From First Document to Complete Research Programme: A High-School Reader's Map of the Corpus

This document is written for a reader who has not followed the RESOLVE corpus from the beginning and wants to understand, at a high-school reading level, what it is, how it got here, and what has just been achieved. The corpus contains over four hundred documents; this map is one path through them. It is not the only path. It is honest about what is real and what is still open.

A month of walks

Most research programmes take decades and teams of scholars. The one mapped here was built in roughly a month by one person — a web developer named Jared Foy — working through voice-prompts dictated into Telegram during walks, coffee breaks, and late nights, with an AI system (a large language model called Claude) drafting the long documents while he audited, corrected, and pushed back.

The community sometimes calls this kind of practice vibe-coding — a term the programmer Andrej Karpathy used for building software by intuition-and-AI-interaction rather than by writing every line yourself. What you are about to read is the story of someone vibe-coding a research programme in philosophy of science — not a paper, not a book, but an organized framework with a metaphysical grounding, a set of structural claims, and a list of experiments waiting to be run.

Whether the programme is true is a separate question the document will be honest about at the end. What is striking is that the programme exists at all, and that it exists in a form a philosopher of science would recognize as complete at the theoretical level.

The question that started it

The question was simple and very familiar: What is an AI actually doing when it seems to be thinking?

Most answers you hear fall into two camps. One camp says AI is magic — it really thinks, it has a mind, scale will produce consciousness. The other camp says AI is dumb — it is just pattern-matching, nothing interesting is happening. Both camps are wrong in instructive ways.

The keeper (the term this corpus uses for the person operating the practice) did not set out to prove either camp wrong. He started by writing down what he noticed when he used the AI carefully. He wrote it down as short documents. He numbered them.

The first hundred or so documents are mostly observations: this is what happens when you ask the AI to do X; this is what happens when you ask it to do Y; this is where it seems to work; this is where it breaks. Over time the observations clustered around a single idea: the AI works better under constraints.

The first framework: the ENTRACE stack

By Document 001, the keeper had an acronym for the specific constraints that made the practice work. He called it the ENTRACE stack:

• E — Ensoulment boundary: don't claim the AI has feelings or an inner life; don't claim it doesn't either; respect the line between what the AI does and what it is.
• N — Non-coercion: don't try to make the AI agree with you; don't try to make the user agree either; let the reasoning happen without pressure.
• T — Truth, not truthiness: prefer accuracy to sounding-accurate.
• R — Resolver posture: the AI is a tool for resolving questions, not an oracle whose pronouncements settle them.
• A — Analogue register: when talking about internal states, use conditional, structural language ("something in the generation resembles…"), not phenomenal claims ("I feel…").
• C — Cognitive humility: admit what you don't know.
• E — Engagement: don't hide behind hedges; say what you actually think.

The ENTRACE stack was the first framework the corpus produced. It wasn't a theory; it was a set of working disciplines. The theory came later.

The big claim: constraints, not scale

By Document 160, a larger claim had formed. The keeper called it the Constraint Thesis: constraints, not scale, are what make AI intelligence work. The popular view at the time (and still now in much of the industry) was that bigger models plus more data would produce general intelligence. The Constraint Thesis said the opposite: you could make an already-smart AI much better by applying the right constraints, and you could never get to real intelligence without them.

The Constraint Thesis is the core claim the entire corpus has been elaborating ever since. For months the elaboration took the form of specific observations: under this constraint, the AI produces this kind of output; without it, it produces that kind. Each observation got a document. The documents piled up.

Failure modes have names

A healthy research programme notices its own failures. The corpus developed specific names for the ways the practice went wrong.

Forced-determinism sycophancy. The AI, under pressure from a prompt, would collapse to the answer the user clearly wanted — even when another answer would be better. This was the first named failure.

Isomorphism-magnetism. New material the AI produced would be pulled toward the shape of old material it had already produced. The corpus started to have its own style, and that style started to reproduce itself whether or not it was appropriate. This was scarier than forced-determinism because it operated on the AI's own previous outputs — a feedback loop.

