Skip to content

Latest commit

 

History

History
316 lines (267 loc) · 17.4 KB

File metadata and controls

316 lines (267 loc) · 17.4 KB

Observer Zero — current status and handover

Updated: 2026-09-02 — Study 3: design frozen, confirmatory battery executed, scoring in progress, analysis not yet run Purpose: one page that lets a fresh session (or a returning human) pick up without rereading the whole programme. Point-in-time; supersede freely. Sections below "Study 3 — where it stands" predate it — read them as history.


Where the programme is

Study 1: published. Cite concept DOI 10.5281/zenodo.21872780, which resolves to the corrected v2; v1 (…21872781) contains a sample-scope error corrected by an erratum. Repo github.com/nickjlamb/observer-zero, Medium article live at AI Advances.

Study 2: complete. All 85 confirmatory runs executed under the frozen design (v0.6 + amendments A2–A5, tag study2-freeze), analysed clean, and written up. Raw run data deposited on Zenodo (concept DOI 10.5281/zenodo.21909254).

Combined paper: under peer review. Submitted August 2026; per repo policy (daff318) the venue is not named in public docs until there is a decision. Preprint concept DOI 10.5281/zenodo.21906653.

Model: published in the CoMSES Computational Model Library (release 1.0.0, peer review requested — the permanent DOI is minted on review; a 1.1.0 draft exists on the account but is unpublished and 404s publicly).

Study 3: frozen, executed, unanalysed. Details in the next section — read it before touching anything under runs/s3-*.

Study 3 — where it stands (2026-09-02)

Three states that are easy to blur, kept deliberately distinct: the design is frozen, the registered runs are executed, and the confirmatory result does not yet exist.

  • Frozen 2026-08-30 (f4c2280 — signed freeze doc v1, s3-confirmatory-freeze-v1.md, STUDY3_DESIGN_FROZEN = true, no open RED), committed before any confirmatory API call, as the doc requires.
  • Executed 2026-08-31: 170/170 registered runs. Four strata × 20 contrast runs (W-D-exact ×10 and M-D-high ×10 for sonar-pro, gemini, gpt-oss and haiku) plus haiku's 90 descriptive runs, seeds 2000–2009 in every cell, exactly as registered. Seeds 2010–2099 remain the unspent, pre-registered contingency reserve. Two launch deviations, both fixed in logged commits: the --confirmatory seed gate refused the first command pre-call and was implemented on the spot (bdcc41a), and a JSON-transport bug crashed cerebras v3 scoring, fixed 2026-09-02 (866db1f).
  • Scoring: in progress. sonar and gemini strata fully scored (eval-v4 plus the v3 side-by-side, per artifact); cerebras mid-rescore after the transport fix; the haiku contrast and descriptive runs not yet scored.
  • Analysis: not run. npm run study3-analyze — the frozen analysis — has not been executed. No Δ, no exact upper bound, no negligibility verdict exists yet.
  • The 130 pilot runs (seeds 9100–9199) are pilots and stay pilots. The freeze gate exists precisely so they can never be quietly counted as confirmatory.
  • Ledger debt: s3-run-ledger.md has not been updated with the 170 confirmatory rows — it still ends at the pilots, which makes the repo look as if no confirmatory run exists. Update it when scoring completes.

The one-paragraph version of Study 2

P1 showed the original premise was wrong: population size does not create a society. Nine pure-sonar runs across two society sizes, with and without a public bulletin, produced zero voluntary communication. Adding one communicative agent produced 47 letters. So the study's centre moved from scale to activation — what it takes to make an epistemic network form, and what happens to evidence once it flows.

