Project overview

Scientific record

Role: Active first project and primary research program of Hermes Labs.

Working definition

Innovation under fixed informational conditions: the ability to recombine, restructure, or transform already-available concepts, relations, memories, or representations in a novel way to produce a new actionable concept, strategy, interpretation, or solution, without materially new external information entering the system.

Shorthand: same information state -> novel recombination/restructuring -> new actionable representation. What remains fixed must be operationalized; internal representations can change. Retrieval is part of the substrate. The contrast is mere reuse or reproduction versus innovative transformation, not a binary opposition between construction and retrieval.

Project Fiction is unfinished and under active stress-testing. Develop, attack, narrow, modify, test, and falsify hypotheses; preserve what changes and what survives. Do not defend a claim because it belongs to the current project.

Current research direction

Observe and operationalize the human capability; test whether innovation events or dimensions can be identified independently; investigate behavioral, linguistic, acoustic, semantic, or other accompanying signals while controlling difficulty, uncertainty, hesitation, retrieval effort, wording novelty, and cognitive load.

The proposed Invention Classifier is a research instrument, not invention itself. Phase One prioritizes domain generalization with modality held constant. Modality translation comes later. A directional pre-test asks whether further resources are warranted; a small success does not confirm the thesis.

Runtime assistance and reward-signal use are separate hypotheses. Detector performance establishes neither intervention benefit nor reward suitability. Ultimately, investigate reproduction or encouragement of the capability in AI models, first assessing relevant behavior in existing systems.

Canonical working documents

Also read methodology, research status, and the provisional process model (not included in this public selection: theory/provisional-process-model.md). Supporting artifacts retain the existing project folders; use the lab-wide experiment registry for actual registered designs and runs.

Evidence and history

The migration handover (not included in this public selection: ../../docs/handovers/2026-10-06-migration-handover.txt) supplies current prior-work context; the older conversations, original protocols, and raw data remain unavailable. The pre-handover snapshot (not included in this public selection: ../../archive/migration-snapshots/2026-10-06-before-handover/projects/fiction/README.md) preserves the initialization state. No human study dataset, collection-ready human protocol, classifier implementation, frontier-model run, or empirical capability finding is recorded. The separate Fiction-CWR branch now includes synthetic showcase specifications, a token reference fixture/interpreter/oracle and offline engineering verification.

The baseline separates established related science from plausible extensions and a candidate distinctive event/label/control/transfer claim. It does not establish scientific novelty or validate the program. Project Imagination and Project Daydream retain distinct research questions even where mechanisms overlap.

Existing-model capability evaluation branch

CURRENT HYPOTHESIS / NEW PROPOSAL: Frontier language models may show a measurable deficit in valid closed-world recombination despite primitive understanding. Fiction-CWR v0.1 operationalizes the narrow branch FH-08 with P primitive probes, R supplied-handbook retrieval, E execution and primary F familiar versus C withheld synthesis. Novelty is relative to supplied evidence, never asserted absence from pretraining. EXP-0002 is proposed with zero model runs.

The package supplies eight concrete task specifications and one implemented reference family; private matched stimuli and structural holdouts remain a run gate. It complements EXP-0001, preserving human construct/label/signal, runtime/reward and transfer hypotheses. Integration (not included in this public selection: evaluations/fiction-cwr-v0.1/INTEGRATION.md) and DEC-0003 (not included in this public selection: ../../decisions/decision-log.md#dec-0003--add-fiction-cwr-as-a-complementary-ai-capability-evaluation) record the scope.

CURRENT HYPOTHESIS / NEW PROPOSAL FH-09: Workflow organization may partly explain an apparent composition deficit; performance belongs to a specified model–workflow combination. EXP-0003 prepares a separate small one-model F/C × self-directed/structured comparison with matched iterations, history, evidence, permissions and cumulative resource ceilings. Offline package (not included in this public selection: evaluations/fiction-cwr-workflow-v0.1/README.md) supplies candidate evidence/certificates and stored-record checks, not model findings. Protected execution stimuli, provider/compute/criteria/spend freeze and authorization remain unresolved. The main evaluation and human/runtime/reward/transfer questions retain their independent scopes.

October 14 — Wider thesis and first-pilot role

The lab thesis (not included in this public selection: ../../lab/mission.md) connects Fiction to possible minds through characterized capabilities. Fiction Experiment 001 / Micro1 is the source discussion's first serious pilot and simultaneous Hermes Method test: a forcing function, not all of Fiction and not registry EXP-0001.

Scientific criteria and tiers (not included in this public selection: ../../lab/open-questions-and-next-experiments.md) complement existing gates. Operational retrieval/deduction/recombination comparisons retain retrieval as substrate and separate human-label/runtime/reward/transfer claims. No protocol, hypothesis, gate, or run state is silently replaced by source shorthand. The retrieved conversation supplies thesis context, not raw data or missing protocols.

Public-facing artifact site prototype

A dependency-free local homepage and visual direction brief (not included in this public selection: public-site/ART-DIRECTION.md) begin the public-facing Project Fiction work. The prototype links to the canonical evaluator record and surfaces current evidence limits. It is presentation work only; no research hypothesis, result, gate, or run status changes.

The production host and public artifact URL remain unresolved. Local links point to the evaluator sources in this repository.

2026-10-23 — Construct-validity source integration

Construct validity and problem transformation integrates the readable *Fiction Research Program* conversation and its explicit user corrections. Separate innovation outcomes from mechanisms; keep fixed-goal solutions and valuable problem reframing in scope. Situated Functional Innovation is not adopted as the complete definition. The canonical question and CWR branch remain intact. Source excerpts (not included in this public selection: ../../docs/handovers/2026-10-23-fiction-research-program-source.md) disclose missing report bodies and attachment content; this is partial source recovery, not an empirical result.

PT-001 — Executed measurement rehearsal

The PT-001 packet and results implement the user's Cycle 3 review locally. Fixed-goal innovation versus valuable problem transformation is provisionally accepted as an experimental distinction, not separate mechanisms. Four resource-distribution worlds, independent exhaustive reference outcomes and constructed strategy/adversarial tests were executed. The apparatus supports bounded feasibility/value measurement; reframing-specific attribution is not identified because objective optimization/search reproduce the gains. No human or model inference was run. The canonical question and CWR implementation remain intact.