Fact-Check-X

September 11, 2026 · View on GitHub

中文 · SkillHub · skills.sh · Awesome Skills · Agent-Skills.md · Releases · Security · Privacy

Fact-check complete answers and citations from one or more AI services: capture the originals without loss, compare atomic claims, verify authoritative evidence point by point, and evaluate platform performance. The final answer is one deliverable in a fully traceable evidence chain.

Fact-Check-X multi-platform fact checking: complete capture, claim comparison, authoritative verification and answer generation, platform evaluation

Quick start

skillhub install fact-check-x --namespace user_186e37d0

Install through the Agent Skills CLI:

npx skills add ASI2030/Fact-Check-X --skill fact-check-x-complete

The public listing is available on skills.sh.

The complete skill is also indexed by Awesome Skills and Agent-Skills.md. Both directories install from this repository's fact-check-x-complete source tree; each release verifies the public installation tree against the release source.

Claude Code users can install the repository as a versioned plugin marketplace:

claude plugin marketplace add ASI2030/Fact-Check-X
claude plugin install fact-check-x@fact-check-x-marketplace

Or download fact-check-x-complete.zip from GitHub Releases and install it in WorkBuddy, Codex, Claude Code or another Agent Skills-compatible host. Verify the download against the SHA256SUMS file from the same release.

Then ask naturally:

Use Fact-Check-X to verify this question: ...
Platforms: DeepSeek, Qwen and Doubao.

Officially supported platforms

Platform ID	Name	Capture
`dknowc-chat`	DKnow Chat / 深知晓	Standard answer, citations and official sources
`dknowc-deep-research`	DKnow Deep Trace	Runs from the normal answer in the same conversation and is saved as a separate platform result
`doubao`	Doubao	Complete answer, citations and page evidence
`yuanbao`	Tencent Yuanbao	Complete answer, citations, opened source bodies and page evidence
`deepseek`	DeepSeek	Complete answer, citations, opened source bodies and page evidence
`qianwen`	Qwen	Complete answer, citations and page evidence

The selected set is dynamic. N=1 runs a complete single-platform verification. N≥2 adds agreement, conflict and citation comparison. Platforms not listed here are not part of the current support commitment.

Four independent deliverables

Answer collection: complete answers, references, screenshots and HTML evidence.
Answer aggregation (unverified): atomic facts, claims, agreements, conflicts and source faithfulness, with direct answers and supplemental references separated in both overview and per-point detail sections.
Authoritatively verified final answer: direct answers supported by evidence, evidence boundaries, and a compact source index. Supplemental findings are stored separately and never mixed into the final answer.
Answer performance report: reads the locked stage-three result without recomputation, uses one per-point verification area with separate direct-answer and supplemental-reference subsections, then shows platform verdicts, accuracy, completeness, source quality and differences.

Every run uses machine-enforced acknowledgement tokens and artifact hashes between stages. Each report must be opened and presented before the user chooses whether to continue, revise, or stop; there is no mode that bypasses these confirmations.

Capture and comparison require no API key. Authoritative verification first reuses official material already attached by DKnow, or gov.cn material attached by another platform, when the captured body actually supports the claim. Trusted Search is used only for the remaining evidence gaps. On first use, the user signs in to the DKnow MaaS page; the skill obtains or creates a dedicated local key without asking the user to paste secrets into chat. Semantic analysis runs in the current host and does not call an external model API.

DKnow Deep Trace is a separate platform result

dknowc-deep-research is not an alias or a longer wait:

submit the original question in the same authenticated session;
wait for the normal answer to finish;
invoke the Deep Trace entry without resubmitting the question;
capture the inline result, or take over a same-product result page if one opens;
wait for it to finish;
save its answer, sources, screenshot and page evidence separately.

A missing button, missing result or incomplete result is a capture failure. When both platform IDs are selected, the question is submitted once while answers, references, screenshots and HTML evidence are stored separately. DKnow Deep Trace does not inherit the normal DKnow result; it can establish an official-material anchor only from its own captured evidence.

