# Narrative Pro methodology: curated evidence summaries Reviewed 2026-09-07. Article version: methodology-2026-09-07. This publication summarizes internal records. It is not raw run data, an independently reproduced benchmark, or a peer-reviewed study. The summaries were deliberately authored for public reading; private source narratives and test artifacts are not included. Missing fields remain **Not recorded**. A record's date is not a production release date. ## length-control-2026-06 - Date: 2026-06-27. Category: writing. Status: reported-observation. - Question: Can measured revision approach a character limit without inventing facts? - Hypothesis: Not recorded. - Method and input: retrospective summary of a generate, measure and revise spike in a historical design note. Original input privacy classification: Not recorded; no input narratives are reproduced. - Sample: six reported sample-by-target runs, with targets of 250 and 350 characters. Unique input count and pairing: Not recorded. - Model: gemini-3-flash-preview. Prompt version: Not recorded. - Counting convention in the original runs: Not recorded. A later specification of normalized Unicode codepoints is not proof of the spike's measurement method. - Observation: the note describes reaching a near-limit band through revision, but also describes naive prompts inventing numeric details. - Limitations: original run-level inputs, outputs, prompts and counting records were not located in this bounded review. This does not prove they never existed. No benchmark rate, average call count, quality rating or customer outcome is published. Length fit alone is not evidence fidelity. - Decision: measure outputs; ask for missing evidence; reject unsupported additions. ## field-fit-2026-09 - Date: 2026-09-06. Category: writing. Status: reported-observation. - Question: Does a live draft meet the near-limit band and remain reviewable? - Hypothesis: Not recorded. - Method and input: local workflow QA with fictional evidence and disposable accounts. No account identifiers or output narratives are published. - Sample: two individual generation observations described in a QA record, not a complete run ledger. - Model: Gemini; exact model version: Not recorded. Prompt: first observation pre-v2 (exact version Not recorded); later observation labeled v2. - Counting convention: the record reports Unicode codepoints with a target of 350 and a zero-to-five-under band. Original normalization and count artifacts were not independently reconstructed for publication. - Observation: both described drafts were labeled short of the requested band. The record separately reports explicit Apply and Undo working. - Limitations: original run-level artifacts were not reconstructed; exact output counts and quotations are withheld. The observations are not a paired A/B test and cannot establish a model success rate or general factual accuracy. - Decision: show short, within-band and over-limit states separately from writing feedback and factual review. ## synthetic-walkthroughs-2026-09 - Date: 2026-09-05 (remediation ledger, referring to evaluation on 2026-09-04). Category: usability. Status: reported-observation. - Question: Where can a career-writing journey lose evidence, edits or intent? - Hypothesis: Not recorded. - Method and input: synthetic consumer-persona walkthroughs summarized by a remediation ledger. Only generalized usability findings are included here. - Sample: 15 synthetic walkthroughs, not 15 recruited users. - Model and prompt versions: Not recorded. Character-count convention: not applicable to the walkthrough sample unit. - Observation: the ledger identifies lost unsaved work, signup-handoff fragility, unclear field-fit feedback, target-role confusion and evidence-preservation problems. Numeric preservation alone did not resolve unsupported semantic claims. - Limitations: these were not interviews or recruited participants. No real-user prevalence, satisfaction or customer outcome can be inferred. Local software closure checks are not a follow-up usability study or production delivery. - Decision: retain source evidence and edits, ask neutral questions, and separate desired roles from actual employment history. ## resume-reliability-2026-09 - Date: 2026-09-07. Category: engineering. Status: reported-observation. - Question: Can resume editing preserve intent across delays, save failures and navigation? - Hypothesis: Not recorded. - Method and input: local execution ledger reporting unit regressions, controlled browser fixtures, type checks, static audits and a production build. No raw fixtures or personal records are included. No checks were rerun for this summary. - Sample: regression scenarios for stale responses, recovery, target-role persistence, cancellation and optional Air Force entry. Assertions are not participants or model runs. - Model and prompt versions: not applicable to controlled provider fixtures. Character-count convention: not applicable to these software checks. - Observation: the ledger reports passing checks for preserving newer edits, recovering unsaved work and retaining target-role changes. Browser checks also found authentication-handoff and route-initialization defects missed by focused unit mocks; the recorded revision corrected them. - Limitations: this is recorded local verification, not an independent article rerun. Software checks do not establish writing quality, live-model accuracy, customer outcomes or production deployment. Recovery is a time-limited latest copy per tab, not permanent backup. Independent release review remains required. - Decision: keep proposals reviewable, reject stale results and test whole journeys. ## scoring-history Reviewed 2026-09-07. Method description, not an empirical study. The historical C6 framework classified data maturity on six levels: basic metric mention, comparison, multiple-point analysis, multi-faceted statistical analysis, prediction, and causal analysis with attribution. Its military formula was: `score = 1.25 * (impactOrgLevel - currentOrgLevel) + c6Level` The organizational-level difference represented rank-relative reach. The 1.25 multiplier was a chosen scoring weight, not measured evidence of 25 percent more impact. The historical corporate formula was `score = c6Level + impactLevel`. These definitions are historical product rules, not current military regulations or evidence that wording establishes causation. The inspected det-v1 implementation uses five dimensions, each from 0 to 2: specificity, quantification, impact scope, structure and language. Their sum is from 0 to 10. The deterministic editor path uses text heuristics, including keyword scope and clause structure. Server grading can instead use a language model for impact scope and coaching, with a deterministic fallback. Neither path verifies the underlying accomplishment. The old and new scores are not a validated common scale or probabilities of employment, promotion or awards. ## fictional-illustration Reviewed 2026-09-07. Illustration only; not part of the experiment register. The original, clarified and refined example was authored as fiction. Clarification supplies new fictional evidence; refinement changes wording without adding facts. The observed time change in that example is not proof of causation. No experimental grade or measured quality improvement is assigned to it. ## next-protocols Recorded 2026-09-07. Status: planned. Neither protocol has been run. 1. Paired evidence and field-fit evaluation: freeze synthetic inputs, counting rules, prompts and models; retain all attempts and failure types; distinguish unique inputs from repeats. Sample size and versions: Not recorded. 2. Human reading and reuse: define consent, recruitment and rating criteria; blind and randomize comparisons where feasible; retain disagreements. Sample size and recruitment records: Not recorded. No hiring-outcome claim is proposed. ## change-history 2026-09-07: first curated article record. Separates reported writing observations, synthetic usability evaluation, engineering verification and fiction. Withholds unsupported rates and exact historical output counts. Documents missing evidence and planned protocols without presenting them as completed results.