# Experiment registry Updated: 2026-10-23 **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](../projects/fiction/experiments/pilot-pre-test-redesign.md) | 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](../projects/fiction/experiments/fiction-cwr-exp-0002.md) | Zero model runs; [offline engineering verification](../projects/fiction/evaluations/fiction-cwr-v0.1/VERIFICATION.md) 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](../projects/fiction/experiments/fiction-cwr-workflow-exp-0003.md) | Zero model runs; [offline verification](../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](../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](../projects/fiction/experiments/pt-001/README.md); 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//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](../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](../projects/fiction/experiments/fiction-cwr-exp-0002.md) 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](../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](../projects/fiction/construct-validity-direction.md) 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.