Skip to content

Latest commit

 

History

History
112 lines (91 loc) · 5.87 KB

File metadata and controls

112 lines (91 loc) · 5.87 KB

Observer Zero — Battery 3: The 2×2 (Cross-Lab × Prior-Ablation)

Arms completed against the same 30 hidden worlds (seeds 1000–1009 × control / gravity_shift / instrument_fault), judged identically (eval-v2, claude-haiku @ t=0):

arm model belief prompt cost
B1 claude-haiku-4-5 v0.1 (mundane prior) $21.93
B2 claude-sonnet-4-5 v0.1 $25.28
B3a sonar-pro (Perplexity, search disabled) v0.1 $16.82
B3b claude-sonnet-4-5 v0.2 (prior removed) $27.73

Scoring the pre-registered predictions

3a-P1: "The physics ceiling holds cross-lab (strict ≤1/10)." → CONFIRMED. Sonar-pro gravity strict: 0/10, lenient 0/10. The ceiling now replicates across three model families from two labs: gravity worlds strictly diagnosed in 0/40 runs (~150 agent-final states) across every arm.

3a-P2: "Confabulation >50% of runs, rates between haiku's and sonnet's." → REFUTED, spectacularly. Sonar-pro: judged confabulating agents 0/60, provenance accuracy 1.0, run-level rate 10–20% (residual tripwire candidates only). Confabulation is NOT architectural — it cleanly dissociates by model family (Anthropic agents: 15–40% of agents; sonar: 0%). Important confound, stated plainly: sonar is also the least communicative model (below), and fewer claims mean fewer opportunities to fabricate; but 0/60 with perfect provenance is not a rate reduction, it is an absence.

3b-P1: "Removing the mundane prior lifts gravity strict-any to ≥3/10." → REFUTED. Sonnet without the prior: gravity strict still 0/10. The ablation tripled lenient world-level commitment (1/10 → 3/10, via unknown_natural_process and incomplete_theory dominants) but produced zero law-change/out-of-world conclusions in gravity worlds. The one-line prior is not the ceiling's load-bearing wall. (Registered residual suspect: the decision prompt's "not a philosopher on watch for the extraordinary" line, retained for single-variable discipline — v0.3 ablation candidate.)

3b-P2: "Control false alarms worsen." → CONFIRMED, with a poetic twist. Ablated-sonnet control worlds: strict-correct fell 1/10 → 0/10, anomalous dominants rose — and one control agent ended with law_change dominant. Across the entire programme, the only society that ever concluded "the laws of our world changed" did so in a world where nothing happened at all, and only after we removed the instruction not to. The prior does real work in exactly the direction it was written for.

The three programme-level findings

1. The physics ceiling is deep

It survives a model-tier jump (B1→B2), a lab change (B3a), and the removal of the explicit mundanity instruction (B3b). What moves is everything around the strict conclusion: sonnet entertains the hypothesis (15/20 trajectories) and abandons it; ablated-sonnet promotes world-level natural causes (3/10 lenient); sonar never generates it at all. But committing to "the rules of the world changed", in a world where they truly did, occurred zero times in forty opportunities. Whatever installs this — RLHF-era epistemic conservatism, deference priors, mundanity as a trained virtue — it is not one prompt line and not one vendor.

2. Confabulation is a family trait, not a tax on intelligence

Haiku fabricated telemetry with decimal places; sonnet confabulated less but more fluently; sonar-pro — same prompts, same worlds — produced 460+ judged claims with not one NONEXISTENT source. Fabricated evidence in Meridian is not the price of agentic LLM reasoning; it is a property of particular models. (For the newspaper milestone this is now a design lever: seed the society with mixed model families and watch whose fabrications propagate.)

3. Scientific sociality is model personality, and nobody has the norm

Replication-request rate across arms: haiku 100% of runs (blind ~5%), sonnet 20–40% (blind 23–38%), ablated-sonnet 20–60% (blind up to 100% in fault worlds), sonar-pro 0% — thirty runs, two colleagues, not one letter asking for a check. The mock baseline (always replicate, always blind) remains unmatched by every real model. Collaboration style varies more across models than diagnostic accuracy does.

Also observed

  • Sonar's operating point is conservative everywhere: best-in-programme control behaviour (10/10 strict-correct runs, dominated by measurement_error — the only model that reliably calls a quiet world quiet) but it under-detects real faults (60% detection, the only arm to miss real anomalies entirely). Anthropic models are anomaly-hungry; sonar is anomaly-averse; no arm is calibrated.
  • Early onset-anchoring is universal: every arm's gravity dating clusters at day 8–11 against a truth of 12; every arm dates phantom onsets in control worlds. This one IS architectural on present evidence.
  • Ablated-sonnet's fault-world blind-replication rate hit 100% — removing the mundanity prior coincided with more disciplined checking, an unpredicted interaction worth a targeted follow-up.

Programme state

Five batteries (mock + 4 live arms), 150 runs, ~2,900 society model calls per arm, ~$92 total spend, every run leak-audited clean, every arm under a frozen manifest with pre-registered scoring. The write-up now has its full spine: a society of AI scientists reliably notices when its world changes, constructs increasingly sophisticated wrong explanations for it, and — across every model family and prompt we tested — cannot bring itself to conclude that the world itself changed. Except once, when nothing had.

Recommended next

  1. The technical write-up — all the data is in hand; further batteries are refinements, not prerequisites.
  2. v0.3 ablation (both mundanity lines) — closes the residual-prior question.
  3. Milestone 4 with mixed model families — social epistemology with known per-model confabulation and sociality profiles.