Blank falsifiable study
Name one comparison, one held-fixed boundary, one falsifier, and the smallest test that could change a decision.
Protocol starter · Evidence none · No run performed
Use this starterResearch
Turn one question into a test someone else can inspect.
Learned or varied attention? Separate the calculation you can explain from the question you still need to test. Start with a worked framing, then open a separate evidence-free local draft.Frame a question from attention Available here: page-memory drafting and manual JSON export, not study execution or independent review.Inspect the current research state
Current conditional objective
SBJ-001 is the sole conditional scientific objective, not an admitted study. Its only executable work is the model-free SBJ-001-P0 feasibility preflight; no protocol is locked, no participant study is executing, and no scientific evidence or result exists.
Mapped · P0 feasibility only · as of 2026-08-16
Source-bound judgment · SBJ-001
Among predeclared, domain-qualified reviewers making bounded promote / hold / reject decisions on independently adjudicated AI-generated ML or safety evidence, and among case families where a matched full derangement is constructible, what is the causal decision cost of complete claim-to-source misbinding when mechanically valid claim-level exact-source-span pointers are compared with the same pointers reassigned by a fixed-point-free, label-blind, within-case full derangement, while report assertions, artifact and source bytes, navigation shell, and review budget are fixed?
Current claim boundary: The objective is conditional on case families where a matched full derangement is constructible. It cannot claim a generic provenance-interface benefit, citation-validity result, behavioral-verification result, expert-use effect, or platform efficacy.
P0 failure boundary: The P0 must return KILL or REVISE if the direct-prior, exact-source, rights, treatment-invariance, ethics, reviewer-route, endpoint, power, or resource gates cannot be closed without acquiring study materials or involving participants.
Scientific state: mapped / none / none / none; no study is executing.
Exact next gate: Complete only the model-free SBJ-001-P0 feasibility record and obtain a disposition from a named human who did not design or implement the preflight. Even ELIGIBLE-FOR-SEPARATE-PROTOCOL-ADMISSION would not admit a study.
How this record fits the project
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.
Verify — inspect submitted matched evidenceBrowser-local · current product wedgeReads paired scores and declared design assumptions after a run. It does not observe, execute, certify, or persist the experiment.SBJ-001 — decision cost of complete source misbindingConditional objective · P0 feasibility only · evidence noneSBJ-001 remains mapped / none / none / none. SBJ-001-P0 is the sole executable research preflight and may test only whether a future model-free, zero-copy qualified-human study is ethically, legally, statistically, and operationally specifiable. No participant study, protocol lock, independent human review, result, or publication exists.Open question map — ideas before admissionOpen to critique · not an execution queueNew questions, objections, and alternative framings may be mapped and compared. A named human decision is still required before one becomes an active study.
Human-controlled, cumulative researchDirection · not current capabilityLearning, explicit predictions, governed experimental labor, evidence inspection, and reviewed carry-forward should eventually form one loop while humans retain authority over questions, thresholds, interpretation, and publication.
Learn → Question → Assumptions → Local draft
What question does your attention change leave open?
Separate the mechanism you can explain from the claim you still need to test.
Hold one attention row fixed: query q = [1], keys kA = kB = [0], and scalar values vA = 2 and vB = 8. The key dimension is 1, so both scaled dot-product scores are 0. Change only which sources are allowed.
| Allowed sources | Weights (A, B) | Output |
|---|---|---|
| A and B | (0.5, 0.5) | 5 |
| A only; B masked | (1, 0) | 2 |
Check why the output changes
Each allowed score contributes exp(0) = 1 to normalization. With both sources, each weight is 1 / (1 + 1) = 0.5, giving 0.5 × 2 + 0.5 × 8 = 5.
Exclude B before normalization: A gets weight 1 and B gets exactly 0, giving 1 × 2 + 0 × 8 = 2. Setting B’s already-zero score to zero would not mask it. At least one source must remain allowed.
Toy result, not research evidence. The mixture changed from 5 to 2. No model was run, and this does not show that an answer became better.
- Question · still open
- For one fixed model and task, does masking a distractor improve answer accuracy without harming cases that need that source?
- Assumptions to state
- Define a distractor before inspecting answers. Fix the model, inputs, mask location, and decoding; name the task and the limits of the claim.
