A lab runs end to end from a question the product thinker asks

Press Release

A product thinker names a question an experiment could answer, and abcd runs the lab end to end: it drafts the intention (the hypothesis, what would show it wrong, the measures and the stop conditions) for them to confirm, prepares the world, briefs the session that runs it, keeps the log as the experiment goes, and brings the harvest back to the project's record for them to adopt.

Filed by the facilitator on 2026-10-04 at the product thinker's request, verbatim: "Do we need to file anything so that we can build a lab session automation once the manual run for pge is complete? In future, I want to use labs more often and ideally automated, too. We only need an intent or something so we won't forget what we already had to setup manually." To be shaped after the downstream brief lab's harvest.

Why This Matters

Facilitator-written.

abcd's lab verbs (itd-2609212137128014: mint, preflight, record, sweep, harvest) mechanise the lab's stages, but a person and an agent still drive every step, and a lab in a project other than the one asking, or in the real project rather than a throwaway copy, is set up entirely by hand. The product thinker wants labs used more often, and ideally run on their own. This draft keeps the record of what the first such run needed by hand, so the automation is built from what was actually done.

What the downstream brief lab of 2026-10-04 needed by hand, in order:

1. The question came from a design discussion here (the brief's tense), not from the lab: the product thinker's hypothesis was recorded verbatim in the decision log before anything ran.
2. A state-of-the-art pass ran first and fed the lab (research note 2026-10-04-brief-to-record-traceability-sota).
3. The facilitator wrote the intention by hand: the question, the hypothesis, four falsifiers, the measures, the stop conditions and the harvest's destination.
4. The world is the downstream project itself, not a snapshot, which the lab discipline does not yet cover (iss-2610040758486433); this was recorded as an operating constraint.
5. The abcd version was pinned once for the whole lab, chosen against a release cut the same day.
6. Mechanics abcd lacks were specified for hand-running (section markers, realises links with fingerprints, a coverage table, a suspect set per brief edit).
7. The session that runs the lab is another project's session, briefed by a cross-session message and a local brief file; the product thinker works through it with that session over several days.
8. The record rules (no project names, no local model server name, no absolute paths) were restated for the lab's filings.
9. The harvest returns to this project's record, where the drafts it tests (itd-2610040817105016, itd-61) are revisited.

Typed links: builds on itd-2609212137128014 (the lab verbs); held to the discipline itd-2609251624540864; the real-project lab gap is iss-2610040758486433.

Decisions

1. 2026-10-04, the product thinker: file it now so the manual setup is not forgotten; shape it after the downstream brief lab completes.

Mechanism

Prompted (the claim-recording gradient): why the authors expect this to work, as a falsifiable "we expect X because Y" — not the outcome restated. Replace this line with the claim, or with the exact token None stated. alone on its line to record the claim as considered and declined.

Scope Conditions

Required (the claim-recording gradient): the population, platform, scale, or assumptions this claim holds under, one per top-level bullet — abcd intent plan stamps each with a persistent identity. Replace this line with those bullets, or with the exact token None stated. alone on its line.

Acceptance Criteria

Required (the itd-1 discipline): add at least one Given-When-Then bullet describing the verifiable bar for "shipped" before this draft can be planned.

SOTA

Surveyed on 2026-10-23. Repeats, power and the "re-runnable, not reproducible" stance for a lab's measures are in the model-matrix SOTA note. The durable step record a long unattended run needs is in the autonomous loops SOTA note.

• The intention the facilitator wrote by hand (hand step 3) has a standard shape. AsPredicted's pre-registration asks eight questions: whether any data has been collected already; the main question or hypothesis; the dependent variables and how they are measured; the conditions; exactly which analyses; how outliers and exclusions are handled; what determines the sample size; and anything else, such as exploratory analyses (AsPredicted example form). The record's intention already covers hypothesis, falsifiers, measures and stop conditions. The form adds three fields a drafted intention should carry: "has the world already been looked at" (which matters most for a lab in the real project, hand step 4), the analysis fixed in advance, and the exclusion rule.
• End-to-end automated research works poorly without a person at the stages. An independent evaluation of an "AI Scientist" system found that "42% of experiments failed due to coding errors", that "some papers contained hallucinated numerical results", and that its literature check "often misclassif[ied] established concepts … as novel" (Beel, Kan and Baumgart, arXiv 2502.14297). Agent Laboratory reports that "human involvement, providing feedback at each stage, significantly improves the overall quality of research" (Schmidgall et al., arXiv 2501.04227). This informs the "what runs unattended" question, which stays open until the downstream harvest per decision 1: the evidence comes from paper-writing systems, not pinned-version lab runs, and it suggests two human gates, confirming the intention and adopting the harvest, with a failed or empty stage reporting itself loudly rather than producing a result. The product thinker's wish for labs "ideally automated, too" is the tension the harvest has to settle.
• The state-of-the-art pass (hand step 2) is the stage most prone to a confident wrong answer. The novelty misclassification above is that failure. An automated lab keeps the pass's cite-only-what-was-opened rule and its "not verified" list.
• A harvested number needs its repeats stated. The model-matrix note finds spread within one model-and-harness pairing larger than spread between pairings, so a lab's stop condition names a repeat count from a pilot, not a single run.

Path 2: native, on the shipped lab verbs, with the intention drafted as a pre-registration-shaped record the person confirms, the run's stages as durable steps, and two human gates (intention, adoption) as the default the evidence suggests, to be confirmed after the downstream harvest per decision 1. No dependency. The independent fit-challenge runs at plan time, after the downstream lab's harvest, per decision 1.

Open Questions

• Which of the nine hand steps abcd drafts for the person to confirm, and which it performs.
• How a lab in another project is started from this one, and how its harvest comes back.
• A lab in the real project: what replaces the snapshot's isolation, and what makes it safe to discard.
• What runs unattended, and what always waits for the product thinker (the intention, the adoption of findings).
• What the downstream brief lab's harvest says about each of the above.

Audit Notes

Empty. Populated by intent-auditor when intent moves to shipped/.