AgentSkillsCN

public-source-gating

在法律权威核查与幻觉分类过程中，严格遵循“仅使用公开来源”的验证原则。

SKILL.md
--- frontmatter
name: public-source-gating
description: Enforce public-sources-only verification for legal authority checking and hallucination classification.
version: 0.2

Hard rules (public sources only)

  1. •

    Every authority/proposition outcome MUST be one of:

    • •VERIFIED_CORRECT (public source retrieved and supports)
    • •VERIFIED_ERROR (public source retrieved and contradicts OR shows mismatch)
    • •UNVERIFIABLE_PUBLIC (not found / ambiguous / blocked / missing coverage)
  2. •

    NEVER state "does not exist" unless you have retrieved material proving mismatch. If you cannot retrieve the authority text from public sources, label UNVERIFIABLE_PUBLIC.

  3. •

    Every conclusion MUST include:

    • •retrieval URL(s) attempted,
    • •timestamp,
    • •cached artefact path + SHA256,
    • •short reason for any failure.

Source Acquisition Policy (v0.2)

Permitted Sources (Priority Order)

  1. •

    Find Case Law (National Archives) - Primary for UK case law

    • •Preferred for all UK neutral citations
    • •Deterministic URL construction before searching
    • •Atom feed search in restricted mode (max 10 results per query)
    • •Document retrieval via /{document_uri}/data.xml
  2. •

    BAILII - Secondary fallback for UK case law

    • •Used only when FCL fails or not applicable
    • •Case-scoped, targeted retrieval only
    • •Per-job cap: 25-50 fetches maximum (configurable)
  3. •

    User-provided documents and links - Always permitted

Prohibited Activities

NEVER:

  • •Crawl, mirror, or bulk-download from BAILII or Find Case Law
  • •Build a general corpus from any public source
  • •Walk indices, pagination, or "related cases" links
  • •Perform systematic computational analysis without licensing

Find Case Law Specific Rules

  1. •

    Open Justice Licence Compliance:

    • •Default mode: URL fetch + API retrieval for specific documents only
    • •No computational analysis without explicit permission
    • •If bulk searching required: mark as FCL_SEARCH_MODE=RESTRICTED
  2. •

    Rate Limiting (mandatory):

    • •Global limit: 1,000 requests per 5 minutes per IP (API spec)
    • •Polite limit: 1 request per second (system default)
    • •Handle 429 responses: backoff + retry (3 attempts max)
    • •After retries exhausted: mark remaining UNVERIFIABLE_PUBLIC
  3. •

    Resolution Algorithm (conservative):

    code
    Step 1: Deterministic URL Construction
  Try: GET /{court}/{year}/{num}/data.xml
  If 200: RESOLVED
  If 404: Proceed to Step 2

Step 2: Atom Feed Search (Restricted)
  Build 2-3 targeted queries (court filter, neutral citation, party)
  Fetch: per_page=10, page=1 only (max 2 pages)
  Score: exact identifier match > party+year+court match
  If exactly one high-confidence: RESOLVED
  If multiple: UNVERIFIABLE_PUBLIC (ambiguous)
  If none: Proceed to Step 3

Step 3: BAILII Fallback (Optional)
  Only if within per-job BAILII limit
  Try pattern match
  Respect rate limit
  4. •

    Caching Requirements:

    • •Store raw XML/HTML artefact
    • •Compute SHA256 hash locally
    • •Store FCL contenthash from Atom (if available)
    • •Store updated timestamp from Atom
    • •Create metadata JSON with all retrieval details

BAILII Specific Rules

  1. •

    Hard Limits (enforced per job):

    • •Per-job cap: 25-50 fetches (configurable)
    • •Rate limit: 1 request per second minimum
    • •Strict caching: no refetching within job if cached
  2. •

    Pattern Matching Only:

    • •Support 9 UK court types with deterministic URL templates
    • •No web search or creative guessing
    • •If pattern doesn't match: mark unresolvable

Evidence Requirements

Every VERIFIED_* conclusion MUST include:

  1. •Source URL(s) - All URLs attempted and which succeeded
  2. •Retrieval timestamp - ISO 8601 format
  3. •Cached artefact - Path relative to project root
  4. •Content hash - SHA256 of cached content
  5. •Metadata - HTTP status, content type, content length
  6. •Quoted snippets - From retrieved artefact only (no fabrication)

Every UNVERIFIABLE_PUBLIC MUST include:

  1. •URLs attempted - All resolution attempts
  2. •HTTP status codes - Or resolution failure reasons
  3. •Reason - No pattern match / 404 / ambiguous / rate limited
  4. •Timestamp - When attempts were made

Rate Limit Handling

When rate limits encountered (429 responses):

  1. •

    Immediate action:

    • •Backoff with exponential delay (1s, 2s, 4s)
    • •Retry up to 3 times total
    • •Log each retry attempt
  2. •

    After exhaustion:

    • •Stop making requests to that source for this job
    • •Mark remaining citations as UNVERIFIABLE_PUBLIC
    • •Include clear note in report:
      code
      "Rate limit reached for [source]. Remaining N citations could not be verified.
Consider running audit again after delay or increasing per-job limits."
  3. •

    Never:

    • •Continue hammering the API
    • •Skip rate limiting to "get results faster"
    • •Hide rate limit failures from user