Topology Repository

This repo contains three completed experiment phases. The front-door claim is:

On a held-out procedural micro-world task, Unknown / non-entailment is recoverable from verdict-region hidden states, while decoder outputs under-express it.

The old “global topology of truth” headline did not survive.

Start Here

• Paper source: paper/main.tex
• Paper PDF: paper/paper.pdf
• Claim-to-file mapping: CLAIMS_AND_EVIDENCE.md
• Repository file map: REPO_MAP.md
• Experiment index: docs/experiments/README.md

Three Phases

1. Phase A (negative): global H0/H1 topology on GSM8K traces
   docs/experiments/01_global_h0_topology/README.md

2. Phase B (positive): fixed-decoding convergence failure on GSM8K
   docs/experiments/02_convergence_failure/README.md

3. Phase C (main positive): procedural micro-world representation–decoder dissociation
   docs/experiments/03_micro_world_semantics/README.md

Main Numbers

• Phase A: corrected topology-only AUC = 0.20 (controls = 0.60)
• Phase B: cap→wrong AUC = 0.839; OR ≈ 26.40
• Phase C: decoder Unknown recall (Qwen2B=0.000, Qwen4B=0.000, Gemma-it=0.0125) while verdict-token probe Unknown recall (0.7375, 0.1292, 0.5625)
• Phase C add-ons:
  • latent pre-readout steering (Gemma-it raw) improves macro-F1 0.339→0.395 with Unknown recall 0.375→0.363
  • shallow MLP sensitivity shows strong hidden Unknown signal on Qwen4B no-think (verdict_token recall 0.129→0.638)

Repo Notes

• Legacy root Python scripts were moved to phase_a_global_h0_topology/scripts/.
• All committed artifacts are kept under artifacts/ and outputs/.
• .npz artifacts are committed (not ignored).

Reproducibility

• Locked deps: repro/requirements.lock.txt
• Rebuild guide: repro/README.md
• CI build/release: .github/workflows/paper.yml