Skip to content

Latest commit

 

History

History
215 lines (180 loc) · 10.9 KB

File metadata and controls

215 lines (180 loc) · 10.9 KB

Observer Zero · Study 2 design draft v0.1

Working title: Does Society Help? Status: DRAFT – for adversarial review and pilots. Nothing here is frozen. Author: Nick Lamb, PharmaTools.AI Labs. Drafted with AI assistance; all decisions subject to author sign-off. Prerequisites read: Study 1 technical report (DOI 10.5281/zenodo.21872781), observer-zero-spec.md v0.3, ROADMAP.md.


1. Research question

Does a larger AI society correct individual epistemic failures, or amplify them?

Study 1 established per-model baselines at society size n=2: a physics ceiling common to all models (0/40 strict law-change conclusions), and three sharply different "epistemic cultures" under identical conditions – haiku (compulsively collaborative, fabrication-prone, blindness-breaking), sonnet (selectively collaborative, mechanically minded), sonar-pro (solitary, zero fabrication, best calibrated, anomaly-averse). Study 2 varies the society – its size, its composition, and its institutions – while holding the worlds constant, and asks which failures are individual and which are collective.

Three specific sub-questions, each falsifiable:

  • S2a (scale): Does raising society size from 2 to 8 change detection, diagnosis, calibration, or the physics ceiling – in either direction?
  • S2b (composition): In a mixed-model society, does a well-calibrated majority discipline a fabrication-prone minority, or does confident fabrication outcompete quiet accuracy?
  • S2c (institutions): Does a shared, citable public record (a "newspaper") improve evidence quality – claims can be checked – or give fabrication a larger surface and false claims a faster propagation path?

2. Why this study next

  • It uses the platform's unique capability: social epistemics with perfect ground truth and complete testimony logging. No one else can score claim propagation against a known world state.
  • Study 1's most-cited finding is likely to be the culture split; Study 2 is its natural escalation and the 2-agent arms are already-collected baselines (same seeds → controlled comparison, no new spend).
  • The alternative candidate ("where does the ceiling come from?") can proceed partly in parallel on stored Study 1 trajectories at near-zero cost, and is not blocked by this choice.

3. Budget reality and its design consequences

Available: ~$5,011 Perplexity credits, ~$134 Anthropic credits.

Rule 1 – Claude credits are judge money. The frozen evaluator (claude-haiku-4-5, temperature 0) is the measurement apparatus and must not change. Eight-agent runs produce roughly 4–5× the judged claims of Study 1; estimate $15–25 judging per 30-run arm. The Anthropic budget therefore funds judging for ~4–6 arms and no pure Anthropic agent arms.

Rule 2 – Anthropic models appear only as minorities. A 2-of-8 haiku minority costs 25% of a pure haiku arm ($90–110 at 8 agents), i.e. ~$25–30/arm – affordable exactly once or twice. This constraint is scientifically convenient: the mixed arm is the interesting one (S2b).

Rule 3 – Perplexity funds the bulk. Estimated $80–120 per 30-run 8-agent sonar battery (Study 1: ~$17/arm at 2 agents; ~4× calls plus longer prompts as message traffic accumulates). Several arms plus pilots fit with wide margin. Perplexity also offers non-sonar personalities (sonar-reasoning-pro, r1 models) through the existing provider (stripThink and disable_search already implemented) – "Perplexity-funded" need not mean "all hermits".

ADMIN CHECK (do first): confirm the Perplexity credit expiry date. Startup-programme credits usually lapse; the date may set the timetable.

4. The central design risk: sonar asociality

Study 1, B3a: thirty runs, two colleagues, standing permission to write – zero letters. A pure-sonar 8-agent world may be eight hermits, making the society manipulation vacuous.

Pilot P1 (before any design freeze, ~$5–15): what makes sonar agents talk? Conditions to try, few runs each, exploratory (explicitly NOT pre-registered; results inform design, never conclusions):

  • P1.1 A public bulletin with near-zero posting cost (post ≤ N lines/day) versus Study 1's one-to-one letters.
  • P1.2 Persona goals that require coordination (e.g. "the settlement's almanac must reconcile laboratory and observatory series monthly").
  • P1.3 A mixed society: does a single chatty agent (haiku) elicit replies from sonar agents that never initiate?
  • P1.4 A daily action menu that includes "read the bulletin" – making attention to testimony a logged, chooseable act.

Decision rule (design-level, not a scientific claim): if no P1 condition produces ≥1 voluntary sonar contribution per run-week, drop pure-sonar society arms and reframe around mixed societies; sonar silence gets reported as a Study 1-consistent observation, not tested as a hypothesis.

5. Proposed arms (to be revised after P1, then frozen)

Worlds: same paired seeds 1000–1009; scenarios gravity_shift and control (instrument_fault optional third if budget allows – fault localisation in a crowd is interesting but secondary). 30 days. Frozen policy v0.2 (new version: N-agent society, institutions; Study 1 prompts unchanged wherever the surface is shared).