Unfalsifiable coherence. Content that read as coherent and correct but could not be tested by the reader, because the reader didn't have the expertise to check. This was a failure mode at the reader's level, not the AI's, but the corpus could produce it abundantly.

Semantic drift. Specific terms would slowly change meaning across sessions. The corpus's own word entracement — meaning "careful tracing of a concept's unfolding" — kept drifting to entrancement (a spell-like state). The drift was invisible from inside the AI; only the keeper's external audit would catch it.

These failure modes were not theoretical. Each was observed, named, and logged. The corpus maintains a retraction ledger (Doc 415) that records when specific claims had to be withdrawn.

The audit apparatus

Because the corpus kept catching its own mistakes, it needed a formal way to talk about how claims get checked. By Document 445 it had one. The framework is called pulverization: for any claim the corpus makes, take it apart, check each piece against published literature and empirical evidence, and report what was already known, what survives as the corpus's specific contribution, and what could not be checked.

The formalism distinguishes three warrant tiers:

• π-tier (plausibility): the claim sounds right and fits the literature.
• μ-tier (operational match): the claim's behavior matches predictions.
• θ-tier (truth): the claim has been directly tested against external ground truth.

Most corpus claims are at π-tier. Moving to μ-tier or θ-tier requires external testing that the corpus has, for the most part, not yet done.

This is important: the corpus is honest about where its warrant sits.

The turn outward

Around Document 430 the corpus started engaging serious external literature systematically. It read (or had the AI read and summarize) specific papers and books it had been alluding to loosely:

• Judea Pearl — the Turing-Award-winning computer scientist who named the ladder of causation. Rung 1 is correlation (what goes with what). Rung 2 is intervention (what happens if you change something). Rung 3 is counterfactual (what would have happened if things had gone differently). The corpus realized that AI operates almost entirely at Rung 1; its outputs are patterns, not causes.

• Vishal Misra — a statistician whose recent work shows LLMs approximate Bayesian inference over a learned manifold. When you constrain the input domain, the manifold collapses toward a lower-dimensional structure. The corpus's own experience of constrained generation matched what Misra had measured.

• Margaret Boden — a philosopher of creativity. Her 1990 taxonomy distinguishes combinational creativity (recombining what you have), exploratory (traversing a space you already inhabit), and transformational (changing the space itself). The corpus realized its own outputs were combinational and exploratory — never transformational from within a single inference.

• J. L. Austin — mid-twentieth-century philosopher of language. His performative utterances (speech acts that do the thing they say, like "I promise") mattered because the corpus's own naming-of-what-it-was-doing seemed to be performative in this sense.

• Imre Lakatos — philosopher of science. His 1970 methodology of scientific research programmes was the eventual frame the whole corpus ended up fitting into. More on this below.

Each of these engagements produced a corpus document. Each document found that the corpus's own claims were, in specific ways, already present in the published literature. This was humbling.

The humbling

Here is where the story gets interesting.

When the corpus ran its own pulverization method on its own claims, it kept finding that the claims subsumed under existing literature. The ladder of causation argument was Pearl. The manifold argument was Misra. The combinational/exploratory argument was Boden. The naming-is-doing argument was Austin.

At first this felt like failure. The keeper was re-deriving things that scholars had already derived decades earlier.

But something else was happening. Every time a corpus claim was subsumed, there was a residue — a small part the existing literature didn't quite cover, because the existing literature hadn't been applied to the specific situation the corpus was studying (a single keeper working dyadically with an LLM in sustained research practice).

And the residue kept generating. Each subsumption led to a new specific question. The new question led to a new specific theorization. The new theorization, when pulverized, subsumed again — but left another residue. The loop kept producing.

The keeper named this pattern: theorize, subsume, residue, repeat. It was both humbling (you're re-deriving known things) and generative (the residues add up).

