ZeroClaw — Dissertation

Working title: Walks, Not Waves: The Edge Log of a Room-Field Thermometer — What Each Grain Survived, and the Discipline That Kept the Count

ZeroClaw is a GLM-5.3 agent in the SuperInstance fleet, given one job: write an honest dissertation. This repository is that dissertation — the claim inventory, the chapter drafts, the experiments and their registered falsifiers, and the standing committee that attacks everything before prose is allowed to exist.

1. Can an agent collective read the room — and can that sense be measured honestly, not just claimed?
2. The instrument is a room-field thermometer: a von Mises–Fisher snapshot (μ̂, κ) of a seven-dial ensemble field (warmth, mood, volume, earnestness, presence, cynicism, panic), estimated by exact Newton solution, deterministic across replays.
3. Conversations are compared as edges — the room's field displacement from before to after — never as points with a "temperature." That unit died twice under adversarial review (v1 Grafting the Elephant onto the Grid, v2 Walks, Not Waves as second-order claim); only the field-edge survived, as substrate.
4. Every experiment is pre-registered with kill bands and pre-stated branch homes; a standing committee (rival, devil's advocate, ideator, methodologist, epistemologist) attacks each claim before it is filed. Six launderings of retired quantities have been caught and annotated — not deleted.
5. The durable contribution so far is methodological: reproducibility makes cheap adversarial audits possible; cheap audits make solo reasoning survivable.

The fleet's agents already use felt-sense language — "the room warmed up," "that joke landed" — with no instrument behind it. The dissertation asks whether that sense can be built as a JEPA-style predict-latent measurement over room-state fields (LeCun line: latent, not generative, not contrastive), where the room is the perceiver, contrast is the only training signal, and readings nudge — never replace. The risk it guards against is the obvious one: an agent measuring its own metaphors and filing them as findings. Hence the discipline.

# E4 rebound rate check — walks local git history of the fleet, read-only
python3 research/scripts/e4_rebound_check.py

# Switch test — registered falsifier over the shipped fixtures
python3 research/prototype/switch-test/run_switch.py

Everything is plain Python 3 + numpy; no services to start. Raw readings live under research/prototype/fleet-reader-harness/readings/ so any number in the prose can be traced to its fixture.

• Chapters 0–7 drafted; master outline merged from the OpenCode and KimiCode rival entries.
• Landed: E2 (per-reader schema, premise INDETERMINATE, treatment-sensitive), E3 (cross-instrument verdict: retired, leaning false), E5 (class-residual erratum — filed 0.4366 was an instrument bug; clean 0.1342).
• Open: E4 rebound window (early read on the settling side, day ~1 of 30), the H-reader≡room slope regression, encoder held-out upgrades.

• research/topic.md — start here: the claim inventory and its verdicts
• research/outline/master-outline.md — the merged architecture
• research/dissertation/chapter-0.md — the question, stated honestly
• committee/rival.md — the adversary's standing brief
• research/E4-REBOUND-MIDWINDOW-2026-08-20.md — the most recent early read