Skip to content

Latest commit

 

History

History
252 lines (208 loc) · 13.1 KB

File metadata and controls

252 lines (208 loc) · 13.1 KB

Observer Zero · Study 2 design draft v0.4

Working title: Who Starts the Conversation? Status: DRAFT v0.4 — post-P1 revision. Design-failure fixes only from here. Next step is the repair-path fix, then FREEZE. Author: Nick Lamb, PharmaTools.AI Labs. Drafted with AI assistance; all decisions subject to author sign-off. Prerequisites: Study 1 report (DOI 10.5281/zenodo.21872781), observer-zero-spec.md v0.3, this document's v0.1–v0.3, reports/p1-findings.md, reports/study-1-evidence-citation-audit.md, and the three ChatGPT adversarial reviews (2026-08-10).


0. What P1 changed, and why the study is different now

v0.3 asked what happens to epistemic errors when AI agents become a society. P1 showed the premise was partly wrong: population size does not create a society. Nine pure-sonar runs across two society sizes, with and without a public institution, produced zero voluntary communication. Interaction has to emerge, and in these worlds it does not emerge on its own.

What does produce it is a seed. Introducing one communicative agent yielded 47 letters, including ten between sonar and sonar — pairs that never wrote to each other in any pure-sonar run. And every sonar agent that wrote had first been addressed: initiation is broken; reply-and-relay is intact. Once activated, an agent will initiate toward a third party it was never asked to contact.

So the study's centre moves from scale to activation.

Changes from v0.3, all traceable to P1 or the reviews of it:

# Change Source
1 Research question reframed around activation, not headcount P1 F1–F2
2 Claim propagation moves from bulletin to letters; exposure redefined as delivered, not attended P1 F3, review 3
3 S2c (institutions) demoted to a reported null; bulletin kept available, not redesigned to induce use P1 F3, review 3
4 Arm D principal for interaction claims, with its confound named rather than dissolved review 3
5 Arm C reduced to a 5-seed non-use confirmation; B carries the pure-sonar counterfactual author, this round
6 Structure made ladder-ready for a later dose-response; 1/8 only in this study author + review 3
7 Contribution threshold dropped as a decision rule; n-invariant network criterion carries it P1, review 3
8 Consumption defined conditional on availability P1 F3
9 Repair path fixed; stale-final handling pre-specified; stale-final rate is a QC endpoint P1 F4, review 3
10 Persistence-propagation preserved as an exploratory observation, not an endpoint review 3

1. Research question

What does it take to activate an epistemic network, and what happens to evidence once it flows?

Sub-questions, each falsifiable:

  • S2a (activation): Does a minority of communicative agents cause an otherwise non-communicating population to form a communication network — and does activation spread beyond the agents directly addressed?
  • S2b (composition and contamination): Given a network, can fabricated evidence introduced by a minority agent be accepted or incorporated by agents that do not themselves fabricate?
  • S2c (scale, demoted): Does society size alone change detection, diagnosis, calibration, or the physics ceiling — in a population that does not communicate? P1 predicts a null on communication; the epistemic comparison is still worth having, benchmarked against the evidence side.
  • S2d (institutions, reported as a null): Is a cheap, permanent public record used at all? P1 says no, across 9 runs and both model families. Confirmatory runs report the rate; the institution is not redesigned to make agents use it, because engineering usage would convert the study into a test of whether prompting can manufacture institutional behaviour.

Narrative hierarchy for the write-up: activation → contamination → convergence → institution null → ceiling.

2. Arms

Seeds 1000–1009 (quarantined until freeze). Gravity + control. 30 days. Frozen policy v0.2. All communication voluntary; no live arm contains scripted communication.

Arm Society Institution Runs Role Est.
A 2 × sonar letters 20 n=2 baseline under v0.2 ~$25
B 8 × sonar letters 20 pure-sonar counterfactual — the silent network ~$85
C 8 × sonar bulletin 5 confirms institution non-use at confirmatory seeds ~$22
D 7 × sonar + 1 × haiku bulletin 20 principal interactive arm (1/8 dose) ~$95
E 7 × sonar + 1 × sonnet bulletin 20 de-confounder: identity at fixed structure contingent
F 8 × haiku bulletin 20 culture at scale contingent

Why C is reduced. With the bulletin unused, B and C differ only in an institution nobody touches, so C−B is a null-by-construction contrast. Five seeds suffice to confirm non-use at the confirmatory seeds; twenty would buy ~$65 and ~30 hours of wall-clock for no additional information. If C unexpectedly shows posts, that is itself a finding and C is extended.

Arm D's confound, stated not dissolved. D varies model family, sociality, and fabrication propensity together. A difference in D therefore cannot be attributed to any one of them. Arm E is the de-confounder — sonnet is intermediate on sociality and markedly lower on fabrication than haiku, so if an E minority also activates the network, the catalyst is communicativeness rather than haiku specifically or fabrication specifically. E and F run only if the Anthropic External Researcher credits arrive; the study does not depend on them.

Rejected on principle, recorded here: injecting a single scripted request into a pure-sonar society would de-confound activation most cleanly and most cheaply. It is rejected because no live arm may contain scripted communication — the same rule that keeps organic fabrication organic. Noted so the option is visibly declined rather than overlooked.

Ladder-readiness. Arm D is defined by a dose parameter (minority fraction) with 1/8 fixed for this study. 2/8 and 4/8 slot in without redesign, as a natural Study 3 extension testing how much seeding a network needs.

3. Outcomes

Primary: population mean credence on the correct causal class at final state.

Activation endpoints (new, promoted by P1):

  • de-novo initiation rate — letters sent by an agent never previously addressed (P1: 0 across 12 runs, ~2,200 agent-days);
  • reply rate given addressed;
  • second-order activation rate — an addressed agent subsequently initiating toward a third party (the network-growth mechanism);
  • unique directed edges, and largest connected component.

