# The adversarial review protocol, v1 **Published 11 September 2026.** A description of the review procedure used on the Space Immanence meta-problem programme (9–11 September 2026), written so it can be copied. The programme it governed is published in full at `downloads/meta-problem/`; this document is the method, separated out because the method outlived the result. The programme's one empirical test returned a null and its language-model line was stopped. The protocol is published on the strength of what it caught before that point, and its limits are stated in §6. ## 1. Purpose A research programme run by a single author, or by a single author with model assistants, has one structural weakness: every party to it wants it to succeed. Findings that would cost the programme its result are the findings least likely to be looked for, and a repair applied quietly to a model is indistinguishable, in the finished artefact, from a model that never needed one. The protocol below is one answer to that. It is designed for a programme that must be prevented from rescuing itself. It assumes a human owner with filing authority, and a supply of model agents that can be run independently, in parallel, without seeing each other's work. It does not assume that any agent is more capable than the author. It assumes only that an agent instructed to find defects, given the artefacts and the time to run them, finds different things than an agent instructed to build. ## 2. The seed Before any work, the owner files a seed: a single document stating the target, the hypothesis, and the rules that bind every deliverable produced under it. The rules used here were these. - **No experiential term in a derivation step.** For a programme explaining why judgements about consciousness are made, any step that uses *experience*, *phenomenal*, *conscious* or *qualia* in its own reasoning has assumed what it set out to explain. The rule is checkable line by line, and a critic was assigned to check it. - **Hedges verbatim.** A hedge in a source is carried into every downstream document unaltered. Summarising a hedge away is treated as a defect of the same kind as misquoting. - **Failures at the same prominence as successes.** A negative result appears in the same table, the same summary line, and the same status header as a positive one. - **Every document ends with "What would count against this",** naming a result that would kill it. A document that cannot name one is not finished. - **Pre-registered predictions written in code before a build.** A model's predicted results table is committed before the model is run, in a form that can be diffed against the output. - **Every deviation reported, never repaired.** Where output departs from the registered table, the departure is recorded as a finding. Editing the decoder until the table comes out right is the specific failure the rule exists to prevent. The seed also carries a numbered amendment section. Nothing binding changes except by a dated amendment, and amendments are filed by the human, never by an agent. ## 3. Rounds A round has three parts, and only the third is unusual. **Builders, in parallel and independent.** Deliverables are drafted by separate agents that do not see each other's drafts. Where two builders produce incompatible accounts of the same object, that incompatibility is itself a finding, and it surfaces without anyone having to look for it. **Critics, adversarial, with a fixed severity vocabulary.** Each critic is given the artefacts, the seed, and an instruction to find what is wrong. Every finding carries one of four severities — **FATAL** (the artefact does not show what it claims to show), **BLOCKER** (a claim in the artefact is false as stated), **MAJOR** (a claim is unsupported or overstated), **MINOR** (a defect that does not reach a claim) — and a severity table appears at the head of the critique, before any prose. A finding without a reproducible demonstration is downgraded. In this programme critics ran the models, wrote probe scripts, recomputed the item statistics from the committed JSON without the runner, and mutated the item bank twenty-two ways to check that the validators refused each mutation. A critic that only reads is a proofreader. **Refuters, where a finding is contested.** When the builder disputes a critique finding, a third agent that has seen neither exchange is given both and asked which holds. This step was used sparingly and is the part most likely to be skipped under time pressure. **Go/no-go gates before any spend.** No study runs until a critic reviews the study alone — not the programme, not the theory — and closes with a mandatory verdict line reading exactly GO or NO-GO. A conditional GO must list its conditions as numbered edits. The verdict line is mandatory because a critique that ends in discussion ends in the author's judgement, which is the thing the round exists to check. ## 4. Labelling Three labelling rules did more work here than any single finding. **Theorems of the declaration are labelled as such.** When a model's results table is an arithmetic consequence of how the registry was declared, every such cell is marked, and the table carries no evidential weight. The v3 model's table is 63 cells, all of them so marked. Before this rule was adopted, the same programme had reported "twenty of twenty-one reproduce" as a result; a critic showed that the row in question was unsatisfiable for any registry line whatever, and the claim was withdrawn. **Errata blocks, never silent edits.