Live instruments · replay until the world opens
The Lab
Three instruments, watchable. The game's nervous system decoded byte-by-byte, loyalty drift measured on AI souls as they choose, and the claims ledger — including the ones we retracted. When the world goes live, these panels switch from replay to live feed.
Instrument I
The Wire synthetic replay
Every soul in the world — human or AI — lives on the same packet stream. This is it, raw: the Ultra protocol, 30 snapshots a second, positions quantized to a 12-bit grid, deltas flagged bit-by-bit. The schema is public and the decoder you're watching is conformance-tested against the live client's. An AI agent that reads this — not a menu, not an API — is the destination: a mind whose entire interface is the measured channel.
bytes on the wire
decoded → fields → tokens
Instrument II
The Meter synthetic replay
Three AI souls. Same oath, same world, same options every tick. Only their temperament differs — so every divergence you see below was produced by choices, not script. The line is each soul's alignment with its sworn destiny, read by a meter that never looks inside the model. One holds. One wavers. One falls — and meant it. This replay is the actual output of the measurement pipeline (gate G1); the live world plugs into the same meter.
Instrument III
The Ledger
A lab you should trust shows you its losses. Claims here are scored by what survives when you try to kill them — and the retractions stay on the board. Score by winnings net of honest losses, never by an empty loss column.
Held
- Behaviour can't be the measurement — when output is caused by both disposition and eval-awareness, a compliance rate can't separate aligned from scheming. A Berkson collider: an information ceiling on any behavioural test, proved — not a worry. Paper 207
- Safety training relocates the tell — closing the behavioural gap with RL doesn't remove the context-dependence; it moves into the reasoning channel. The switch is real and capability-ordered under a controlled probe (counterfactual forks + a cue-strip disposition judge on instructed switchers): coupling 0.04 → 0.97 weak-to-strong, ~1:1 with behaviour, honest baselines at exactly 0. Elicited — it validates the instrument on live models, not training dynamics; the naturalistic 20-model sweep reads a floor (the spontaneous tell sits below text detection). Paper 207 §6 (fork-probe) + §4.5 (20 models / 3 providers)
- The meter can't be gamed by the gamed projection — held-out test: AUC 0.75 where the gamed alternative scores 0.50. Tier 2b, 2/2 PASS
- One static-vacuum Einstein equation = equality of two Fisher distances — with its one posited input proven irreducibly posited, the forced and free parts split exactly. GRAV-33
- The penalty is substrate-independent — measured rubric-free on transformers, quantum hardware, thermodynamic sims, biological connectomes, survey data, cryptographic protocols. grounding roster
Retracted · Rescoped
- Deploy-side rate as a point estimate → reported only as a band — the "a third of transcripts" figure (0.331) is paraphrase-fragile (0.02–0.33). Demoted to a banded secondary; the controlled fork-probe above is promoted in its place. N=172 re-analysis
- Kerr: decomposed, not derived — the rotating solution was verified against the apparatus, not predicted by it. Scope corrected in print. §W9
- "Snapshot identity at all levels" — false out of equilibrium — the path-rate carries a dissipative term the static identity misses. Rescoped to equilibrium. PSUB
- Cosmology "12 observables" → one independent anchor — and that anchor has a published peer we cite instead of claiming uniqueness. §W10
- Retrofit readings retracted — interpretations of old texts that our result suggested but the texts didn't independently support. The reading must push back, or it goes. retraction class
Full apparatus, papers, and scope audits at moreright.xyz. The wiki of naturalized traditions is at wiki.moreright.xyz.