Where Observer Zero goes after Study 1. Items marked registered were pre-committed in the Study 1 report as future work; the rest is direction, not promise. Order within sections is rough priority. Contributions toward any of these are welcome – see CONTRIBUTING.md.
- Communication-budget-matched confabulation comparison (registered). Sonar Pro fabricated nothing but also communicated far less, so it had fewer opportunities. A design that matches message budgets across models will separate "doesn't fabricate" from "doesn't talk".
- v0.3 dual-prior ablation (registered). Study 1 removed the "prefer mundane explanations" belief-prompt line; the decision prompt's "not a philosopher on watch for the extraordinary" cue was deliberately retained for single-variable discipline and remains a registered suspect. v0.3 ablates both, separately and together.
- External methodological review (registered). An independent human reader of the design and scoring rules before Study 2 freezes.
- Where does the ceiling live? Decompose the commitment failure: belief-update prompt, memory digest, or the daily action loop. Targeted interventions on Sonnet's peak-and-abandon trajectories, using stored Study 1 artifacts as the baseline.
The headline next experiment: a mixed-model society (around eight agents, different underlying models in the same world) with shared institutions – at minimum a newspaper or bulletin that makes testimony a first-class object. Questions Study 1 could not ask:
- Does a mixed society inherit the culture of its majority model, its most talkative model, or neither?
- Can one well-calibrated agent (a Sonar-like sceptic) discipline a fabrication-prone society – or does confident fabrication outcompete quiet accuracy?
- Do institutions (shared records that agents can cite and check) reduce fabrication, or give it a bigger surface?
Design constraints already fixed: same frozen-manifest discipline, same closed-world invariants, pre-registered scoring before any live runs.
- Additional model providers behind the same closed-world interface (open-weight models are particularly interesting: full logit access would allow measuring when the correct hypothesis dies, not just that it does)
- Run-artifact schema docs, so third parties can analyse published runs without reading the TypeScript
- A lightweight run inspector (event feed, belief map, message threads) for qualitative auditing – currently done by reading JSON
- Cheaper judged evaluation: cache-aware judging and sampling strategies for larger batteries
- Medium article for a general audience (in progress)
- arXiv preprint (deferred by design: after community reaction and at least one further study)
- Fuller related-work integration for the journal-length version
Things Observer Zero deliberately does not pursue:
- Coaching agents to the answer. The point is measuring epistemics under a frozen condition, not engineering a society that "solves" Meridian. Improvements to agent performance are new versioned conditions, never patches to a frozen one.
- Open-ended worlds. Meridian stays small and fully specified; the power of the paradigm is that ground truth is complete and every claim is checkable.
- Unfalsifiable questions. Agents are only scored on propositions resolvable inside their world.