Adversarial registry. Plan

Status 2026-10-30: protocol, template and first-pairing proposal drafted; emails awaiting sending. Third spine line; low compute. Context: ../../research-directions.md section 2c; Tree G supplies the checklist (../../trees/02-g-theory/).

Goal

A public registry of pre-registered claims about machine minds. Each claim carries divergent predictions from at least two camps and a neutral executor, and its result is published whatever the outcome. The institution runs the first entries and hosts the rest.

Deliverables

1. protocol.md, the rules, adapted from the Templeton World Charity Foundation's structured adversarial collaboration format and the Cogitate consortium's published methods. It covers the eligibility of a claim (it must yield a divergent, measurable prediction on an open model), the roles (proposer camps, executor, reviewer), what a pre-registration contains, the freeze date, the analysis plan, the publication rule, and dispute handling. Under the publication rule both camps sign the pre-registration, and the executor publishes the result with both camps' pre-committed interpretations.
2. entry-template.md, with fields for the claim, the theories, models and versions, each camp's prediction with a numeric threshold, the executor, the analysis plan, the freeze date, status, the result, and both interpretations.
3. The registry itself, an OSF project (or a static page mirroring it) that lists the entries.
4. A first completed entry.

First pairing

Revised 2026-10-30; full proposal in first-pairing.md.

Claim: after a concept injection is removed, an introspective query recovers the injected concept better than a matched non-introspective query over the same cache.

• Camp A (Pearson-Vogel, Vanek, Douglas, Kulveit, arXiv 2602.20031) recovers which of nine concepts was injected into an earlier turn of Qwen2.5-Coder-32B-Instruct, at up to 1.36 bits.
• Camp B (Lederman and Mahowald, arXiv 2603.05414) holds that detection is content-agnostic, and that steering effects persist without being introspectively accessed.

The divergent prediction is ΔMI, the concept information gained by asking the introspective question over the non-introspective one. Nulls are no injection, random directions of matched norm and shuffled labels. Cleanstone executes, Eleos is asked to host, and Lindsey is asked to comment on the measurement only.

The 2026-09-30 draft named Berg and collaborators as Camp A. That was wrong: Berg, de Lucena, Rosenblatt (arXiv 2510.24797) study experience reports under self-referential prompting, mainly on closed models, and found that suppressing deception features raises the rate of experience claims. They make no claim about identifying injected concepts. Their result moves to the fallback list.

Fallback pairings if the first declines:

• Self-referential experience reports (Berg, de Lucena, Rosenblatt 2025) against a role-play reading. Needs a counter-camp willing to name a threshold, and an open model on which the deception-feature effect reproduces.
• Workspace-as-access-consciousness (the Eleos reading of Gurnee et al.) against mimicry (Butlin et al. 2026 TiCS). Does workspace structure appear in a model trained only on data not generated by humans (code, synthetic)?
• IIT's substrate claim (Findlay, Koch, Tononi 2025) against functionalist readings. It is hard to operationalise on software, so it goes into the registry as a documented open problem and is not a candidate first entry.

Steps

1. Draft protocol.md and entry-template.md (two weeks).
2. Tree G reviews both against the checklist, then a style pass.
3. Take the protocol and the first-pairing proposal to Eleos (Long, Butlin) and ask whether they will host or co-host.
4. Send Camp A and Camp B a one-page proposal with the template pre-filled.
5. If both accept, freeze the entry, execute it under the ../02-open-replication/ protocol, and publish.
6. If either declines, publish the protocol and the invitation, with the institution's own claims as the first entries. Tree A predictions P1 to P5 go in marked as retroactive, and developmental-atlas predictions go in prospectively.
7. Present the registry at the next Eleos meeting or the next AMCS Models of Consciousness meeting. AMCS 7 (Copenhagen, 12 to 16 October 2026) passed before the registry was ready.

Decision gates

• If no camp accepts within six months, the registry carries on as a self-pre-registration standard for the institution's own work, and we revisit the recruitment goal yearly.
• Once one entry completes, recruit the second pairing from the fallback list.

Cost

Documents and meetings, plus one replication-scale run per entry.