Research with AI, without giving away human judgment.

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: What is the causal decision cost of complete claim-to-source misbinding when valid claim-level exact-source-span pointers are replaced by a matched full derangement while the report, artifact and source bytes, navigation, and review budget stay fixed?

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.

Archit Khare is an applied researcher and machine-learning engineer with more than seven years of experience across deep learning, foundation models, language systems, information retrieval, model fine-tuning, recommendation systems, and large-scale experimentation. His work spans applied research and production machine learning, with an emphasis on turning research ideas into reliable, testable systems. He is currently developing Continuous Function, an independent research project focused on careful AI evaluation, human oversight, and inspectable research methods.

founder@continuousfunction.ai · LinkedIn · GitHub

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. SBJ-001 is the sole conditional scientific objective at mapped / none / none / none. SBJ-001-P0 may test only whether a future model-free, zero-copy qualified-human study is feasible; it cannot admit that study. No protocol is locked, no participant study is executing, no independent human review has occurred, and no empirical evidence exists. AE-001 is retained only as a parked conditional draft whose dated sequencing is superseded. The prior Atlas-64 record is a maintainer-authored internal case withheld from normal production builds. It is not an admitted model result. Any future capability must pass its own evidence and authorization gate before this page describes it as current.

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.

SBJ-001-P0 is the sole executable research preflight. It must fail closed on direct prior work, exact sources, rights, treatment invariance, ethics, reviewer route, endpoint, power, and resource feasibility. A named human who did not design or implement it must disposition the exact preflight; even an eligible disposition would not admit a protocol or study. AE-001 remains parked and is not gated or reactivated by the separate protected Verify source decision. The Research agenda exposes that absence of evidence as prominently as the question and keeps conditional, parked, and future work visibly separate.