Scientific record
Project: Project Fiction, Hermes Labs | Proposed: 2026-10-11
Design version: 0.1 | Epistemic label: NEW PROPOSAL
Model runs: 0 | Human runs: 0 | Paid calls: 0 | Spend authorization: none
Protocol and seven required statements (not included in this public selection: ../evaluations/fiction-cwr-workflow-v0.1/PROTOCOL.md) are canonical for this design. Package (not included in this public selection: ../evaluations/fiction-cwr-workflow-v0.1/README.md), proposed plan (not included in this public selection: ../evaluations/fiction-cwr-workflow-v0.1/config/pilot-plan.json), source boundary (not included in this public selection: ../evaluations/fiction-cwr-workflow-v0.1/SOURCE.md) and actual verification (not included in this public selection: ../evaluations/fiction-cwr-workflow-v0.1/VERIFICATION.md) supply reproducibility and limitations.
CURRENT HYPOTHESIS FH-09: An apparent withheld-composition loss can partly depend on workflow organization. This is a provisional alternative explanation for FH-08, not evidence that such a loss or workflow benefit exists. Measured outcomes belong to a specified model–workflow combination.
Proposed inputs/allocation: four related token structural bases, independent F/C nonce target replicas with audited familiar/withheld source evidence; crossed SD/SW; one fresh trial per cell and three calls per session (48 synthesis calls), plus 84 P subitems in four separate fresh control sessions: 52 calls total proposed. Correlated subitems are controls, not independent tasks. This is a finite directional panel, not independent confirmation. Development materials are public/local candidates; fresh protected execution copies and review/freeze remain required.
Controls: Frozen existing Fiction-CWR exact checker/oracle, whole bounded-solution motif checks, source-copy and exposure/difficulty audits, independent primitive probes, same evidence/permissions/settings/history/call opportunities and cumulative resource ceilings. No external retrieval, tools, cross-trial memory, checker feedback or oracle traces. Final strict JSON alone is scored; raw stages and all errors/usage retained.
Metrics: Per-base and finite-panel F/C exact validity by workflow; gaps, gap shrink, withheld gain and familiar change; schema/semantic/provider/resource failures; actual input/output/reasoning usage, latency and cost. Report P and unconditional performance separately; qualification is not yet selected. Distinguish improved validity from equal validity at lower cost. Gap reduction only from worse F is not the intended gain.
UNRESOLVED execution fields: named pinned model/API/version; provider-supported seeds/temperature/reasoning settings and tokenization; caps/total spend/current pricing; protected bundle/assignment/ordering; transport policy; P mastery criterion; continuation/failure numeric values and resource tolerance. Not executed fields remain unavailable. Reproduction requires Python 3.10+ standard library for offline artifacts; the actual runtime/checks are recorded in Verification.
Decision boundary: Research preparation and offline checks are authorized. Paid APIs, training, publishing and submission are not. Resolve/review gates and freeze this complete design before requesting bounded execution authorization. Failed measurement gates block attribution; null, reversed, floor/ceiling and primitive failures are preserved. A persistent deficit does not establish architectural impossibility; successful SW does not establish a human-like mechanism.
Scope separation: EXP-0002's main evaluation remains intact. EXP-0001 human construct/label work, runtime utility, reward suitability and human-to-model transfer retain their independent hypotheses and gates. No model architecture is adopted. Prior records are recoverable in the pre-workflow snapshot (not included in this public selection: ../../../archive/research-snapshots/2026-10-11-before-fiction-workflow/SNAPSHOT.md).