# Project Fiction — Chronological Hypothesis Change Log

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](../../archive/Hermes_Project_Index.pre-initialization.md).

**Consequences:** Foundational questions remained unresolved. The entire pre-handover documentation state is preserved [in the migration snapshot](../../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 Record narrative-only fine-tuning mechanism FM-R01 as REJECTED / SUPERSEDED according to the handover.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](../../docs/handovers/2026-10-06-migration-handover.txt); [baseline review](literature/cognitive-science-baseline-review-v0.1.md). 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](research-question.md). Preserve the human-to-model gap, label validity problem, event ontology uncertainty, and unselected numerical criteria.

**Decision:** [DEC-0002](../../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](evaluations/fiction-cwr-v0.1/DESIGN.md) defines objective executable outcomes and registers [EXP-0002](experiments/fiction-cwr-exp-0002.md) 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](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](../../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](../../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](experiments/fiction-cwr-workflow-exp-0003.md): 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](evaluations/fiction-cwr-workflow-v0.1/SOURCE.md). Industry commentary motivates the attribution question, not an empirical claim. [Offline verification](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](../../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](../../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](evaluations/fiction-cwr-v0.1/MATCHING-AUDIT-v0.1.1.md). 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](evaluations/fiction-cwr-v0.1/verification/prior-records-v0.1.0.json) retained.
**Decision:** [DEC-0005](../../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](../../docs/handovers/2026-10-14-hermes-discussion-source.md#s33), current user request, and [DEC-0006](../../decisions/decision-log.md#dec-0006--integrate-the-hermes-thesis-and-preserve-its-evolution). [Pilot tiers](../../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](construct-validity-direction.md) 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](../../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](../../decisions/decision-log.md#dec-0007--integrate-fiction-construct-validity-direction); [prior files](../../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](experiments/pt-001/README.md) 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](../../decisions/decision-log.md#dec-0008--execute-pt-001-and-retain-its-attribution-failure).