Scientific record
Append changes without rewriting earlier positions. Dates below are repository record dates; dates of unavailable older work are not invented. Empirical support and a position's provenance must be recorded separately.
2026-10-06 — Initialization state
Issue: Only the initial index and initialization brief were available.
Previous/new position: Preserved the provisional innovation-under-fixed-information target. No testable hypothesis or protocol was recorded in the scaffold. This was a description of repository contents, not a rejection of hypotheses outside it.
Evidence: Initial brief and original index (not included in this public selection: ../../archive/Hermes_Project_Index.pre-initialization.md).
Consequences: Foundational questions remained unresolved. The entire pre-handover documentation state is preserved in the migration snapshot (not included in this public selection: ../../archive/migration-snapshots/2026-10-06-before-handover/README.md).
2026-10-06 — Import migration handover and baseline v0.1
Issue: The new handover supplies prior positions absent from the scaffold.
Previous position: Broad recombination target; no current recorded hypothesis map. Original narrative-training mechanism and classifier design history unavailable.
New position: Record transformation rather than mere reuse as the contrast, with retrieval as substrate; retain a provisional process model; decompose construct, labels, local signal, same-modality domain generalization, runtime, reward, and AI reproduction into FH-01 to FH-07. Record narrative-only fine-tuning mechanism FM-R01 as REJECTED / SUPERSEDED according to the handover.
Reasoning: Preserve the supplied research context while separating claims that do not entail one another. The literature baseline supplies close precedents and constraints, not confirmation of the event or transfer hypothesis.
Evidence: Handover (not included in this public selection: ../../docs/handovers/2026-10-06-migration-handover.txt); baseline review. Older formulations, dates, protocols, and raw data remain unavailable. The narrative rejection's rationale is anticipated style transfer, not a verified experimental result here.
Consequences: NEW PROPOSAL identifiers, audit, pre-test sketches, and gates now make falsification paths explicit. No claim has passed an experimental gate. The baseline's MemiC comparison narrows the candidate distinctiveness to the event/label/control/transfer conjunction, with scientific novelty still unestablished.
Open questions: See FQ-01 to FQ-13. Preserve the human-to-model gap, label validity problem, event ontology uncertainty, and unselected numerical criteria.
Decision: DEC-0002 (not included in this public selection: ../../decisions/decision-log.md#dec-0002--incorporate-the-migration-handover-and-constrain-fiction-with-literature).
2026-10-11 — Add complementary existing-model recombination branch
Issue: The long-term AI reproduction aim needs a narrow existing-model capability test.
Previous position: FH-01–FH-07; human construct/label/signal outline EXP-0001; no implemented evaluation or observed model deficit.
New position — CURRENT HYPOTHESIS / NEW PROPOSAL: Add FH-08: pinned models may lose valid-synthesis performance for withheld versus familiar supplied composition motifs despite independently measured primitive competence. Candidate δ=0.20/τ=0.90 are provisional design decisions, not facts. Fiction-CWR defines objective executable outcomes and registers EXP-0002 as proposed with zero model runs.
Reasoning/evidence: Supplied current task; finite semantics and objective fidelity can test a limited behavior without innovation-event labels. Prior art (not included in this public selection: evaluations/fiction-cwr-v0.1/PRIOR-ART.md) motivates methods, not current-model failure.
Consequences: Eight specs and one reference checker; independent oracle and engineering tests; matched stimuli/necessary motif gates, prospective confirmation, all-model/qualified reporting. Preserve FH-01–FH-07, FM-R01 history, EXP-0001 and separate human/runtime/reward/transfer branches. Prior canonical snapshot (not included in this public selection: ../../archive/research-snapshots/2026-10-11-before-fiction-cwr/SNAPSHOT.md) preserves earlier formulations.
Open questions: Demonstration-relative novelty, residual evidence/difficulty matching, valid familiar alternatives, provider/budget freeze, sample power, practical/human transfer and persistent deficit.
Decision: DEC-0003 (not included in this public selection: ../../decisions/decision-log.md#dec-0003--add-fiction-cwr-as-a-complementary-ai-capability-evaluation).
2026-10-11 — Prepare a workflow attribution alternative
Issue: A future apparent familiar/withheld loss could belong partly to workflow organization rather than to the model generally.
Previous position: FH-08 and EXP-0002 already identify prompt, budget, load and representation alternatives; no model loss is observed. Main private matched stimuli remain unresolved.
New position — CURRENT HYPOTHESIS / NEW PROPOSAL: Add FH-09 and EXP-0003: one pinned model, F/C × self-directed/structured, equal stage opportunities and same-trial history, common evidence/permissions, cumulative ceilings and actual usage. Candidate token evidence pairs, bounded certificates and offline stored-record analysis are separate from the main arm.
Reasoning/evidence: Current supplied preparation brief; inspected Hermes canon and frozen CWR package; article source boundary and Hermes inference (not included in this public selection: evaluations/fiction-cwr-workflow-v0.1/SOURCE.md). Industry commentary motivates the attribution question, not an empirical claim. Offline verification (not included in this public selection: evaluations/fiction-cwr-workflow-v0.1/VERIFICATION.md) concerns apparatus only.