Requirements

Python 3.10+
Node.js 20+
macOS, Windows or Linux
Google Chrome, Microsoft Edge, Brave or Chromium
authorized accounts for every selected website
authorized 深知 MaaS access for non-exempt authority retrieval

Inside an extracted fact-check-x-complete skill:

python3 scripts/fact_check_x.py locate
python3 scripts/fact_check_x.py prepare-runtime

prepare-runtime is idempotent. On a clean install it uses the bundled lockfile to install dependencies and probes the collector CLI; when the runtime is already complete it exits without reinstalling. It does not reuse another skill's stale node_modules and skips Playwright browser downloads.

The visible workflow prefers an installed system Chromium browser. Install Playwright's test browser only for headless or CI validation:

cd modules/llm-answer-reference-compare/assets/tool
npx playwright install chromium
cd ../../../..

Installation

Download fact-check-x-complete.zip from GitHub Releases. Verify it against the SHA256SUMS file published with the same release.

WorkBuddy

Open Skills → Add skill → Upload skill, then upload fact-check-x-complete.zip.

The helper validates the archive and prints the manual upload instruction:

python3 scripts/install_skill.py \
  --source fact-check-x-complete.zip \
  --target workbuddy

Official reference: WorkBuddy Skills

CodeBuddy

# User scope
python3 scripts/install_skill.py \
  --source fact-check-x-complete.zip \
  --target codebuddy

# Project scope
python3 scripts/install_skill.py \
  --source fact-check-x-complete.zip \
  --target codebuddy \
  --project-dir .

The destinations are ~/.codebuddy/skills/ and .codebuddy/skills/.

Official reference: CodeBuddy Code Skills

Codex

python3 scripts/install_skill.py \
  --source fact-check-x-complete.zip \
  --target codex

The installer honors CODEX_HOME, defaulting to ~/.codex/skills/. Start a new task or refresh skill discovery after installation.

Reference: Using skills

Claude Code

# User scope
python3 scripts/install_skill.py \
  --source fact-check-x-complete.zip \
  --target claude-code

# Project scope
python3 scripts/install_skill.py \
  --source fact-check-x-complete.zip \
  --target claude-code \
  --project-dir .

The destinations are ~/.claude/skills/ and .claude/skills/.

Official reference: Extend Claude with skills

The installer rejects traversal paths, symlinks, runtime state and multi-root archives. It stops on collision; --replace moves the old skill to a timestamped backup before installing.

Trusted Search Key

Only knowledge points whose captured official material is insufficient require Trusted Search. Never paste a Key into chat, an issue, a report or shell history.

Getting a Key

Trusted Search Keys are provided by DeepKnown MaaS:

If you already have a MaaS account, sign in through the official login page, open API Key management and create a dedicated Fact-Check-X Key. Then run the automatic onboarding command below; the component will reuse the Key without requiring you to copy it manually.
If you do not yet have an account or Trusted Search access, visit the DeepKnown product site and follow its API access / beta participation path. Account eligibility, quotas, service scope and review requirements are controlled by the current MaaS service.
Do not reuse a shared production Key. Create and revoke the Fact-Check-X Key independently and grant only the permissions it needs.

Fact-Check-X uses DeepKnown MaaS, built by Beijing Caizhi Technology, because it provides traceable knowledge search for regulations, public services and industry standards through API and MCP interfaces. MaaS retrieves candidate authoritative material; the host agent still decides whether that evidence supports each claim.

Before authority retrieval, Fact-Check-X checks the process environment, the shared credential file and whether all knowledge points have qualified anchors. If a non-exempt point exists without a valid Key, prepare-authority returns configuration_required, a user prompt and the exact configuration command.

Standard onboarding:

python3 scripts/trusted_search_config.py configure

The component opens the 深知 MaaS page, waits for the user to complete identity checks, reuses an existing complete Key or creates a dedicated Fact-Check-X Key, validates it, and stores it at:

~/.fact-check-x/credentials/trusted-search-key

Check status without displaying the Key:

python3 scripts/trusted_search_config.py status

Clear the shared credential:

python3 scripts/trusted_search_config.py clear

Only 401/403 invalidates an existing Key. A timeout, network failure or service error preserves it and moves the run to review.