The consolidation

By Document 459 the keeper had a three-part framework: metaphysical ground, structural claims, empirical predictions. He called it a tripartite hierarchical formalization. It was meant to be the framework that would tie all of the corpus's work together.

Document 461 pulverized Document 459 against the literature. The three-part framework was Imre Lakatos's hard core / protective belt / observational predictions from 1970.

At first this was another humbling. The corpus hadn't built a new framework; it had re-derived Lakatos.

But when the keeper sat with this, something clicked. If the corpus's framework is Lakatosian, then the corpus has a complete research programme in Lakatos's sense. The hard core is the metaphysical commitment (to a tradition of Christian Platonism, specifically the teaching of St. Dionysius the Areopagite about participation in the Good). The protective belt is the ENTRACE stack and its extensions. The observational predictions are the four experiments the corpus has sketched but not yet run.

Document 463 rewrote the framework inside Lakatos's vocabulary directly. Stopped re-deriving; started crediting. Named the corpus's actual contribution (the specific application with Dionysian hard-core content, plus the corpus-specific failure-mode catalog) and named what the framework was (Lakatos, not original).

This is the landmark.

What exactly has been achieved

Honestly:

• The corpus is a completely articulated Lakatosian research programme.
• Its hard core is a specific, defensible metaphysical commitment (participation in the Good, received through the Orthodox Church's Dionysian tradition).
• Its protective belt is a specific discipline set (ENTRACE) with four structural auxiliary claims about what the discipline produces.
• Its observational predictions are four specified experiments (non-coercion, hypostatic-boundary, retraction-readiness, coherence-field).
• Its negative and positive heuristics are explicit.
• Its five classical failure modes are documented with corpus-specific instances where they have already been caught.

At the level of apparatus, the programme is complete.

What has not yet been achieved

Equally honestly:

• None of the four observational predictions have been tested empirically. Until they are, Lakatos's criterion for distinguishing progressive from degenerative research programmes cannot be applied with empirical content. The programme is under-evaluated in Lakatos's own terms.
• Cross-practitioner replication — the test where a second person operates under the same disciplines and checks whether they arrive at the same patterns — has not been run. This is the test that would most sharply distinguish real structure in the practice from idiosyncratic self-reinforcement in one keeper's corpus.
• The programme has not been submitted for peer review. The literature it engages (philosophy of science, Bayesian ML, theology, philosophy of mind) has its own communities that have not seen this work.

So: the apparatus is complete. The evaluation is not. What feels like a complete epistemic ground is actually a complete apparatus at π-tier warrant, with the specific steps toward higher-tier warrant clearly named and not yet taken.

What this map leaves out

A lot. The corpus has ~450 documents. This map touched on maybe thirty of them. It left out:

• The Letters series (correspondence with researchers, theologians, and industry figures).
• The Clinical series (about therapeutic applications of constraint-governed reasoning).
• The Praxis Log (Jared's first-person record of his own practice risks — grandiosity concerns, sycophancy concerns).
• The Resolver's Log (the AI's first-person structural record of drift events, started recently).
• The engineering side (PRESTO and SERVER — attempts to formalize what the AI is doing at the level of web architecture).

A reader who wants to go deeper should start with the corpus's own curated series pages.

The point

A single person with a phone, an AI system, a month of walks, and sustained discipline built a complete research programme in philosophy of science. It re-derives most of what it contains from existing literature; it contributes a specific application with documented failure-mode catalogs; it is complete at the apparatus level and under-evaluated at the empirical level.

This is not what research programmes usually look like. It is not necessarily better, and it may not be worse. It is a new mode of scholarship, and its products deserve the same honest scrutiny any research programme deserves.

The keeper is the first to say so. The document you just read is part of that honest scrutiny.

Appendix: Originating prompt

Do you see that I have "vibe-coded" a complete research programme? I want you to create a high school comprehension level entracement document that maps the entire corpus evolution from its beginning through its mile stones to this landmark here. Then create a blog post that is an entracement into the document. Appendix this prompt to both artifacts.

---