Consequences: Performance is scoped to model–workflow conditions. Require C improvement alongside gap shrink, retain F change and actual cost; F-only deterioration is not support. Equal validity at lower cost is efficiency. Related public development cases do not become private independent tasks through nonce renaming. Retain all failure/null results and source-matching limits.
History/open questions: Sixteen prior records preserved (not included in this public selection: ../../archive/research-snapshots/2026-10-11-before-fiction-workflow/SNAPSHOT.md). Human/runtime/reward/transfer hypotheses and EXP-0002 main arm remain intact. No numeric criterion, model, hidden-compute equivalence, spend or empirical execution is adopted. Model/human runs remain zero.
Decision: DEC-0004 (not included in this public selection: ../../decisions/decision-log.md#dec-0004--prepare-a-separate-workflow-attribution-diagnostic).
2026-10-12 — Refine FH-08's token evidence boundary
Issue/previous position: Current main records still described token pair/necessity apparatus as absent, though EXP-0003 had public candidate certificates; broad “motif withheld” could imply withholding primitive/rule information.
New position — NEW PROPOSAL / working definition: Reuse one public pair in EXP-0002 with strict bounded certification. C withholds realized multistep production dependencies from its compound trace; rules/one-step examples expose the dependency for permitted deduction.
Reasoning/evidence: Current saved sources inspected; stronger partial-state/failure checks and 12 new tests pass. The old goal-only pruning comparison was insufficient assurance. Target spaces match; source F/C valid-plan counts 2/13 are an explicit residual.
Consequences/open questions: Preserve all valid answers, frozen checker/oracle, EXP-0001/EXP-0003 and separate gates. No empirical hypothesis support or numeric criterion change. Source matching, private/execution freeze and transfer remain unresolved. Exact prior records (not included in this public selection: evaluations/fiction-cwr-v0.1/verification/prior-records-v0.1.0.json) retained.
Decision: DEC-0005 (not included in this public selection: ../../decisions/decision-log.md#dec-0005--reuse-token-evidence-and-strengthen-bounded-certification).
2026-10-14 — Lab thesis context and pilot naming
Scope: Conceptual integration, not hypothesis outcome. Fiction is the first serious pilot of a possible-minds program; Micro1 supplies a forcing function for Fiction Experiment 001 and a method test.
Preserved: Current definition, retrieval-as-substrate, FH-01–FH-09, separate EXP-0001/2/3, detailed gates, zero empirical-run state. Source retrieval/deduction/recombination shorthand is operational comparison, not a new ontology.
Source/reasoning: Discussion S33–S49 (not included in this public selection: ../../docs/handovers/2026-10-14-hermes-discussion-source.md#s33), current user request, and DEC-0006 (not included in this public selection: ../../decisions/decision-log.md#dec-0006--integrate-the-hermes-thesis-and-preserve-its-evolution). Pilot tiers (not included in this public selection: ../../lab/open-questions-and-next-experiments.md) preserve redefining/abandoning a failed formulation. No empirical hypothesis acceptance or rejection occurred.
2026-10-23 — Clarify outcome measurement without adopting a new construct
Previous position: FH-01–FH-09, process/classifier hypotheses and narrow CWR measurements; later source review absent.
New position: Construct-validity direction separates the outcome from its producing mechanism and keeps problem transformation in scope alongside fixed-goal solutions. Search/analogy can constitute the target's mechanism. Definition B is not adopted as Fiction's complete operational definition. No hypothesis is empirically accepted/rejected and the canonical research question is unchanged.
Evidence/reasoning: Explicit user corrections in retrieved source (not included in this public selection: ../../docs/handovers/2026-10-23-fiction-research-program-source.md); assistant reports/proposals do not establish findings. Missing Cycle 2/3 bodies and CV-001 materials remain visible gaps.
Consequences/open questions: Preserve all registered branches, gates and zero-run state; recover original materials and test independent outcome validity, usefulness and goal/reframing controls. DEC-0007 (not included in this public selection: ../../decisions/decision-log.md#dec-0007--integrate-fiction-construct-validity-direction); prior files (not included in this public selection: ../../archive/research-snapshots/2026-10-23-before-fiction-source/manifest.json).
2026-10-23 — PT-001 exposes outcome-only attribution failure
Previous/new position: The user provisionally accepts fixed-goal versus problem-transformation as an experimental distinction and requests local execution. PT-001 preserves FH-01–FH-09 and the canonical question. It rejects, for this operationalization, treating a gain over proxy optimization or weak search as evidence of formulation-specific innovation.
Evidence: Actual constructed-strategy outputs, complete bounded enumeration, independent replay and adversarial tests. Direct value optimization reaches all primary ceilings; search-only R-order exactly reproduces R. Value criteria change conflict labels, including in hypothesized no-conflict worlds. No human/model mechanism observed.
Consequences/open questions: Retain outcome/mechanism separation and strengthen objective/search controls. The provisional experimental distinction survives as a question, not a demonstrated mechanism or measured cognitive effect. Recover original source materials separately; next cognitive design must justify formulation coding, resources, usefulness and criteria. DEC-0008 (not included in this public selection: ../../decisions/decision-log.md#dec-0008--execute-pt-001-and-retain-its-attribution-failure).