Playwright sessions and Computer Use recovery

Persistent browser state defaults to:

~/.fact-check-x/browser-profiles

Profiles reduce repeated login but never bypass login-state, ready-page, answer-completion or all-platform-success checks. login and run must execute in the foreground with their real exit status.

When automation cannot complete, capture-recovery.json records the selected platform, original question and recovery action. A host with Computer Use takes over the same visible page and profile, reuses the recorded question, lets the user handle password, SMS and CAPTCHA, waits for generation to stop, then reruns capture.

Do not replace this recovery with a headless browser, a new temporary profile, lock-file deletion or a partial-result shortcut. A host without Computer Use must stop at the original-answer capture stage.

CLI example

FCX_QUESTION="What is the current rule for the policy in question?"
FCX_RUN_DIR="./fact-check-x-run"

node modules/llm-answer-reference-compare/assets/tool/dist/cli.js run \
  --question "$FCX_QUESTION" \
  --platform dknowc-chat \
  --platform dknowc-deep-research \
  --platform doubao \
  --out "$FCX_RUN_DIR/capture" \
  --headed \
  --interactive \
  --timeout 180000 \
  --retries 2

python3 scripts/fact_check_x.py prepare-comparison \
  --results "$FCX_RUN_DIR/capture/results.json" \
  --run-dir "$FCX_RUN_DIR"

# Interactive mode returns a one-time token. After the user reviews stage one:
python3 scripts/fact_check_x.py acknowledge-stage \
  --run-dir "$FCX_RUN_DIR" \
  --stage capture \
  --token "<token returned for this stage>" \
  --decision continue

# The current host agent writes comparison-analysis.json.

python3 scripts/fact_check_x.py complete-comparison \
  --results "$FCX_RUN_DIR/capture/results.json" \
  --run-dir "$FCX_RUN_DIR"

python3 scripts/fact_check_x.py prepare-authority \
  --run-dir "$FCX_RUN_DIR"

python3 scripts/fact_check_x.py search-authority \
  --run-dir "$FCX_RUN_DIR" \
  --max-workers 12

# The current host agent writes one assessment per knowledge point.

python3 scripts/fact_check_x.py finalize-authority \
  --run-dir "$FCX_RUN_DIR"

python3 scripts/fact_check_x.py deliver \
  --results "$FCX_RUN_DIR/capture/results.json" \
  --run-dir "$FCX_RUN_DIR"

Stop on every non-zero exit and read the structured error. Never hide an exit code with a background job or permissive shell wrapper.

Output contract

The run directory contains:

capture/results.json
capture/report.html
capture/report.md
capture-gate.json
comparison-task.json
comparison-analysis.json
comparison.json
comparison.html
comparison-gate.json
authority/requests/
authority/evidence/
authority/assessments/
authority/results/
authority-gate.json
stage-checkpoints.json
verification.json
pipeline.json
01-capture-report.html
02-comparison-report.html
03-authority-report.html
04-final-report.html
05-complete-report-package.zip

The capture layer never rewrites an answer or replaces an original citation. The comparison layer never searches for final truth. The authority layer handles one atomic point per request. The final package is not produced unless all four reports exist, and it uses only relative paths.

Configuration

Variable	Purpose
`FACTCHECK_SKILLS_DIR`	Override the parent directory containing the four modules
`FACT_CHECK_X_BROWSER_EXECUTABLE`	Select a system Chromium executable
`FACTCHECK_BROWSER_PROFILE_DIR`	Override the persistent browser-profile root
`FACT_CHECK_X_TRUSTED_SEARCH_KEY_FILE`	Override the shared credential file
`TRUSTED_SEARCH_KEY`	Controlled process injection only; do not persist in chat or history
`FACTCHECK_TRUSTED_SEARCH_URL`	Controlled Trusted Search endpoint override
`FACTCHECK_TRUSTED_SEARCH_TIMEOUT_SECONDS`	Authority retrieval timeout
`CI`	Select CI browser behavior

Security and privacy

Review every third-party Skill before enabling it.
The question is sent only to platforms explicitly selected by the user.
Authority retrieval sends one atomic point at a time.
No external model API is called.
Browser state and the Trusted Search Key remain in the local user profile.
The repository and Release exclude profiles, Cookies, Keys, real sessions, real reports, caches and node_modules.
Fact-Check-X adds no telemetry; selected services apply their own policies.