- Discriminating test · proposed, not run
- Plan matched masked/unmasked cases and a control where B is needed. Predeclare accuracy, a useful-change threshold, stopping and invalidity rules. Worse or inconclusive outcomes must remain possible; a lower toy output is not higher accuracy.
Continue in the existing local workbench
Write the question in the workbench; edits live only in that page’s memory. Export JSON before leaving or refreshing. Drafting runs no study, creates no review request, locks no protocol, admits no evidence, approves no claim, and publishes nothing. Human review happens elsewhere; model advice is advisory, never independent named-human review.
One question · five honest capability phases
Research is a chain of decisions, not a run button.
Keep the question visible while the instrument changes. Every stage leaves an inspectable artifact and ends at a human gate; negative, inconclusive, and invalid outcomes stay in the record.
02 · Available here
Commit
What comparison, held-fixed conditions, measurement, falsifier, and decision rule will you state before seeing an outcome?
Draft the study- Artifact you leave
- A versioned local study draft and deterministic JSON representation.
- Human gate
- Approve the draft for independent methods criticism. This site cannot approve or lock it.
- Later bounded help
- Completeness checks, alternative controls, and protocol criticism may help; approval and lock remain human decisions.
- Next handoff
- Export the draft for independent methods review and separately authorized execution.
- Capability boundary
- The browser-local workbench drafts and format-checks the packet. It uploads nothing and creates no evidence.
Choose the shape of the first question.
Each starter is editable, browser-local, and deliberately empty of evidence. It gives you a stronger first draft, not an admitted study.
Behavior change, not wording
Compare action-grounded behavior against matched language-only and sham controls while protecting collateral utility.
Protocol starter · Evidence none · No run performed
Use this starterPermissioned handoff integrity
Test bilateral permissions and content-addressed lineage against metadata-only logging, then measure bad-artifact acceptance and useful completion.
Protocol starter · Evidence none · No run performed
Use this starterReproduce a claimed effect
Specify an exact independent reproduction against the released environment, primary output, tolerance, and decision rule.
Protocol starter · Evidence none · No run performed
Use this starterHow to read the state
Work state, evidence disposition, and validity scope are separate. Completing a study does not make its hypothesis true; a clean contradiction or inconclusive result is as visible as support. Sources are provenance, never evidence.
| Field | Value |
|---|---|
| Source-mapped · not admitted | A topic and exact admission gate are recorded. No bounded study, protocol, run, or result exists. |
| Protocol draft | A bounded draft record exists. Current authorization and sequencing must be stated separately; no model call or result claim is allowed. |
| Preregistered · locked | A named version is frozen and dated before eligible evidence collection. Locking does not imply execution or support. |
| Evidence available | Actual artifacts receive a bounded supported, contradicted, inconclusive, or invalid disposition. A source map never qualifies. |
Source-mapped, not admitted
These are possible future directions, not active programs. Each needs the stated human admission decision before protocol work begins.
Behavioral Specification and Human ControlSource-mapped · not admittedExplicit specifications, adaptive-data policies, and human-AI interfaces under a named context shift. Missing decision: Select one specification, context-shift population, comparator, machine-behavior outcome, and human-control decision outcome that can be studied offline.
Multi-Agent Systems SafetySource-mapped · proposal notePopulation-level coordination, failure, attribution, and control in realistic multi-principal agent networks. Missing decision: Identify an existing safe, reproducible testbed or trace corpus with distinct principals, system-level units, a non-single-agent threat model, and an external-validity map.Public Evidence and Technical GovernanceSource-mapped · not admittedEvidence that could change one named risk-management, audit, governance, or public decision. Missing decision: Name the decision-maker, decision, uncertainty to reduce, lawful fixed corpus, analysis method, conflicts, and causal-versus-descriptive boundary.
Methods and prior work
The records below provide methods depth and learning context. They are not additional active programs in the current research agenda.
Atlas-64
Attention serving systems
Methods track · no benchmark claim
A public learning and methods track for cache, batching, and serving tradeoffs. It does not claim production superiority or a completed institutional evaluation.
| Field | Value |
|---|---|
| Identifier | ATTN-SERVING |
| Work state | Concept and protocol development |
| Evidence disposition | None at program level; no model result is published here. |
| Exact next gate | Attach a governed experiment and admissible evidence before promoting any performance claim. |