Scientific record
Registry state: Three proposed empirical designs; zero recorded human or frontier-model runs. Offline engineering verification is recorded separately. The registry was empty at initialization. Registration records proposed designs from the migration review, Fiction-CWR architecture and workflow diagnostic preparation; it does not imply approved empirical execution, participant recruitment, data, or results. Work outside the supplied records remains unknown.
Registered proposals and runs
| Identifier | Project | Question · hypothesis | Status | Proposed | Design record | Runs · results |
|---|---|---|---|---|---|---|
| EXP-0001 | Project Fiction | Construct/label feasibility and controlled signal testing; FH-02/FH-03, with FH-04 conditional on earlier gates | proposed; incomplete design outline v0.1 | 2026-10-06 | Pilot / Pre-Test Redesign | None recorded |
| EXP-0002 | Project Fiction | Existing-model closed-world recombination; FH-01/FH-07/FH-08, primary familiar versus withheld synthesis | proposed; evaluation architecture v0.1; matching/provider/statistical freeze pending | 2026-10-11 | Fiction-CWR design record | Zero model runs; offline engineering verification separately |
| EXP-0003 | Project Fiction | Workflow attribution alternative to apparent loss; FH-09/FH-08, one-model F/C × SD/SW | proposed; offline candidate stimuli/apparatus; protected execution panel, criteria/model/resource/spend freeze pending | 2026-10-11 | Workflow diagnostic | Zero model runs; offline verification (not included in this public selection: ../projects/fiction/evaluations/fiction-cwr-workflow-v0.1/VERIFICATION.md) separately |
EXP-0001 must resolve its event unit, rubric, evidence access, tasks, modality, domains, sample design, metrics, numeric continuation criteria, and procedure before collection. Candidates A-C (not included in this public selection: ../projects/fiction/experiments/next-experiments.md) are conditional components of this outline, not three separately approved experiments. Later studies require their own registrations and complete records.
Completed offline measurement diagnostics
| Identifier | Project | Question | Status · scope | Record and results |
|---|---|---|---|---|
| PT-001 v0.1 | Project Fiction | Can outcome gains distinguish formulation change from value optimization/search? | Completed local constructed-strategy rehearsal, 2026-10-23; not a human/model experiment | Protocol and actual results; run-001 and deterministic reproduction; bounded apparatus checks pass, reframing-specific attribution fails |
PT-001 is a new local realization of the supplied review, not recovery of the missing original Cycle 3 protocol. Four worlds, 4,372 bounded sequences and 480 constructed strategy records per rehearsal were evaluated. Direct value reaches all ceilings; an identical-order search-only strategy reproduces R exactly. No human/model gate passed. The three empirical proposals above remain unexecuted; PT-001 is separate from Fiction Experiment 001, EXP-0001 and CV-001. Its conditional cognitive-system follow-up is not an approved run or complete frozen protocol.
Where records belong
Project-specific designs and runs belong in projects/<project>/experiments/. Cross-program experiments belong in experiments/shared/ and must state their relevance to each participating project. Link evaluations, model configurations, datasets, and results from the complete experiment record rather than maintaining conflicting copies.
Future registry entries
Give each experiment a stable identifier and register its project or shared scope, question or linked hypothesis, actual status, design-record path, and any run/result paths. Record relevant dates and changes in status. Register actual documented proposals, clearly marking incomplete designs; do not turn placeholder identifiers or speculative mentions into experiments.
Use proposed, underway, completed, blocked, or cancelled according to the actual state. A proposed design is not an executed run. Completion does not imply success; preserve negative and inconclusive results. If a design changes across runs, identify the version used for each run.
Complete design and run records
Record the applicable details needed to interpret and repeat the work:
- Research question or hypothesis, design rationale, conditions, controls, and intended evaluation criteria.
- Inputs, dataset identity and version, selection and preparation steps, and relevant information constraints.
- Model identity and version, configuration, environment or dependencies, prompts, and random seeds where relevant.
- Procedure or linked scripts, actual run conditions, and deviations from the design.
- Metrics and how they were computed, outputs and artifact locations, and the run's actual state.
- Observations, Hermes' interpretation, limitations, uncertainty, and unresolved questions.
Separate proposed methods from executed conditions and observed results. Mark unavailable or inapplicable details explicitly. Prefer scripts and configuration over undocumented manual steps; explain reproduction limits when they remain. Substantive design or evaluation changes must update canonical project documents and the research log (not included in this public selection: ../logs/research-log.md).
EXP-0002 execution boundary
Eight public showcase specs and one token reference implementation do not constitute model runs or a completed reusable generator release. F/C matching and necessary-motif certification, remaining selected checkers, P counts, provider settings/prices, spending ceiling and prospective confirmation design are unresolved. No paid calls are authorized by registration. EXP-0001's human labels/signals and all independent runtime/reward/transfer branches remain unchanged.
EXP-0003 execution boundary
The separate token workflow diagnostic prepares a finite panel of development candidates and offline exact audits; it does not complete EXP-0002's private benchmark gates. Proposed 48 synthesis calls plus bundle-determined independent P probes authorize no spend. Model identity, protected execution material, provider tokenization/compute access, caps/criteria/qualification and budget require review/freeze. Report exact-validity gain separately from equal validity at lower cost. A smaller gap only from worse familiar performance is not success. No external retrieval, cross-trial memory, validator feedback, witnesses, training or publication is authorized.
2026-10-12 — EXP-0002 token engineering scope
Version 0.1.1 now links a complete original token audit and one imported public F/C candidate with stricter full-state rejection checks and 12 passing new tests. Source matching/private main-arm freeze remain pending. No new empirical experiment or independent base is registered; all three existing empirical proposals retain zero runs. EXP-0003 source/artifact files are preserved; reuse imports no workflow allocation.
2026-10-14 — Pilot nomenclature clarification
Fiction Experiment 001 / Micro1 is the source discussion's program-level first-pilot label and Hermes-method test. It does not rename EXP-0001 (human construct/classifier pretest), EXP-0002 (CWR), or EXP-0003 (workflow attribution). No new empirical design is registered; all proposed states, gates, and zero human/frontier-model run counts remain. Pilot criteria and tiers (not included in this public selection: ../lab/open-questions-and-next-experiments.md) complement the designs without replacing their unresolved numerical criteria.
2026-10-23 — Source-reported CV-001 and construct-validity work
Source reconciliation records CV-001 as a proposed sixteen-case measurement diagnostic referenced by the user in the retrieved conversation. Its original complete design/stimuli and Cycle 2/3 reports were not recovered. This is an incomplete source mention awaiting recovery, not a new complete registry design, an alias for EXP-0001, or a run. The three registered empirical designs and zero recorded human/frontier-model runs remain unchanged. Source requests for additional pilots/reformulation comparisons remain design work, with no inferred execution or spend authorization.