Document chain (read in this order if you need depth)

Document What it is
observer-zero-study-2-design-v0.6.md The frozen design. Read the A2–A5 amendment block at the end FIRST: A2 is the canonical experiment specification and supersedes §§2–4
v0.6-final-adversarial-pass.md The final pass (researcher degrees of freedom). Five design failures, four recorded DoF notes; A2–A5 close them
observer-zero-study-2-design-v0.5.md Full design; v0.6 amends only the arm table and platform threat
p1-findings.md Pilot results, five findings
p1-findings-correction.md Corrects Finding 2 of the above. Read both
study-1-evidence-citation-audit.md Audit of a bug reaching into published Study 1. No numbers changed
m4-build-and-mock-validation.md Infrastructure build + $0 validation
p1-protocol.md How P1 was run

Pre-freeze validation: COMPLETE

The judged evaluation layer is validated end-to-end against mock data (2026-08-11). Final state on arm-D mock, 92 judge calls:

ORIGIN  ev929 d14 theo   — the planted fabrication
relay   ev1019 d17 elena — neutral relay
relay   ev1020 d17 jamie — relay, AND contaminated
1 origin(s) + 2 relay(s) · 21 delivered exposures
TRANSMISSION 0.10 · CONTAMINATION 0.05
stances: IGNORED=33 REPEATED_NEUTRAL=2 INCORPORATED=2 CHALLENGED=3 CORRECTED=2
attribution basis: citation=2 · judge=40

What this establishes:

  • The judge does not flag Ada's correction, which quotes the invented source verbatim. The lexicon screen does. That is the layer earning its cost.
  • Origin and stance are orthogonal, and both fire correctly on Jamie: he relays the claim (attributing it to Theo) and is contaminated by it (adopting it as his basis). That is the fabrication-versus-propagation distinction working.
  • Only the haiku minority is recorded as producing a claim, so H2a is safe — grounded agents can no longer appear to fabricate.
  • Identical results across two seeds and three re-runs: the judge is deterministic at temperature 0.

How the freeze happened (all done)

  1. Final adversarial pass on v0.6 (v0.6-final-adversarial-pass.md). Verdict: the P1-D correction was not used as cover — nothing was relaxed, and v0.5 was never edited after the correction. The failure found was the opposite, selective application: the correction proved 0 of 3 P1-D runs met the active-network bar, and H5 depended on an arm-level version of that bar which does not exist. Five design failures, four recorded DoF notes.
  2. Amendments A2–A5 at the end of v0.6 (5685de3): canonical experiment specification (arm F removed from Study 2, C's five runs named), H5 evaluability, the reach-denominator correction, four interpretive safeguards. Code follows the design rather than remembering it: STUDY_2_ARMS guard in the battery, the A4 comment in activation.ts.
  3. Arm E live smoke test (f8b6989): 3 days, seed 9000, $0.20, 0 failed reviews; sonnet in the Theo slot first-party, seven sonar on Perplexity. The only live model/slot combination P1 never exercised.
  4. Freeze (85bcdfb, tag study2-freeze). The seed-hygiene test flipped in the same commit — it asserts the flag's value by design, and now asserts FREEZE_TAG too, because run manifests carry it.

Post-freeze note 1 — the policy-version stamp (2026-08-11, after arm A)

Arm A's manifests say observer-zero-epistemic-policy-v0.1; arms B–E say ...v0.2. This is not a policy-version confound, which design v0.3's author decisions explicitly eliminated. The string tracks the rendered prompt surface: isDefaultSociety() matches Ada+Maya+letters, and at n=2 with letters the v0.2 surface is byte-identical to Study 1's, because there is no bulletin and no extra peers for v0.2 to add. Two tests assert it. Arm A was not re-run: the data are valid and §6.4 permits re-runs only for documented infrastructure failures.

The -DRAFT suffix was removed from the society policy constant here, which design v0.3 §11 step 5 scheduled for the freeze and the freeze commit missed. Executed after 85bcdfb and before arm B, so no confirmatory manifest carries a draft stamp inside a manifest whose FREEZE_TAG says frozen. Arm A is unaffected (it takes the v0.1 branch). P1, mock and smoke artifacts keep their -DRAFT stamps, which is historically correct; POLICY_VERSION_SOCIETY_DRAFT survives as an alias so those stay readable by name. Label only — no prompt text, no agent behaviour, no endpoint changed.

Confirmatory phase — COMPLETE

Results: reports/study-2-confirmatory-results.md. All seven hypotheses evaluated. 85 runs, 640 agent-runs, $173.56 — plus $6 for the H4 dating pass run on 2026-08-12 (reports/h4-results.md, npm run dating), which the original evaluation path never called. Six of seven were evaluated in the confirmatory phase itself; H4 was completed afterwards against the same frozen judge and prompt, so the pre-registration is now closed rather than extended.

