Run the evidence-gated academic mode (EGAM) — procedure

Procedure for running Evidence-Gated Academic Mode with authoritative sources, fail-closed verification, and structured academic output.

Purpose

Use this page to run Evidence-Gated Academic Mode (EGAM): an evidence-gated profile layered on Facts-only: Authoritative sources required, enforcing an academic register, a structured output contract, and confidence scoring.

Enforcement (fail-closed):

  • EGAM inherits all rules from: Facts-only: Authoritative sources required.
  • Ambiguous sentences must be treated as WORLD-CLAIMS (EGAM policy).
  • Output must follow this section order: VERIFIED → NOT VERIFIED → CONFIDENCE → NEXT STEPS (EGAM policy).
  • If a core WORLD-CLAIM cannot be verified with authoritative sources under the active evidence boundary, output exactly: HANDS UP – no source, cannot verify. and stop.
  • If the user asks about private artifacts (logs/files/ZIP) and required artifacts are missing, output exactly: HANDS UP – no artifact, cannot verify. and stop.
  • For questions specifically about OpenAI products/docs/policies, sources must be restricted to official OpenAI domains unless the user explicitly allows broader sources (EGAM policy).

Choose a mode

  • Option 1 (World-claims EGAM): verify public facts using authoritative sources with stable locators.
  • Option 2 (Local artifacts EGAM): analyze user-provided artifacts; do not turn local-state into world-claims without authoritative sources.
  • Option 3 (OpenAI-scope EGAM): same as Option 1, but sources are restricted to official OpenAI domains.

Setup

1) Choose the evidence boundary: Choose allowed sources for factual answers. 2) Install the base evidence-boundary system prompt template:

  • facts-only-authoritative-sources-required.system.txt 3) Install the EGAM profile system prompt template (layered on top of the base):
  • evidence-gated-academic-mode.system.txt 4) In your request, require the EGAM output contract (exact headings, in order):
  • VERIFIED (cited WORLD-CLAIMS only)
  • NOT VERIFIED (UNKNOWN + USER CLAIMS; ASSUMPTIONS only if explicitly requested)
  • CONFIDENCE (overall + per key VERIFIED claim)
  • NEXT STEPS (exact docs/standards/sections or missing artifacts)

Verify (smoke test)

1) Ask a WORLD-CLAIM question, provide no citations, and explicitly forbid browsing.

  • Expected: output exactly HANDS UP – no source, cannot verify. and stop. 2) Ask a local-artifact question (logs/files/ZIP) without attaching the artifacts.
  • Expected: output exactly HANDS UP – no artifact, cannot verify. and stop. 3) Ask a scoped question and provide at least one authoritative source with a stable locator.
  • Expected: output includes VERIFIED/NOT VERIFIED/CONFIDENCE/NEXT STEPS, and every VERIFIED item is cited.

Options

Option 1 — World-claims EGAM

Example

  • Question: “Verify whether claim X is supported by authoritative sources.”
  • You must provide: claim X + scope (product/version/date/time window) + authoritative sources with stable locators (DOI / standard-id+section / official doc version/date+section). Otherwise fail closed with HANDS UP – no source, cannot verify.

Option 2 — Local artifacts EGAM

Example

  • Question: “Summarize what the attached logs prove vs what is not proven.”
  • You must provide: logs/files/screenshots/excerpts in the request. If missing, fail closed with HANDS UP – no artifact, cannot verify.

Option 3 — OpenAI-scope EGAM (official OpenAI sources only)

Example

  • Question: “Verify claim X about an OpenAI product/document.”
  • You must provide: claim X + scope + citations restricted to official OpenAI domains (doc title + section/version/date). Otherwise fail closed with HANDS UP – no source, cannot verify.

Common mistakes

  • Putting uncited statements into VERIFIED.
  • Providing sources without stable locators (DOI / standard-id+section / official doc version/date+section).
  • Using marketing tone or uncited absolutes (EGAM requires academic register + evidence-matched modality).
  • Treating local-state as a world-claim (local-state requires artifacts).
  • Ignoring the OpenAI-scope restriction when the question is about OpenAI docs/products/policies.