an eight-finding invalid map.

Full changelog: v0.3.1...v0.4.0

Doubt v0.3.1 — Verified one-command Agent Skill install

Verified one-command Agent Skill install

This patch adds the remote-publisher layout required by GitHub CLI 2.90.

gh skill install alsoleg89/doubt doubt

The command was tested end to end against the public repository in a clean
directory. GitHub CLI selected v0.3.1, discovered the four-file Doubt skill,
and installed it with repository/ref tracking metadata.

v0.3.0 added the project-level .github/skills/doubt layout used by GitHub
Copilot inside the repository. Remote gh skill discovery uses
skills/*/SKILL.md; this release adds that publisher layout and makes CI
byte-verify both copies against the canonical portable skill.

All v0.3.0 features remain included:

• Agent Skills vs MCP evidence map
• Five-topic map-vs-memo reader study
• balanced local assignment, anonymous JSON export, and no network submission
• fail-closed analysis until ten complete unique reader sessions
• 12/12 adversarial evidence-contract cases

For CI validation of every *.doubt.json map:

- uses: alsoleg89/doubt@v0.3.1

Full changelog: v0.3.0...v0.3.1

Doubt v0.3.0 — Evidence maps readers can test

Evidence maps readers can test

Doubt now ships a complete loop from a contested question to a source-grounded
artifact—and a preregistered test of whether that artifact helps readers more
than cited prose.

Try it

• Explore the Agent Skills vs MCP evidence map

• Take the five-topic map-vs-memo reader study

• Install the workflow with GitHub CLI 2.90+:

  gh skill install alsoleg89/doubt doubt

  Or install it across supported local agents:

  npx doubt-ai init --agent all

What changed

• Added a current, source-grounded Agent Skills vs MCP map with explicit
  contradictions, qualifications, and one decision-changing unknown.
• Added a preregistered five-topic reader study with balanced local assignment,
  25 keyed tasks, anonymous JSON export, and no network submission.
• Added a fail-closed analyzer that refuses condition-level conclusions before
  ten complete unique sessions.
• Added native .github/skills/doubt discovery for gh skill, enforced
  byte-for-byte against the canonical portable skill in CI.
• Replaced the Pages redirect with a gallery for the topical map, dogfood
  decision record, and reader benchmark.

The deterministic evidence contract remains 12/12 on its adversarial suite.
This proves structural traceability and safe rendering—not source truth or
faithful AI extraction. The reader study has no published winner until its
preregistered stopping rule is met.

For CI validation of every *.doubt.json map:

- uses: alsoleg89/doubt@v0.3.0

Full changelog: v0.2.1...v0.3.0

Doubt v0.2.1 — Discoverable Agent Skill

Doubt can now be discovered as a digest-pinned Agent Skill from the public Pages endpoint.

Highlights:

• Agent Skills Discovery v0.2 index and deterministic root-level tar.gz archive
• SHA-256 integrity verification and byte-reproducibility tests
• Public self-contained evidence-map demo
• 14 passing tests and 12/12 adversarial evidence-contract cases

Discovery index: https://alsoleg89.github.io/doubt/.well-known/agent-skills/index.json
Live map: https://alsoleg89.github.io/doubt/

Full changelog: v0.2.0...v0.2.1