See PRIVACY.md and SECURITY.md.

Troubleshooting

Problem	Correct response
Modules cannot be located	Check the installation depth and `FACTCHECK_SKILLS_DIR`
`npm ci` fails	Use Node 20+, preserve lock integrity and check the selected registry
Browser does not open	Read the real foreground exit; do not claim it launched
Input box appears while still logged out	Wait until the login entry disappears
CAPTCHA or SMS challenge	Keep the page open and let the user complete it
One platform fails	Repair or recapture it; the capture gate must stay closed
`capture-recovery.json` appears	Use Computer Use on the same page or stop at original-answer capture
`configuration_required`	Run the returned configuration command in the foreground
Trusted Search times out	Preserve the Key and retry up to three times; if it still fails, stop with a technical error
`insufficient_evidence`	Keep the evidence boundary explicit, exclude it from deterministic answers and accuracy denominators, and continue to the fourth report
Local report link cannot be shared	Upload the report package or individual report files

Validation

python3 scripts/audit_public_tree.py
python3 -m unittest discover -s tests -p 'test_*.py'

python3 skills/llm-answer-reference-compare/tests/smoke_test.py
python3 skills/fact-check-x-knowledge-compare/tests/smoke_test.py
python3 skills/fact-check-x-authoritative-verify/tests/smoke_test.py
python3 skills/fact-check-x-unified/tests/smoke_test.py
python3 skills/fact-check-x-complete/tests/smoke_test.py

Build reproducible public assets:

FCX_RELEASE_OUT="./release-assets"
FCX_VERSION="<release version>"
FCX_CANDIDATE_SHA="<candidate complete-package SHA256>"
FCX_OFFICIAL_SHA="<official complete-package SHA256>"

python3 scripts/build_release.py \
  --out-dir "$FCX_RELEASE_OUT" \
  --version "$FCX_VERSION" \
  --upstream-candidate-sha "$FCX_CANDIDATE_SHA" \
  --official-complete-sha "$FCX_OFFICIAL_SHA"

python3 scripts/verify_release.py manifest \
  --manifest "$FCX_RELEASE_OUT/release-manifest.json" \
  --asset-dir "$FCX_RELEASE_OUT"

Builds remain awaiting_promote without an official SHA and fail closed when candidate and official SHA differ. Exact hashes for a published version are recorded in its release-manifest.json and SHA256SUMS.

Version and SHA verification

Verify the formally promoted input:

python3 scripts/verify_release.py sha \
  --file fact-check-x-complete.zip \
  --sha256 <SHA256 published with the release>

Verify downloaded public assets:

python3 scripts/verify_release.py manifest \
  --manifest release-manifest.json \
  --asset-dir .

For an immutable GitHub Release:

gh release verify <tag> --repo ASI2030/Fact-Check-X
gh release verify-asset <tag> fact-check-x-complete.zip \
  --repo ASI2030/Fact-Check-X
gh attestation verify fact-check-x-complete.zip \
  --repo ASI2030/Fact-Check-X

fact-check-x-complete.zip is the formal installation baseline. Versioned module and suite ZIPs are deterministic distributions from the same public source tree and have independent hashes. release-manifest.json records both categories; their hashes must not be interchanged.

Contributing and license

Original project code and documentation are licensed under Apache-2.0. The existing MIT license for the llm-answer-reference-compare runtime is preserved. See THIRD_PARTY_NOTICES.md.

Apache-2.0 does not grant project-brand or third-party trademark rights. See TRADEMARKS.md.

Before contributing, read CONTRIBUTING.md. Public issues, pull requests, fixtures and screenshots must use synthetic or explicitly authorized data.