Human judgment stays in charge.

Continuous Function is an independent research project for questions that can be made explicit, challenged, and tested. The public record separates a possible direction from an admitted study, a protocol from a run, and a source from empirical evidence.

Research thesis: When an intervention makes a model look safer or better, what evidence shows that the target behavior changed rather than only the surface signal?

What open means here: Open here means inspectable and criticizable: questions, assumptions, falsifiers, evidence states, and next gates are visible. It does not mean public model execution, automatic research intake, or that every suggested idea becomes active work.

These are four existing surfaces in one connected-path hypothesis, not an established end-to-end research environment. Preparation is under way for a future consented manual walkthrough that will test task legibility only: whether a person can build intuition in the Atlas, commit a prediction in an applicable Lab, sharpen the unresolved point into a local Research draft, and identify the appropriate Verify or external-evidence boundary and next human decision without losing the exact object. The walkthrough has not happened, and a maintainer rehearsal is not participant observation, a learning result, or product evidence. Today Verify asks what submitted paired evidence supports; Research records expose what is still only a question, what is blocked, and what exact gate comes next.

Some Lab investigations start with a question you answer before their result appears. On those surfaces your prediction is locked in first, so the calculation has something honest to compare against. Other Lab surfaces ask you to inspect, manipulate, complete, compare, reproduce, or explain without claiming the same prediction gate. There is no score. Verify is different: you may calculate the submitted rows first, but its separate design questions stay visibly unresolved until you answer them.

In those deterministic Lab investigations, each toy result is calculated in your browser from the settings displayed; it is not a measurement from a production model or server. Research records are different: they separate proposals, retained evidence, and publication boundaries. Review-only material stays withheld until its exact source and publication decision are established. Where a page cites a paper or course, the source is linked next to the claim it supports.

Verify creates no account, durable record, receipt, or export. Lab and learning retention is surface-specific: some predictions, saved comparisons, learning-route state, source-boundary handoffs, and object-attached Research action drafts or resolutions can remain in this browser, while other surfaces save nothing. None is shared or canonical research memory, and those records are separate from Verify.

A Research Workbench study draft exists only in page memory and is lost on navigation or refresh unless you manually export it. That manual export is a downloaded local JSON file under your control, not site-held state. Clearing this site’s browser data removes the browser-stored records, but it does not remove files you downloaded.

Research pages distinguish source maps, protocol drafts, locked protocols, evidence dispositions, and publication decisions. AE-001 is admitted, but protocol work is parked behind two gates: the protected exact Verify founder-decision P0 must close, and AE-001 must be explicitly selected as the sole cycle objective. It is not under development now. It has not received independent human methods review, been preregistered, or been run, and no empirical evidence exists. The Atlas-64 evidence record remains a maintainer review preview and is withheld from normal production builds; no Atlas-64 model result is admitted. When any of this changes — an assistant, an account, an authorized run, or a research result — it will be described here before it appears anywhere else on the site.

Learning, sources, predictions, protocols, evidence, decisions, and reviewed corrections should stay attached to the same research objects. That continuity could let later teaching and separately approved research cycles build on what survived review instead of beginning from an unexamined answer. AI may eventually propose, challenge, and carry out bounded delegated work; humans retain authority over the question, assumptions, evidence standard, interpretation, decision, and publication.

If real use earns the required gates, those objects could support human-human, human-agent, and permissioned agent-agent collaboration and, over time, an open research institution. Its knowledge commons would consist of inspectable, human-reviewed objects and explicit disagreements — not a feed of conversations or unreviewed agent assertions.

None of that collaboration infrastructure exists today. There are no accounts, identities or permissions, shared community, live on-site AI, agent orchestration or execution, canonical shared memory, independent-review workflow, publication connector, or automatic curriculum change. The role lenses shown in current object rooms are deterministic perspectives, not people or contributions. Research pages can orient a journey; the Workbench can edit a page-memory draft and manually download or import JSON. It creates no run, evidence, approval, durable research record, or publication.

Verify now has explicit metric direction, matched reruns, fail-closed design questions, and no receipt promise. Its immediate gate is a founder approve-or-reject decision on the protected exact source tuple. Approval would not itself create a candidate or authorize deployment; those mutations and the required independent human reviews need separately admitted work. Only after an exact deployed build and the remaining release and accessibility gates should someone other than the maintainer bring a real result and test whether the utility changes an actual decision.

AE-001 has an admitted protocol-forming record, but all further protocol work remains parked until the protected exact Verify founder-decision P0 is closed and AE-001 is explicitly selected as the sole cycle objective. It cannot run until its model, units, outcomes, numeric margins, analysis, exclusions, stopping rule, rights, budget, and independent human methods review are fixed and the protocol is locked. The Research agenda exposes that absence of evidence as prominently as the question and keeps future directions visibly separate from admitted work.