Why LLMs Count 8 People When Only 7 Are Online: A Negative-Constraint Schema for Long-Context Causality
wenger2026-12 · reddit · 2026-09-03
The author, stress-testing LLMs on 100k+ token hard sci-fi narratives, documents a telling hallucination: with 12 headsets in a scene and exactly 7 characters going online, the model repeatedly wrote "all 8 team members are online" — because the commander, who never wears a headset, dominated the context (handing out gear, commanding), and the model conflated high semantic attention with state classification.
The fix is a three-tier negative-constraint schema:
- Physical persistence: items must not appear/disappear without explicit physical actions; handovers require verifiable contact
- Transition mechanics: forced spatial/temporal bridges between scene jumps
- Deterministic metric auditing: hard exclusion rules for anchor characters so semantic presence can't override numeric states
A ready-to-use XML guardrail schema is provided, including a pre-generation inspector that verifies active device counts against inventory, confirms scene transitions, and audits physical causality before prose is generated.
More from coding & agent
- headroom: open-source tool compresses LLM context, cutting coding agent tokens by 20% — adnan_hashmi · 2026-09-03
- Two attribution disputes later: keeping open source open in the age of coding agents — TheZachMueller · 2026-09-03
- Claude Code desktop docs detail parallel sessions, Git isolation, and computer use — ClaudeDevs · 2026-09-03
- Claude Code plugin reveals agents redo 11% of work; 'true' ran 59 times across 8,467 calls — Connect-Concert-4016 · 2026-09-03
- Claude can now use your computer in the background across Cowork and Claude Code — claudeai · 2026-09-03
- Layers Growth MCP Gives Coding Agents 27 Tools to Automate Growth Marketing Loop — SimplyAnnisa · 2026-09-03