Where the model is. And where it is not.

The prose sections

A model writes them from your sentence. They are a draft in the shape of a protocol: no concentration, volume or incubation time in that text was looked up, measured, or checked against a source before it rendered. Read it the way you would read a first draft from a colleague who has not run the experiment either.

The editable step graph

stepsFromSections() derives the draggable step graph from the procedure text by rule, not by a second model call. The same prose yields the same graph, and the derivation has its own unit tests.

The score and its 7 signal checks

analyzeRigor() scans the section text for 7 named signals: controls, replication, sample size and power, a statistical plan, quantitative parameters, randomization and blinding, and safety. It reports the share that are present. Pure function, no model, no network, no clock. The count on this page came from calling it on this request.

The rpf_ identifier, the entity counts, the compression ratio

canonicalize() reduces a protocol to its measurable backbone: quantities normalized to canonical units, techniques, reagents, statistical methods, referenced standards. hash128() then content-addresses that form. 5 mL, 0.005 L and 5000 µL collapse to one token, so wording and unit choice drop out. No model, no network, no clock.

The severity verdict and the finding list

A model reads the protocol and returns an overall severity (none, low, medium, high or critical) plus findings in six named categories: controls, sample size, safety, reagents, statistics, feasibility. Nothing is counted to arrive at that severity. It is one reviewer’s judgement, produced by a model, stored beside the protocol and rendered with the protocol.

The records it reads

Your own rows: papers, findings, protocols, read through tools against the database under row-level security. The assistant can see what your account can see and nothing beyond it.

The reply, and any draft it produces

The answer is a model’s judgement over those records. Its own instructions forbid it from claiming that anything was saved, created, applied or ordered: a draft is presented for review, and you are the one who commits it.

Citation metadata for a DOI

Fetched from CrossRef, falling back to PubMed. No model participates in that lookup, so the title, authors and journal are the registry’s values rather than a reconstruction of them.

The methodology pulled out of a paper

resolveGrounding() picks the deepest source available before the model is asked anything: the paper’s stored full text, then the attached PDF handed over as a document block, then the abstract. Which of the three was used is part of the request rather than a preference. When a raw PDF is attached the route forces a model with reliable PDF handling instead of the cheaper one the router would otherwise select.

A synthesis or a gap analysis across several papers

Comparing papers and naming what the literature is missing is a judgement. It is not tied to a span in any one document, and it belongs on your reading list rather than in your results.

The numbers: lane density, cell counts, areas

Classical computer vision running in your browser: Otsu thresholding, Gaussian smoothing, connected components, distance transform and marker-controlled watershed, the same algorithm family as ImageJ and CellProfiler. Pure functions over a grayscale buffer, unit-tested on synthetic images whose right answer is known in advance.

The interpretation panel

A separate route sends the image to a vision model for quality control: what kind of image it is, which artifacts are visible, which analysis mode to try. It does not produce or adjust a measurement. Invoking it is the point at which the image itself is sent to the model provider, and it is a separate action from the measurement.

Every classification, dose and score across 15 engines

Each engine is a pure function carrying the published work it implements, a version, and a SHA-256 over both its constant tables and a fixed grid of its own outputs. The public registry re-runs all of them when you load it and prints whether each still reproduces its recorded hash. A drifted hash fails the test suite, and the test suite gates the build.

The literature corpus behind an evidence score

Real PubMed retrieval through the NCBI E-utilities, esearch then efetch, stored as an immutable corpus snapshot alongside the query that produced it, so the same search can be re-run against the same identifiers.

Each claim taken out of an abstract

The model proposes a claim and the quote that supports it. The claim is persisted only if that quote is verified to be a contiguous verbatim span of the real abstract. A paraphrase is rejected. The verification is a deterministic string check, not a second model, so the model cannot invent a source.

The direction label on a grounded claim, and the case brief

Two judgements, declared rather than buried. Whether a grounded claim supports or refutes the hypothesis is labelled by the model, and that label feeds the deterministic strength engine. The case brief narrates a workup that has already been computed: it is instructed never to emit a number, and the numbers are returned whether or not the brief succeeds at all.

Every calculator result across 61 tool routes

Those 61 routes resolve to 51 computation libraries, and all 61 of them sit on a module with unit tests. Searching that code for a call to a model returns 0. That figure was measured over this tree in August 2026 and covers this tree only.

A standards record

Read from the seeded Standards Library table, which holds at least 490 records across 16 seed migrations. The count is reproducible from the repository without touching a database, because the seeds are the source.

Which standards might apply to your experiment

The standards finder maps a description or an uploaded document to candidate standards. It is a shortlist to go and check, never a compliance determination, and it is the single route in this tree that reaches a model.

Result flags and QC verdicts

A result is flagged against its reference range in code, and control measurements are evaluated against 6 Westgard rules in the database, mirrored by a pure module that drives the live preview and is the tested specification.

The plain-language read of an order

1 of 72 LIMS routes calls a model, and this is the one. It narrates results that were already released, it is instructed not to alter a number, it persists nothing, and it only ever receives rows the viewer is already entitled to see.

The statistic, the p value, the effect size, the interval

The statistical engine runs in your browser on distribution primitives verified against published reference values. Your raw data does not leave the machine in order to produce them.

Which test to use, and what the result means in words

Two separate metered operations: a method recommendation before the test, and a plain-language reading of an already-computed result after it. Both are advice about a number. Neither is the number.