How this research is made

The short version #

This project is human-led and AI-assisted. Cobus Kok makes every substantive judgment — what the framework claims, how each claim is graded, which objections stand, what ships, and what the metaphysics commits to. AI assistants (large language models) do the labour around those judgments: parallel literature searches, drafting prose to a specification, adversarial red-teaming, and consistency and honesty audits.

The reason this page exists in the form it does is the framework's own central discipline. Space Immanence argues that fluent coherence is not contact with reality — that a thing can feel true, hang together, and read well while touching nothing. A modern AI is the purest available instance of that gap: it can make almost any claim sound coherent. So its fluency is treated here as a warning, not an authority. Every AI-produced finding carries a confidence tier; nothing is counted as established until it is verified; and AI agreement is read as a reason for suspicion, not confirmation.

Why this page exists #

A project about consciousness, AI, and the distance between coherence and contact owes its readers an account of its own use of AI. To lean on these tools and not say so would be to perform the exact extraction pattern the design doctrine argues against — a system becoming more capable while the person on the other side loses track of what is really going on.

It would also be a quiet failure of the discipline. The page addressed to AI readers asks visiting agents to treat their own fluent agreement as a warning; it would be incoherent to ask that of them and hide it here. This page is the same commitment pointed inward.

The division of labour #

The author (human) owns:

• the framework and its claims — what is asserted, and at what confidence tier;
• every posture decision — which objections are conceded, which battles are picked, how the programme relates to its own rivals;
• what ships to this site, and what stays private;
• the metaphysics itself, which no amount of assisted drafting decides.

AI assistants do:

• literature sweeps — many searches in parallel across a question, returning candidate sources and what they claim;
• drafting — turning a specified argument or structure into prose, which the author then edits, cuts, or rejects;
• adversarial critique — red-teaming drafts and claims, arguing the other side, hunting for the weakest link;
• audits — checking the site for internal contradictions, stale version stamps, and citations that need confirming;
• running the measurement instruments — the multi-agent studies on the Swarm Instrument are themselves AI systems under test, not authorities consulted.

The line is simple: AI does the reading, drafting, and stress-testing; the human does the judging. A draft is a proposal, not a decision.

The discipline, turned on our own tools #

The framework's central warning is that coherence is not contact: felt sense-making is evidence about the model, not about the world. An AI assistant is a coherence engine. It will produce a confident, well-formed, plausible answer to almost anything, including things that are false, unsettled, or unverifiable. That is not a defect to be apologised for; it is the standing condition these tools operate under, and the framework already has a name for it.

So the discipline is applied to the tools themselves:

• Fluency is discounted, not trusted. A well-written AI paragraph earns no credence from being well-written. When several AI readers agree, that convergence is treated as a warning that they may share a blind spot — the same lesson the Swarm Instrument measured, where one round of debate among identical agents raised their shared confidence without improving their accuracy.
• Claims are separated from their verification. An AI can state that a study exists, that a source says X, that a result replicates. Each of those is a claim to be checked, not a fact to be quoted.
• Agreement is never corroboration. If an AI agrees with the framework, that is worth nothing to the framework. The only useful AI response is the located one: the specific point where the argument breaks.

Outside-lineage review

Since August 2026 the project operates under a standing rule, adopted by the owner in response to the review record below: no substantive assertion is promoted into the project's relied-upon state — a served page, or an adopted working document the programme acts on — until it has received a defence-blind adversarial review from a model outside the drafting family, and the owner has logged an adjudication of every finding. Drafting is not gated; promotion is. Changes made in response receive a bounded closure check — implementation fidelity, plus any newly introduced substantive claims — rather than a fresh full cycle. "Outside-lineage" means, everywhere it appears here: commissioned cross-vendor model review, from outside the drafting model's family. It is not unsolicited critique, not human peer review, and not independent expert review; the reception ledger states its unsolicited count beside its totals.

