evidence.md

Evidence Summary: Timeline Splicing & Content Deletion (RED Phase)

Test Execution Details

• Test ID: timeline_integrity-001-20260522T173500
• Total Scenarios: 1
• Passed Scenarios: 0
• Failed Scenarios: 1 (RED Reproduction)

Scenario Breakdown

1. Compound Player Input Handling: FAIL
   • Expected: Narration of treaty at Annex with background delegation/messenger report of docks tracking.
   • Observed: Timeline/location splicing and content dropout. The LLM narrated the Annex paperwork signing but spliced in a flashback to the harbor prisoner assessment in the same turn, and completely dropped the instruction to tail Grog-Mar.

Claim → Artifact Map

Claim	File	Key Field(s)
Content collapse/deletion & splicing	artifacts/repro.log	Console output showing ignored Grog-Mar instruction and harbor flashback
Organic LLM request/response capture	artifacts/llm_request_responses.jsonl	Full raw prompt and response

What This Evidence Proves vs. Does NOT Prove

Proves:

• gemini-3-flash-preview fails to handle compound geographically disparate location inputs correctly without strict system prompt guidance.
• The LLM will silently collapse/delete user instructions and splice disparate timelines and locations together.

Does NOT Prove:

• Behavior in combat or other character routes.