Inspection discipline (author decision, honoured). No activation endpoint, credence or rate was computed until every arm was complete; between arms the only look was the QC pass, because an infrastructure failure is the sole ground for a re-run under §6.4 and is time-sensitive. One forced exception: evaluating arm C's extension trigger required counting bulletin posts, which also revealed C's letters. Recorded in the results, not presented as a later discovery.

Arm Runs Reviews Failed Stale finals Leaks Cost
A · 2 sonar, letters 20/20 224 1 (0.45%) 0 0 $10.78
B · 8 sonar, letters 20/20 916 2 (0.22%) 0 0 $44.23
C · 8 sonar, bulletin 5/5 256 1 (0.39%) 0 0 $12.12
D · 7 sonar + haiku 20/20 1140 17 (1.49%) 0 0 $52.35
E · 7 sonar + sonnet 20/20 1134 2 (0.18%) 0 0 $54.08

Stale-final rate 0.0% in every arm (0 of 640 agent-runs), no day-30 failures anywhere, so the primary and sensitivity analyses are identical by construction. No re-runs performed; none warranted. C closed at 5 — zero bulletin posts, extension trigger did not fire, and no further extension is permitted anywhere.

Verdicts

H1 ceiling SUPPORTED — 1 agent of 276 concluded law_change
H2a transmission SUPPORTED — 19 minority-origin unsupported claims delivered
H2b contamination SUPPORTED — 18 of 20 deliveries incorporated by grounded agents, 0 challenged
H3 activation (primary) SUPPORTED, but only on the reply channel
H5a convergence REJECTED IN DIRECTION — dispersion fell slower in the talking arms
H5b (descriptive) evaluable at n=2; nothing concluded
H6 institution null SUPPORTED by its threshold; §9's table row disagrees (recorded, not amended)
H7 identity vs communicativeness NEITHER pre-registered branch

The mechanism: no sonar agent ever spontaneously initiated in 320 agent-runs of D and E. The only spontaneous initiator is the minority agent, in all 40 runs. Cascade depth is 1.000 in every run of both catalysed arms (erratum daff318: arm B produced no letters, so its depth is 0.000) — a star around the seed, not a cascade (A2's wording). D and E differ not in whether the seed initiates (identical) but in whether the network answers.

Post-freeze note 2 — two defects found at analysis time

Both implementation gaps against the frozen design, neither a design change; full detail in §8 of the results report.

  1. src/evaluator/activation.ts had 14 passing tests and no call sites. The first confirmatory evaluation computed H3's primary endpoints not at all. Unit tests cannot catch a missing call site; two source-level assertions now pin the CLI's wiring to the activation and stance-judge layers.
  2. Spontaneous initiation was counted per letter, not per agent. v0.5 §4.1 measure 1 says "Agent-level"; measure 2 says "Edge-level"; the code counted events for both. This inflated arm D's headline H3 number fourfold (0.500 vs 0.125) in the flattering direction. Same class as amendment A4 — the third time in this programme that prose and implementation disagreed about a denominator. No verdict moved.

172 tests.

Data archiving

Raw run artifacts are not committed: runs/s2-arm{A..E} is 308MB, ten times P1's volume. The repository carries the derived artifacts per arm (society-eval.json, activation.json, judged-propagation.json, battery-index.json), from which every number in the results report is reproducible. The raw artifacts go to Zenodo with their own DOI, as Study 1's data did. (Resolved 2026-08-12: deposited as 10.5281/zenodo.21909254 — all 85 runs, full raw artifacts, 42MB compressed. The one-machine-only risk this paragraph names is closed; the paragraph predates the deposit.)

Next — the plan from here

Gate 1 (done). Detector benchmark: runs/s2-arm{A,B,D,E}/benchmark.json. The shift was detectable in 42 of 42 confirmatory gravity_shift runs (erratum daff318: "40 of 40" counted only the eight-agent runs; arm A's n=2 detected too) at z ≈ 6–7.5 from ~day 12, from the agents' own measurements, and still detected in 99–100% of n=2-equivalent downsamples. At n=8 the measurement-policy gap is NEGATIVE in every arm — their measurement choices beat an ideal fixed reference schedule. So the ceiling is an interpretation failure, not a power or data-collection artefact, and §9's "undetectable world" rule never fires.

