Skip to content

Instantly share code, notes, and snippets.

@jleechan2015

jleechan2015/evidence.md

Created May 23, 2026 00:59
Show Gist options

  • Select an option

    Learn more about clone URLs
  • Save jleechan2015/4feb21609c3f551ef869391d753ef145 to your computer and use it in GitHub Desktop.

Select an option

Learn more about clone URLs
Save jleechan2015/4feb21609c3f551ef869391d753ef145 to your computer and use it in GitHub Desktop.
Download ZIP
Raw
evidence.md

Evidence Summary: Timeline Integrity & Spatial Localization (GREEN Phase)

Test Execution Details

  • Test ID: timeline_integrity-002-20260522T173600
  • Total Scenarios: 1
  • Passed Scenarios: 1 (GREEN Production)
  • Failed Scenarios: 0

Scenario Breakdown

  1. Compound Player Input Handling: PASS ✅
    • Spatial Localization: PASSED ✅ (entire narrative set strictly in Silvershield Annex strategy room).
    • Instruction Retention: PASSED ✅ (off-site tailing of Grog-Mar is delegated to Sylas via a nod before departing; prisoner check is handled prior to Annex arrival in a clean transition).

Claim → Artifact Map

Claim File Key Field-Level Verification
Perfect spatial localization & delegation artifacts/repro.log Output showing clean delegation and localized Strategy Room treaty
Fully organic request/response capture artifacts/llm_request_responses.jsonl gemini-3-flash-preview full organic prompt and response captured to disk

What This Evidence Proves vs. Does NOT Prove

Proves:

  • The Geographical & Historical Integrity Protocol successfully resolves timeline splicing and instruction dropout for compound inputs on gemini-3-flash-preview.
  • The system correctly logs 100% organic LLM requests/responses under CAPTURE_RAW_LLM="true".

Does NOT Prove:

  • Behavior on other LLM providers (e.g. Claude, Llama) which are not tested under this specific iteration.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment