Say what it does not show
doesNotShow is a required field on every preliminary finding. A result that cannot state what it fails to establish has not been thought through carefully enough to publish.
The problem
Agents write the analysis now. You still sign the submission.
An agent ran the pipeline, produced the figures, and drafted the methods. The claims are probably right. But the person accountable for them did not personally derive them, and there is no structural record of which claims are measured, which are planned, and which quietly lost their evidence three revisions ago.
Writing tools make that faster. They do not make it accountable.
How it works
Three rules, enforced by the type system rather than by review.
Not lint warnings and not a checklist. A record that breaks one of these does not compile, so the discipline survives whichever model, agent, or collaborator produced the text.
Claims carry state
draft → registered → verified, in order, never skipping. A claim reaches verified only with a resolving artifact behind it.
Planned is not zero
A milestone you have not run yet renders as pending, never as a measured zero. Absence and a null result are different findings and the record keeps them apart.
/** One preliminary observation. */ export type PreliminaryFinding = { id: string; label: string; whatExists: string; supports: string; // Required. Not optional, not a convention, not a review step. doesNotShow: string; };
One record, four views
Write the evidence once. Compose it for whoever is reading.
| View | Audience | What it does with the same claim |
|---|---|---|
| Paper | Reviewers | Full register: denominators, intervals, failure accounting. |
| Plain language | Anyone | Ordinary words, same numbers. If it cannot be written without hedging, the claim is weak — that is the point. |
| Grant | Sponsor | Aims and go/no-go gates, with preliminary findings kept visibly apart from proposed work. |
| Deck | A room | Slides built from the record, with the limitations slide generated rather than remembered. |
Because every view reads the same record, a claim cannot appear in the paper without also being accountable in the talk. Change the record and all four move together.
Presentations
The deck is a view of the evidence, not a copy of it.
Slide identity is stable, so inserting a slide does not move everyone else’s comments. Reviewers comment in the deck and export the thread. Decks are self-contained HTML: they open offline, print, and export to PDF and .pptx when a venue insists on a file.
We are connecting data deposits on conductscience.org with paper releases and a presentation for each version. Automatic generation and deposit linking are in development.
Explore the Publishing workflowWho it is for
Teams where overstating a result has a cost.
Regulated writing
Medical affairs and regulatory teams who need the technical and plain-language versions to agree, and who carry real liability when a claim outruns its evidence.
Grant teams
Consultancies and labs that need preliminary data kept visibly separate from proposed work, with provenance for every number.
Agent builders
A schema an autonomous agent can write into, where malformed evidence fails the build instead of reaching a reviewer.
Labs with a long memory
Groups whose claims are revised across years and who need to see what changed between two versions of the same argument.
Questions
Reasonable objections.
Does my data have to leave my machine?
No. Records carry claims, summary numbers, figures and content hashes. Raw datasets, analysis scripts and credentials stay where they are. That is a design constraint, not a setting.
Which AI agents does it work with?
Any of them. Records are typed files in your repository, so anything that can edit a file and open a pull request participates — Claude Code, Codex, Cursor, or a script. There is no plugin to install and no bet on which agent wins.
NIH says work substantially developed by AI is not my own. Where does this leave me?
On the better side of it. The agent writes; the record keeps commit-anchored provenance for every claim and number, and the git history shows who approved what. The audit trail is the product.
Is this a writing tool?
No, and it will lose to one if you want faster prose. It constrains what your text is allowed to claim. Keep writing wherever you like — this holds the evidence underneath.
What does it cost?
Nothing yet, because it is not for sale yet. It is in use on live NIH submissions and manuscripts. Accounts are invite-only until the record store supports more than one tenant.
Look at the worked example.
A fictional study across three releases, with claim states, decision gates, and limitations that survive into the plain-language view. It takes about two minutes.