Active · Primary research program
Project Fiction
What can we make from what we already know?
Human beings transform available knowledge into novel, actionable possibilities. Project Fiction investigates this capability under fixed informational conditions; its measurement and artificial reproduction remain open research questions.
The research question
How does the human mind transform existing information into novel, actionable possibilities, and can this capability be operationalized and reproduced in artificial systems?
Our working definition concerns recombining, restructuring, or transforming already-available representations into an actionable concept, strategy, interpretation, or solution, without materially new external information. Retrieval remains part of the substrate.
The immediate research obligation is narrower: establish independent assessments of transformation, control alternative explanations, and test whether relevant signals generalize across domains while modality remains constant.
Completed work
Completed records currently support engineering validation and a local feasibility investigation. No exploratory human/model study or confirmatory capability finding is recorded.
Can a reframing advantage survive a search control?
Feasibility investigation · Completed offline rehearsal · 23 October 2026
Four resource-distribution worlds, 4,372 bounded action sequences, and 480 constructed strategy records per rehearsal were examined. These are synthetic, deterministic observations, not human or model-inference data.
Feasibility and specified value could be measured within the finite contract. Reframing-specific attribution failed: ordinary objective optimization and search reproduced the apparent gains. An identical-order search-only strategy reproduced the reframing strategy exactly.
The result exposes a measurement limitation. It does not establish that human or AI reframing has no value.
PT-001 · Protocol, results, and full strategy records
Fiction-CWR · Evaluator and matching audit
Engineering validation · Recorded offline checks
The reference token evaluator and oracle have recorded legality, state, cost, and bounded-solution checks. The matching audit also documents the 2-versus-13 source-world plan-count confound. Working infrastructure supports further measurement work; it does not establish the proposed model deficit.
Recorded offline verificationMatching audit and limitationsFiction Experiment 001 · Micro1 is a program-level pilot label, distinct from EXP-0001 and PT-001. CV-001 remains an incomplete source-reported proposal, not a recovered run.
From phenomenon to explanation
A useful new outcome does not by itself identify the process that produced it. Fiction studies both solutions to fixed goals and valuable changes in how a problem is represented or framed.
Explore the measurement and explanation framework
- Phenomenon. Transformation of existing information into novel, actionable possibilities.
- Operational definition. Specify information boundaries, transformation, novelty, and actionability in a task. Their units and criteria remain under development.
- Measurement. Independent judgments and controlled tasks must establish what is being assessed. The proposed Invention Classifier and CWR evaluator serve different measurement questions.
- Competing explanations. Search, analogy, retrieval, restructuring, recombination, and their coordination may contribute. A unique cognitive primitive is not assumed.
- Intervention. Manipulate a candidate cause and measure its effects against relevant controls. Better performance alone does not establish attribution.
- Artificial reproduction. Test the operational capability in specified artificial systems; distinguish function from superficial output similarity.
- Generalization. Examine transfer beyond the initial setting. Domain transfer, modality translation, runtime usefulness, reward suitability, and human-to-model transfer remain separate claims.
Proposed experiments
Three registered empirical designs. No recorded human or frontier-model runs; specifications and working apparatus do not establish execution.
Human construct and signal pre-test
EXP-0001
Can innovation be assessed independently, and can candidate signals contribute beyond effort, novelty, and other confounds?
Hypothesis, procedure, and open issues
- Hypothesis
- FH-02/FH-03: independent assessment and controlled local signal may be feasible in a defined setting. Domain transfer is conditional on earlier gates.
- Proposed procedure
- Define the task and information boundary; establish independent ratings; examine local signal; then test a frozen instrument across domains within one modality.
- Comparison
- Recalled-original, demanding retrieval, difficult non-innovative reasoning, and superficial novelty conditions; nuisance-only and output-only predictors.
- Principal unresolved issue
- Event/window unit, rubric, rater evidence access, tasks, sample design, and numerical continuation criteria remain unselected. Detector telemetry must not define its own ground truth.
Closed-world recombination
EXP-0002
Do existing models lose exact validity when a required composition has not been demonstrated, despite primitive competence?
Hypothesis, procedure, and open issues
- Hypothesis
- FH-08: a substantive familiar-to-withheld composition loss may remain after primitive understanding and measurement controls.
- Proposed procedure
- Use a finite world with strict executable-plan grading. Assess primitive competence independently, then compare familiar and withheld synthesis alongside retrieval and execution controls.
- Comparison
- Paired familiar/withheld tasks; primitive probes, supplied-handbook retrieval, execution, and copy/random/greedy baselines.
- Principal unresolved issue
- Source matching remains incomplete. Protected stimuli, chosen checkers, model/settings, budget, statistical criteria, and confirmation design need review and freeze.
Workflow attribution
EXP-0003
How much could an apparent composition deficit depend on workflow organization?
Hypothesis, procedure, and open issues
- Hypothesis
- FH-09: a familiar-to-withheld loss may partly depend on the model–workflow combination. Neither that loss nor a workflow benefit is established.
- Proposed procedure
- Cross familiar/withheld synthesis with self-directed/structured workflows in a finite one-model token panel, keeping call opportunities, evidence, history, permissions, and cumulative resources matched.
- Comparison
- Frozen exact checker/oracle and primitive probes; compare withheld gain and familiar change, alongside actual usage, latency, and cost.
- Principal unresolved issue
- Protected execution material, pinned model, resource ceilings, qualification, and decision criteria remain unresolved. A smaller gap caused only by worse familiar performance is not the intended benefit.
Experimental apparatus
Fiction-CWR · Offline reference implementation · v0.1.1
Closed-World Recombination is one instrument within Project Fiction. It supplies synthetic task specifications, a token environment, an exact-validity evaluator, and a computational search oracle for investigating executable recombination.
Models would produce executable solutions from supplied primitives, semantics, constraints, and budgets. The proposed primary comparison contrasts familiar synthesis with a withheld composition. This operational comparison does not measure a human innovation event.
The source worlds retain 2 versus 13 valid plans. This difficulty mismatch prevents treating the familiar/withheld comparison as a frozen matched experiment. Seven additional full checkers, private generators, and execution freezes remain unresolved.
Research library
Working notes, a cognitive-science literature review, and methodological records. Source dates and unresolved claims remain visible.