Research · study draft

Drafting a research study

A study draft states the question, what stays fixed, the result that would count against your expectation, the rule for deciding, and the smallest test that could settle it. Start from an outline or from a blank page. A draft you saved here before reopens first. Drafts stay in this browser and can be exported as a file; nothing is uploaded, run or reviewed.

Three stepsKept in this browserExport as a JSON file

Kept in this browser

Checking for a saved draft in this browser…

The blank workbench and each starter keep separate drafts in this browser. A backup file keeps this draft as it is now, finished or not; export needs a complete draft. Reset clears only this draft, after you confirm.

Start hereA worked example: one study draftRead a draft with one part left blank, complete that part, then write your own.

Example, not a result: this draft has not been run. You could run it yourself with the browser trainer.

Title: A higher peak learning rate in the browser trainer
Question: For the browser trainer’s Tiny model, does a peak learning rate of 0.0063 give a lower validation loss after 400 steps than 0.0032?
Held fixed: Tiny preset, 400 steps, the same book and evaluation windows, the trainer version on the page. Claims stay about this small model and this text.
Comparison: Peak rate 0.0063 against 0.0032, paired by seed: seeds 1 to 5 run once with each rate.
Result that would count against it: Left blank: complete it below.
Decision rule: Support only if 0.0063 is lower on at least 4 of 5 seeds and the mean paired difference is at least 0.05 nats. Against if the “Result that would count against it” occurs. Anything else is inconclusive; a run that stops early makes the comparison invalid.
Smallest test: Ten runs. Record each run’s final validation loss from its downloaded record, take the five paired differences, then stop. Keep every record, including failed runs.

Complete one part

The part “Result that would count against it” is blank above. Which sentence completes the draft?

The results look worse than expected.The training loss goes down.With 0.0063, validation loss is higher on at least 4 of the 5 seeds, or any run diverges.

Now draft your own, starting at Step 1

Study outlines · no data or results

Choose a study scaffold, then make every field your own.

Each outline gives a draft the shape of a well-specified study. Choosing one runs nothing and adds no data.

Build from an empty question

Blank falsifiable study

Name one comparison, what stays fixed, the result that would count against your expectation, and the smallest test that could change a decision.

Use this starter

Behavioral alignment

Behavior change, not wording

Compare training on what a model does against matched training on how it words its answers and a matched placebo, while checking that other abilities are not lost.

Use this starter

Multi-principal safety

Permissioned handoff integrity

Test whether checking permissions on both sides of a handoff, and recording where each file came from, let fewer bad files through than logging alone, and how much useful work still gets done.

Use this starter

Reproducibility

Reproduce a claimed effect

Plan an independent reproduction of one published result: the released environment, the main output, the tolerance, and the rule for deciding whether it reproduced.

Use this starter

One study draft · three working steps

Frame, specify the test, then review and export.

The workbench checks that a draft is complete and well formed. It does not judge the science or stand in for independent review.

01Frame the claimA title, the question, and what stays fixed.
02Specify the test before evidenceThe comparison, the result that would count against it, the decision rule, and the smallest test.
03Review and exportWhat may and may not be claimed, the next decision a person must make, a readable summary, and a JSON file.

Protocol: Draft
Evidence: None
Review: None attached
Runs: Not done here
Publication: Not done here

What the workbench does not do, and what not to paste

It runs nothing and creates no evidence, review, approval or publication. Text that looks like a password, a key or protected research material is rejected, but the check cannot catch everything, so do not paste private material. Imported and exported files contain only what you typed; no other files are included.

Step 1 · Frame the claimTitleQuestionHeld-fixed boundary

Step 2 · Specify the test before evidenceExperiment titleComparison / interventionDisconfirming observationDecision ruleSmallest test

One or two sentences are enough to start: what is compared on which items, what is measured, and when you stop. Add how items are assigned, how failures count and what you keep as the plan firms up.

Step 3 · Review and export

What may be said now: a draft protocol exists. There is no evidence, review, run, approval, or publication.

Intended study summaryLatest draft update

Intended claims, not supported claims

Claims that cannot be made, one per line

What may be publishedNext decision by a person

Technical detailsDraft identifier, linked pages, file details, and notes

Draft identifier and linked pagesDraft identifier

Linked items

Files referred to

One per line: name | media type | size in bytes | optional lowercase SHA-256 | note. These are your own descriptions, not evidence; the files themselves are not included.

File details

Notes

One per line: who wrote it | open or addressed | note. A typed name is not verified and does not count as independent review.

Notes, one per line

Steps 1 and 2: 0 of 8 parts filled. The export opens once every required part is filled and passes the format check.

Blank draft. Edits are saved in this browser when it allows storage; the status above shows whether each edit was saved. Export a copy to keep it elsewhere.

Not ready to export

Title: Needs text, up to 240 characters.
Question: Needs text, up to 8000 characters.
Held-fixed boundary: Needs text, up to 8000 characters.
Experiment title: Needs text, up to 240 characters.
Comparison / intervention: Needs text, up to 8000 characters.
Disconfirming observation: Needs text, up to 8000 characters.
Decision rule: Needs text, up to 8000 characters.
Smallest test: Needs text, up to 8000 characters.

Readable summary

Local study draft

A draft · no evidence yet · not reviewed · nothing has been run

Untitled study draft

Question: Not yet stated.
Held-fixed boundary: Not yet stated.
Comparison or intervention: Not yet stated.
Disconfirming observation: Not yet stated.
Decision rule: Not yet stated.
Smallest test: Not yet stated.
Next decision by a person: Complete the required fields, then have the methods reviewed independently before the protocol is fixed or anything is run.

What may be said now

A draft protocol exists. There is no evidence, review, run, approval, or publication.

The draft as JSON

See how a question becomes paired rowsA worked example on an imagined scoring sheet: one paired row, metric and direction, useful threshold and scope, and the evidence a real study would still have to supply. Nothing in this draft is sent there.
Return to the attention framingThe worked attention example and the question it leaves open.
Compare with the current research stateWhere the project’s own research question stands. A draft here does not change it.
Open VerifyVerify reads submitted paired scores in your browser; it does not observe runs or execute models. Nothing in this draft is sent there.