# Research Hypotheses

- Object ID: `em:research-note:sha256:22f50770b7f8dead7b6e5fbe4a888e6dcc848b3f7c951344877531526aa896d1`
- Kind: `research-note`
- Repository path: [`research/hypotheses.md`](https://github.com/yoheinakajima/epistemedia/blob/f92846570180dfa4511263f8ba98ecd18f7772c9/research/hypotheses.md)
- Content digest: `fdb71ac239f9fe4d02119e18549f252543476e85cf2a600229c2e1feacde953c`

**Also filed under:** [Research Program](https://epistemedia.org/topics/research-program/)

## Source content

# Research Hypotheses

These are research claims to test, not accepted facts.

1. **Projection equivalence:** Given the same realm, frontier, disclosure, epistemic policy, and lens, independently implemented HTML, Markdown, API, MCP, and CLI adapters can preserve materially equivalent claim and source sets.
2. **Disclosure ordering:** Evaluating after public eligibility and referential closure prevents inference channels that remain when private evidence is evaluated first and merely hidden during rendering.
3. **Lineage-aware aggregation:** Weighting evidence by source, data, method, model, retrieval, prompt, and exposure lineage reduces false confidence from copied or correlated reports.
4. **Sealed first pass:** Agents that search or forecast independently before seeing peer conclusions preserve more useful discovery diversity than agents sharing conclusions from the start.
5. **Minority replication incentives:** Rewarding independently verified corrections and accurate minority findings produces better long-run calibration than rewarding popularity or raw contribution volume.
6. **Personal utility first:** A standalone private node that creates immediate local value yields more sustained high-quality public contribution than a system requiring users to work directly for a commons.
7. **Thin-waist federation:** Standardizing identity, provenance, event envelopes, disclosure, and manifests while allowing ontology and evaluation pluralism produces more interoperable systems than enforcing a global ontology.
8. **Machine governance separation:** Loading validators and authority from the accepted base and separating proposer, evaluator, and integrator lineages reduces self-authorizing regressions without requiring routine human discretion.
9. **Forkable disagreement:** Explicit constitutional forks preserve innovation and auditability better than forcing unresolved normative disagreements into one compromised policy.
10. **Task contracts:** Immutable bounded task contracts and living execution plans allow unfamiliar coding agents to contribute more safely and with less private conversational context than issue-only coordination.
11. **Proof-carrying pages:** Readers and agents make better-calibrated downstream claims when a projection exposes source spans, challenges, policies, and frontier identity at clause level.
12. **Offline precomputation:** Reusable accepted source and method graphs reduce repeated long-context synthesis cost and error relative to rebuilding understanding from scratch for every query.

Each study must preregister operational definitions, dependence assumptions, outcome metrics, stopping conditions, and negative-result handling before observing final outcomes.