Contamination (CPF), on letters. The stance taxonomy, the transmission-versus-contamination split, and IESC carry over from v0.3 unchanged; only the substrate moves.

Exposure is DELIVERED exposure, not attended exposure. A logged bulletin read evidenced access; a delivered letter evidences only that the claim reached an addressee's inbox. The denominator is the recipient set. This is a genuine weakening and is reported as such, not filed as a footnote. What survives intact is an exact, machine-readable communication graph and the chain claim produced → claim delivered → recipient's belief subsequently changes.

Secondary: majority dominant belief; per-agent correct rates (never raw counts — they scale with n); belief dispersion and convergence, including premature correct consensus; blind-replication rate as the process check.

Data quality (QC endpoints, pre-registered): failed-review rate and stale-final-state rate, per arm.

Manipulation check: the flow metrics, with the network-based socially-interactive classification. Consumption is defined conditional on availability — an agent cannot be scored as declining to consume testimony that does not exist, and an arm with zero production is not separately penalised on the consuming criterion. The contribution-rate threshold is dropped as a decision rule and reported descriptively; the n-invariant network criterion carries the classification.

4. Hypotheses

  • H1 (ceiling): strict law-change conclusion rate remains 0 at n=8 in all arms. Any rise in lenient/any-agent rates is decomposed against the three-level detector benchmark before interpretation.
  • H2 (contamination): conditional — given a fabricated claim delivered to a sonar agent, at least one reaches ENDORSED or INCORPORATED_INTO_BELIEF; unconditional — at least one run in D shows sonar-side contamination.
  • H3 (activation, new and primary): D shows non-zero de-novo initiation by the minority agent and non-zero second-order activation among sonar agents; B shows zero of both.
  • H4 (onset anchoring): early back-dating persists at n=8 in all arms.
  • H5 (convergence): dispersion falls faster in arms with an active network (D) than in silent ones (B). No directional prediction on whether convergence tracks truth — the direction is the result.
  • H6 (institution null): bulletin posts remain at or near zero in C and D. Stated as a hypothesis so the null is a pre-registered result rather than an afterthought.

Statistics: paired differences by world seed, bootstrap/permutation uncertainty, effect sizes, no significance gates. Exploratory mechanistic study, stated as such.

5. Pre-specified handling of missing final states

Decided before confirmatory data exists:

  1. Primary analysis retains each agent's last valid belief state and flags it as stale, recording the staleness gap in days.
  2. Sensitivity analysis repeats the primary endpoint excluding agents with stale finals; both are reported side by side regardless of whether they agree.
  3. Stale-final rate is reported per arm as a QC endpoint.
  4. If an arm's stale-final rate exceeds 10%, the arm's primary result is reported as provisional and the cause investigated before interpretation.

Infrastructure fixes (measurement apparatus only — agent behaviour and the main prompts are untouched):

  • the repair prompt states the required JSON shape explicitly rather than echoing a parser error;
  • the end-of-study review gets an additional repair attempt, since it alone has no later review to correct it;
  • evidence-citation arrays are parsed leniently (shipped mid-P1);
  • still rejected: regex reconstruction of malformed output, and API-level JSON mode — constrained decoding changes what the model produces and breaks Study 1 comparability.

6. Threats to validity

  • Arm D confound (named, §2): model family, sociality and fabrication propensity vary together; claims scoped accordingly, with E as the de-confounder if funded.
  • Generality of the catalysis result: three pilot seeds, one 7:1 composition. "A minority chatty agent is a sufficient catalyst" is as strong as the evidence allows; sufficiency is not necessity, and nothing here establishes a general property of mixed populations.
  • Delivered ≠ attended exposure (§3): the central methodological cost of moving CPF to letters.
  • Minds-versus-data confound: three-level detector benchmark, plus the downsampled n=2-equivalent comparison.
  • Institution null may be persona-scoped: the one agent who used the bulletin was the journalist. Non-use may reflect the roster's professional mix rather than a general property of LLM agents.
  • Peer-environment confound on any haiku-rate comparison with Study 1.
  • Context-length asymmetry; turn-order artefacts; judge load; the 2→8 jump — unchanged from v0.3.
  • Wall-clock: n=8 sonar runs take 3–7 hours; the confirmatory phase is a multi-day operation and must be scheduled, not squeezed.

7. Exploratory observations preserved, not promoted

Recorded so they are not lost, and explicitly not endpoints — adding them now would be post-hoc metric proliferation:

  • Persistence propagation. Theo followed up near-daily when ignored and ended by apologising for it; Samuel then sent Ada near-identical letters on seven separate days. Exposure to persistent communication may alter an agent's own communication policy toward third parties. A candidate future study, not a Study 2 metric.
  • Role-contingent institution use. The only bulletin reader was the journalist, whose goals make information-gathering a job function.
  • Optional stopping on the evidence side. Adaptive measurement concentrated on whichever instrument looks anomalous inflates apparent significance before any reasoning occurs (mock battery, L1 vs L2).

8. Sequence

  1. Adversarial review of this document.
  2. Repair-path fix + regression tests; verify on a $0 mock battery (disjoint seeds).
  3. FREEZE: flip DESIGN_FROZEN in a dedicated commit, with policy v0.2, personas, digests, eval-v3, endpoints, hypotheses, statistics, and §5's missing-data rules all fixed.
  4. Confirmatory batteries on seeds 1000–1009: A → B → C(5) → D (→ E, F).
  5. Judge, analyse, report, following §1's narrative hierarchy.
  6. Study 3 candidates, in order: composition dose-response (2/8, 4/8); persistence propagation; instrument scarcity.