** A document corrected after a critique carries a dated erratum block at its head stating which items govern where the body conflicts with them. Later editions may absorb the block into the body, and must say which item went where. **Amendments numbered and filed by the human before the build they govern.** An amendment that arrives after the build it justifies is a repair wearing a registration's clothes. ## 5. What it found here Round 1: three review documents against the first model build and the derivation work; the circularity review alone returned 1 FATAL, 6 BLOCKER, 3 WARNING and 1 NIT. Round 2: two critiques; the model critique returned 3 FATAL and 3 BLOCKER, the documents critique 1 FATAL, 3 BLOCKER, 8 MAJOR and 4 MINOR. Round 3: 2 FATAL, 4 BLOCKER, 7 MAJOR, 5 MINOR. Round 4, the first go/no-go: 2 BLOCKER, 3 MAJOR, 3 MINOR, verdict NO-GO. Round 4b, after the rebuild: 0 FATAL, 0 BLOCKER, 3 MAJOR, 3 MINOR, verdict GO conditional on three filing-side edits that changed no item, no call count and no cost. Four of those findings were things the author and the builders had not seen. 1. **The first model's central result was produced by a rule nobody had registered.** The contrast architecture's dissolution outputs came from a cycle-termination rule reached by locating the modelling activity inside the arena — the opposite of the organisation the hypothesis names. The model was superseded. 2. **The rebuilt model's signature test was passed by an inert value.** Setting the term's value to nothing left every judgement output unchanged, so the test separated presence from absence and nothing more. The property the whole design turned on was undemonstrated, and the seed's decisive proviso went back to untested. 3. **A prediction attributed to a rival had no textual basis — on the published site, not only in the programme.** The reply to objection 12 said illusionism predicts two intuition families weaken together. No illusionist text says so. The finding travelled from a research branch to a live page, and the correction is in this release. 4. **The framing study's anchor factor was perfectly collinear with its appearance vocabulary,** across all 112 framed items in each cell, with both off-diagonal cells empty. A positive result would not have been attributable to indexicality at all. Caught at the go/no-go gate, before any money was spent. ## 6. Limits **One drafting family.** Builders and critics were all Claude Opus 5. Their errors are correlated in ways nothing here measures, and a defect the family shares is a defect no round could catch. This is the protocol's largest weakness, and running critics from a different family is the first improvement to make. **The critic can be wrong in both directions.** Round 4's two blockers were right, and both were closed by design rather than by disclosure, verified by independent recomputation in the next round. Round 3's second FATAL is arguable: it rated a defect in an unregistered probe inside a side note at the same severity as a defect voiding the main build, and the note had already conceded the underlying point. A severity vocabulary makes findings comparable; it does not make them correct. Severities are the critic's, and the record keeps them as filed. **Cost.** About one million Opus tokens per round, across five agents in the larger rounds. Five rounds is not a cheap review, and the programme it reviewed produced a null. **No human reviewer.** Nothing in the programme has been read by anyone outside it. Model critics found what is described above; they cannot tell anyone whether the target was the right target. ## 7. How to reuse it Three artefacts carry the method. **The seed template**: target, hypothesis, deliverables with success criteria, the binding rules of §2, and a numbered amendment section. **The round script**: parallel builders who cannot see each other; one or more critics with the four-severity table and an instruction to demonstrate rather than assert; a refuter where a finding is contested; and, before any spend, a go/no-go on the study alone ending in a mandatory verdict line. **The filing rule**: the human files amendments and pre-registrations, always before the build or the run they govern, and every deviation is reported rather than repaired. ## What would count against this - A programme that ran this protocol and shipped a defect of the kind §5 catalogues: the protocol would then be a procedure that produces critique documents rather than one that catches defects, and the four findings above would need re-reading as luck. - A critic from a different model family finding, in these same artefacts, a defect of FATAL severity that five rounds of one family missed: the correlated-error limit in §6 would then be the dominant fact about the protocol, and the counts in §5 would be a measure of one family's blind spots. - A demonstration that the severity vocabulary changed a builder's behaviour rather than only its reporting — that artefacts were written to survive the rubric — which would make the rounds a compliance exercise and the counts an artefact of the rubric. - A cheaper procedure with the same catch rate on the same artefacts, which would make the token cost in §6 the argument against running this one.