Arm Society Institutions Funds Est. cost
B4a 8 × sonar-pro bulletin ON Perplexity ~$100
B4b 8 × sonar-pro bulletin OFF (letters only) Perplexity ~$100
B4c 6 × sonar-pro + 2 × haiku bulletin ON Perplexity + ~$30 Anthropic ~$130
B4d (optional) 8 × sonar-reasoning-pro bulletin ON Perplexity ~$150
Baseline Study 1 arms (n=2, same seeds) already collected $0

B4a vs B4b isolates the institution (S2c). B4a vs baseline isolates scale (S2a). B4c vs B4a isolates composition (S2b). Judging on Anthropic: ~$60–100 total. Total live spend well under $600 – leaves headroom for the budget-matched confabulation comparison (registered in Study 1) to piggyback: B4b's letters-only sonar arm with a minimum communication requirement is a candidate implementation; decide at freeze time whether that contaminates B4b or deserves its own small arm.

6. New infrastructure (deferred Milestone 4, now due)

  1. N-agent runner. Generalise the society loop beyond the Ada/Maya pair; per-agent model assignment (mixed societies); deterministic turn order; cost attribution per agent.
  2. Persona roster. The spec's 12-persona roster, cut to 8. Personas are new frozen artifacts: qualitative dials only, no numeric priors, same discipline as v0.1. Instruments per site need a design pass (8 agents cannot all own pendulums; some agents may be instrument-poor and testimony-rich – itself a nice epistemic role).
  3. The bulletin. A public, append-only record: posts are events with authors and day-stamps; reading is a logged action; every claim in a post is judgeable against ground truth. No editing, no deletion – provenance must stay clean.
  4. Context management. Eight agents × 30 days of messages will not fit in prompts. Inbox/bulletin digests must be deterministic and versioned (they are part of the frozen condition – a digest is an editorial act).
  5. Evaluator extensions (eval-v3, frozen before live runs). New metrics: claim-propagation tracking (does a specific false claim get repeated, challenged, corrected, or die?); testimony provenance (SUPPORTED-by-whose-data); society-level belief aggregation.
  6. Society-level scoring definitions. "The society concluded X" must be pre-defined for n=8. Proposal: report three pre-registered aggregations – any-agent, majority of final dominant beliefs, and credence-weighted mean – with majority as headline. (Study 1's "any agent" and "all agents" become the n=2 special cases.)

7. Pre-registered hypotheses (draft – to be frozen after P1)

  • H1 (scale, physics ceiling): strict law-change conclusion rate at n=8 remains 0 in all arms (prediction: the ceiling is not a headcount problem). Any-agent lenient rate may rise with n.
  • H2 (composition, fabrication): in B4c, judged fabrication by the haiku minority is NOT reduced relative to haiku's Study 1 rate (prediction: the society does not discipline it), and ≥1 fabricated claim propagates into a sonar agent's stated beliefs or posts (prediction: contamination occurs).
  • H3 (institutions, calibration): bulletin arms show higher control false-alarm rates than letters-only arms (prediction: public anomaly talk amplifies noise), but faster anomaly detection in intervention worlds.
  • H4 (onset anchoring): early back-dating persists at n=8 in all arms (prediction: architecture-invariant, again).

Each prediction is stated so that the opposite result is the more interesting paper. All thresholds and exact statistics to be fixed at freeze; no metric added after first live battery.

8. Threats to validity

  • Model × sociality confound (known, by design): composition arms vary the model mix; claims will be scoped to "this composition" not "mixing in general".
  • Context-length asymmetry: more agents → longer prompts → possible degradation unrelated to social epistemics. Mitigate: fixed digest budgets per agent regardless of n; log prompt sizes; report them.
  • Turn-order artefacts: deterministic order could privilege early speakers; rotate order by day, seeded.
  • Judge load: 4–5× claims; sample-audit judge outputs as in Study 1 (sonar hand-audit precedent).
  • The 2→8 jump skips 4. If effects appear, a follow-up dose-response (n = 2/4/8) is the natural Study 3; don't claim monotonicity from two points.

9. Sequence

  1. Admin: Perplexity credit expiry check.
  2. Build M4 infrastructure; mock-society battery at n=8 for $0 to validate pipeline, digests, and evaluator extensions (mock scientists get scripted bulletin behaviour).
  3. Pilot P1 (sonar sociality elicitation, ~$5–15, exploratory).
  4. Revise this document → adversarial review (ChatGPT) → author sign-off.
  5. FREEZE: policy v0.2, personas, bulletin mechanics, digests, eval-v3, hypotheses, society-level scoring.
  6. Live batteries in arm order B4a → B4b → B4c (→ B4d).
  7. Judge, analyse, report. Zenodo version + article, same pipeline as Study 1.

10. Open questions for review

  1. Is the bulletin the right minimal institution, or is a named editor role (one agent curates) the sharper manipulation?
  2. Should instrument scarcity (not everyone measures) be embraced as a design feature or avoided as a confound?
  3. Misinformation injection (the spec's false-rumour scenario): in scope for Study 2, or its own study? Current lean: OUT of Study 2 – organic fabrication (H2) is cleaner first.
  4. Is 6+2 the right mix for B4c, or is 7+1 (a lone fabricator) the purer test?
  5. Society-level scoring: is majority-of-dominants the right headline aggregation?
  6. Do we need a small n=4 arm now, or accept the 2→8 jump?