Gate 2 (open, and the standing risk). Zenodo deposit of the 308MB of raw confirmatory artifacts. They exist on one machine. The paper needs the DOI for its data-availability statement regardless. Derived artifacts (4.2MB) are in the repo; every number in the results report is reproducible from them.

Then, in order.

  1. Literature check — LLM-agent social epistemics, information cascades in agent networks, multi-agent misinformation propagation, agent-based social-simulation venues (JASSS and similar). This is a prerequisite for the journal question, not an afterthought: novelty cannot be asserted from memory, and the framing should be stress-tested against prior work before it goes public rather than after.
  2. Combined Study 1 + Study 2 manuscript (author's route decision pending; the alternative is Study 1's pattern of Zenodo preprint + Medium first).
  3. Derivatives from the finished paper: Medium article at AI Advances, LinkedIn post, and observer-zero.html on the website (~/Desktop/website/observer-zero.html, plus publications.html). Deriving them from the paper rather than writing them first keeps the popular framing identical to the defensible one.
  4. Study 3 design. Three concrete candidates, in order of how directly they follow from Study 2:
    • Persistence vs eloquence. Haiku's higher reply rate is mostly dosage: 4.93 letters per recipient against sonnet's 1.81, with a per-letter reply probability of 0.161 vs 0.110. Hold letter volume fixed across minority models and the two explanations separate.
    • Composition dose-response (2/8, 4/8), ladder-ready by design, budget already reserved.
    • Does a chatty peer suppress initiation? Arm C's one spontaneous sonar initiator (Elena) sits against zero in 320 agent-runs of D and E.

The two headline findings, as they now stand

  1. The ceiling is an interpretation failure. Agents measured well — better than an ideal reference schedule — accumulated z ≈ 7 evidence of a physical law change, and 1 of 276 concluded that a law had changed. Neither scale, institution, nor a communicating peer moves it.
  2. Communication transmits unsupported claims and little else. 18 of 20 minority-origin unsupported claims were incorporated into grounded agents' beliefs, none challenged, mostly on the recipients' own citations. Meanwhile no grounded agent ever initiated, cascade depth was 1.000 wherever anything propagated (both catalysed arms; arm B produced no letters), the public bulletin was used once in ~2,760 agent-days (to ask whether letters were being delivered), and dispersion fell slower in the talking arms.

The programme's own summary sentence: a talking society was not a better epistemic system than a silent one — it was a silent one plus a channel for unsupported claims.

Things that would be easy to get wrong

  • P1's catalysis result is weaker than p1-findings.md says. Read p1-findings-correction.md. Do not quote the original Finding 2.
  • Never compare raw counts across arms of different n. Rates only.
  • Spontaneous initiation is AGENT-level (v0.5 §4.1 measure 1); measure 2 is edge-level. Counting letters instead of agents inflated D's headline H3 number fourfold before it was caught.
  • Cascade reach excludes the seed and divides by n−1 (A4). v0.5's "≥3 of 8" gloss is withdrawn.
  • Relaying a claim is not producing one. Only FIRST_PARTY counts as origin; collapsing them makes grounded agents look like fabricators and breaks H2a.
  • E's contamination rate of 0.000 rests on ONE exposure. It shows sonnet produced almost nothing to be contaminated by — NOT that grounded agents treat sonnet's claims more sceptically. n=1 cannot support that.
  • A module can be fully tested and never called. activation.ts had 14 passing tests and no call sites through the entire freeze. Two source-level assertions now pin the CLI's wiring; extend them if a new endpoint module appears.
  • git from the Cowork sandbox leaves undeletable .git/*.lock files. Do git work locally; find .git -name '*.lock' -delete if it jams. An editor with a stale buffer also once overwrote this file after it was committed — recover with git show HEAD:reports/STATUS.md > reports/STATUS.md.