Pass 4: The Scientific State Compiler — A Versioned Model of What We Understand Author: Daniel Kliewer Date: 2026-07-23 Tags: ai-agents, compile-time-ai, knowledge-compiler, local-first-ai Description: Pass 4 of the Scientific State Compiler: we took the map of scientific uncertainty and promoted it into a versioned state of a field — accepted, uncertain, and contested models, with the disagreement preserved. Local-first, running on a 35B model on our own hardware. --- Yesterday we built Pass 2, the **Scientific Question Compiler**, which turns a corpus into a map of what humanity does *not* yet know — the ranked open questions, the contested claims, and the silence where the literature simply does not speak. Today we ran **Pass 4: the Scientific State Compiler**. It takes that uncertainty map and promotes it into something stronger — a *versioned state of a field*: a structured model of what is understood, separating what is **accepted** from what is **uncertain** from what is genuinely **contested**, and preserving the disagreement rather than flattening it into a false consensus. A summary tells you *what the paper said*. A state tells you *what the field believes, and how sure it is.* That distinction is the whole point. ## What "state" means here The roadmap calls Pass 4 the unifying layer: *Knowledge + Evidence + Mechanisms + Questions + Uncertainty = Scientific State*. The output is a single artifact, `scientific-state.json`, with five components: - **Accepted models** — claims with broad support, each carrying a confidence score and the basis for it. - **Uncertain models** — plausible but incomplete; the open issue is named. - **Contested models** — genuine disagreement, recorded as position A vs position B, *never* collapsed into one. - **Conflict ledger** — the specific unresolved disputes, with a resolution status (usually `none`). - **Open questions + silence summary + major changes** — the residual unknowns and the directional shift in the field. The model itself (a 35B parameter model, running on hardware we own) described the single most important structural fact about the field: > The field has shifted from a strictly amyloid-centric paradigm to an integrated > model incorporating metabolic and vascular dysfunction as early contributors, > though the temporal hierarchy and causal primacy between these pathways remain > actively contested. That is the kind of sentence a synthesis *should* produce and a retrieval search *can't*. It is the state. ## The versioned payoff The state is committed to a local-first versioning layer (Scientific Git-lite, reused from the Discovery Compiler). Each compile snapshots the artifact with a content hash and a real `diff` against the previous version. Re-running with new sources — a fresh batch of papers, a new clinical trial — produces a new commit and a diff showing what changed in the field's understanding. Nothing leaves the machine unless you push it. This is "observability becomes the OS" made concrete: the artifact is not a frozen file but a *versioned research substrate*. The loop compiles understanding, and the version history is the record of that compilation. ## What it took (and what broke) Same local-first constraint as Pass 2: a reasoning model emits a long chain-of-thought before its answer, so if you cap `max_tokens` too low the JSON answer gets truncated and the parser finds nothing. The fix is per-pass token budgets that leave the CoT room to breathe. The state pass runs at 9000 tokens and produces clean, parseable output. One call, ~12 minutes, on a 35B model — slow, but sovereign. ## What you can look at The explorer now has a **State** view alongside the Uncertainty Map and the ranked Questions: - Live: **[scientific-question-compiler.vercel.app/state](https://scientific-question-compiler.vercel.app/state)** It renders the accepted / uncertain / contested split, the conflict ledger, and the open questions — all generated by the model from a three-document corpus on *Alzheimer's & metabolic dysfunction*. ## Where the program stands Five compiler directions — Mechanistic Reasoning, Scientific Question, Scientific State, Discovery, and the Unknown Detector — are not competing projects. They are passes in one architecture that compiles the scientific corpus into an executable model of human understanding. We have now shipped: - **Pass 2** — the map of what we do not know. - **Pass 4** — the versioned state of what we do. The next move is to make the loop *turn on itself*: feed the open questions back as research queries, re-compile, and watch the state versions advance. That is when the compiler stops being a tool and becomes infrastructure. You are looking at Pass 4 of the Scientific State Compiler, running on hardware we own. ## What this makes possible next With Question (Pass 2) and State (Pass 4) both producing real, versioned artifacts, the feedback loop closes: the open questions in State become the next corpus's research targets. The compiler begins to direct its own next compile. That is the moment the loop — not the model — becomes the product.