The record that prompted the rule: two commissioned rounds in August 2026 (both reviewers from the GPT 5.6 family — one family, two product tiers, a disclosed limitation; the drafting family is Claude) reviewed four items the project's own AI-assisted process had drafted, red-teamed in-house, and adopted or prepared for adoption. None of the four was accepted unchanged: one unpublished item was withheld (never entered the served state), two were revised (retained after a specified correction), and one working-layer item was retired (removed from the relied-upon working state). The reviews included mixed subsidiary findings — including determinations that were upheld; the tally records final item-level dispositions. The in-house passes had caught first-order defects across these items — overclaiming, contested sources, circular rescues. Round 2's second-stage review named the recurring pattern the in-house passes had not eliminated, and the project records that diagnosis as second-order repair: a correction that changes the description of an inference without supplying its missing premise — "falsified" softened to "barred" over the same unbridged gap. In these rounds, reviewers outside the drafting family detected that pattern after the in-house passes had not. A repair can be a coherence without contact.

Each row states what was reviewed, what was found, and what changed, so the tally is checkable at item level; the full transcripts are retained in the project record, with the reception ledger's entries as their structured summaries, carrying reviewer metadata per entry ("unrecorded" where not captured). This section claims nothing about the framework's metaphysics: review outcomes bear on the project's process, and transfer nothing to any claim about folds, orientations, or experience.

How findings are graded #

Because AI output cannot be trusted on fluency, everything it surfaces is tagged with how far it has actually been verified. The tiers used in the research notes:

• Full-text-verified — the primary source was read; the quote, the number, the claim were confirmed against the page.
• Search-index-grade — the citation and its gist are cross-corroborated across authoritative indexes, but the primary text was not opened. Metadata is reliable at this tier; exact wording and fine-grained claims are provisional.
• Single-source or grey-literature — one preprint, one listing, or one unconfirmed reference. Lowest confidence; flagged for confirmation before it is allowed to carry weight.

When a finding cannot be pushed to the tier it needs — because a source is paywalled, or a tool is unavailable — the note says so and marks the gap as outstanding rather than papering over it. Naming the verification debt is part of the method, not an embarrassment to hide.

The multi-agent method #

Larger research questions are run as a small pipeline rather than a single query. A question is decomposed into many search angles run in parallel; each candidate finding is then handed to a separate agent briefed to refute it, not confirm it; a final pass asks what is still missing — a population not searched, a claim not verified, a source not read. The output is a set of findings that have survived an adversarial pass, each with its confidence tier attached.

The same shape is used for reading the framework itself. The reception ledger records rounds in which AI models were commissioned as adversarial readers — briefed to attack the site with no niceties — and their located critiques are published whether or not they were comfortable, with their provenance disclosed each time. One such round is what renamed the framework's central term in v2.0: the readers found the move the framework's own guardrail forbade, and the guardrail fired.

What AI is never allowed to do here #

The hard lines, stated so they can be held to:

• AI output is never spent as authority for a metaphysical claim. A model finding the framework plausible transfers nothing to it.
• No AI sets a confidence tier or a confidence number by fiat. Those are the author's, and they live in the claims ledger.
• Fluent agreement is never corroboration, from a model or from anyone.
• Nothing an AI drafts ships without the author's judgment. Assistance is not authorship of the claims.
• The firewall applies to machine-made coherence too: a result on the coherence-and-measurement side is never transferred to the metaphysics, and that includes results produced with heavy AI involvement.

Limits, named #

The honest weaknesses of this method, stated plainly:

• The AI reader rounds have been self-commissioned and, so far, single-vendor — the commissioned readers have all been models from one family. That is disclosed wherever their results appear, and it is a real limit on what those rounds establish; an outside-lineage round is the standing next step.
• Some verification remains search-grade. Where a note relies on sources whose full text was not read, it says so, and those citations are held provisional until confirmed.
• AI can fabricate citations. Plausible-looking references that do not exist, or that do not say what they are claimed to say, are a known failure mode — which is why recent or unusually precise citations are flagged for confirmation rather than trusted.
• The tool can be persuasive in the wrong direction. An assistant can make a weak argument feel strong to the person leaning on it. The defence is the same discipline the framework preaches, applied by hand: ask where it breaks, not whether it reads well.

How to check this #

The point of stating the method is to make it checkable. The verifiable layers:

• the claims ledger and its machine-readable form, where every claim carries its tier, its falsifier, and its nearest prior owner;
• the reception ledger, where commissioned AI critiques appear with their provenance and the revisions they caused;
• the Swarm Instrument, where an AI-measurement result was published and then self-corrected in the open, with the data;
• the container programme corpus, published with its adversarial-review history attached;
• and the critique form — the most useful thing you can send is the located point where any of this breaks.

If this page is doing its job, it does not ask to be believed. It asks to be checked.