0010 — Research stays open; claim gates carry scope and updating evidence

Status: current. Consolidates the live governance principle from 0009 and narrows categorical readings of 0002, 0004, and 0005.

Context

Several defensible precautions became broader than their evidence. Source limitations were read as bans on acquisition, a zero-leverage recommendation became a restriction on financing research, fixed numerical reference points became rejection hurdles, and underpowered nulls became apparent closures. The frozen specifications and results remain valid records of what was run. Their interpretation must remain no broader than the design.

The opposite absolute—“measurement is never gated”—is also too broad. Law, licensing, research ethics, security, safety, and destructive acquisition costs can constrain work. The problem addressed here is narrower: an earlier empirical conclusion should not, by itself, foreclose a safe and lawful way to learn more.

Decision

1. Research is open by default. Existing findings do not prohibit registering a candidate, acquiring an appropriate source, changing an estimand, pricing a construction, or running an exploratory comparison. State how the new question differs from the scope already tested.
2. Evidence standards rise with the claim. Exploration establishes mechanism, magnitude, provenance, and next tests. Evaluation freezes decision-relevant choices. Promotion or publication adds independent or forward evidence and implementation review appropriate to the claim. The tiered protocol is in docs/AGENTS.md.
3. A durable claim gate names four things: the claim it limits, the evidence and failure mode behind it, its scope, and the evidence or review trigger that would change the decision. Constraints already enforced by law, licences, code, or tests need not be duplicated as prose policy.
4. Frozen results and current interpretations are separate. Do not rewrite a specification, falsifier, ledger entry, or artifact after seeing results. If a design could not resolve the effect or its rule was defective, preserve the original outcome and publish a corrected claim-level interpretation with an audit trail. A correction may not manufacture positive evidence.
5. Thresholds come from the decision and design. The former 2.0 pp/yr materiality point and 0.30 pp/yr sleeve point are useful sensitivities, not universal bars. A new evaluation derives or justifies materiality, shows its relationship to MDE and arithmetic ceilings, and freezes it before inspecting the result. An exploratory screen may use a rough reference point if labelled as such.
6. Funding is part of the hypothesis. Pro-rata sale, cash funding, leverage matching, and financed overlays answer different questions. The zero-leverage reference portfolio remains a current recommendation for an underspecified investor; it is not a restriction on researching financed constructions.
7. Statuses apply to claims, not whole strategy families. A product can fail a loading test without being a bad product; a vendor index can be contradicted without rejecting the strategy; a premium can remain unresolved while an implementation is observable.

Current interpretation of affected work

• Experiment 010’s frozen outcomes remain in its artifacts. Where its threshold sat above the arithmetic ceiling or below the instrument’s resolution, the portfolio claim is unresolved by that design, not evidence that the sleeve cannot help.
• Experiments 004 and 008 remain about different objects: a vendor trend series and product exposure delivery. Neither is a universal verdict on trend.
• The French-file RMW and CMA work establishes the resolution of those panels and windows. It does not establish a zero premium or prohibit conditional models, different estimands, non-French sources, longer samples, or lower-variance designs.
• The audited free fund-price feeds remain inadequate for confirmatory total-return claims that depend on distributions and corporate actions. That finding does not automatically reach non-distributing assets, reader-supplied data, source reproduction, cross-checks, or sources with different contracts.

No sleeve is promoted by this decision. Promotion status is a current research conclusion, not a permanent governance rule.

Consequences

The repository can search cheaply and creatively while keeping stronger claims expensive. Specifications and artifacts preserve the exact historical result; syntheses carry the best current interpretation. New work must explain scope and evidentiary tier, but it does not need a new decision record merely to revisit an empirical question.

This policy should be reviewed when the status vocabulary, artifact model, or publication contract changes. It is superseded if a later record replaces the tiered evidence model.