Preregistration — Penwright Paper 7: Genre-Specific Effects of AI Writing Systems

Study identifier: penwright-paper-07-genre-effects Repository: people-analyst/peopleanalyst-site (preregistration); people-analyst/vela (Penwright production system) Registration type: OSF-style internal preregistration (upload-ready when external-operator pilot enrolls first cohort across genres) Version: 1.0 — 2026-05-05 (pre-data; locked at first cross-genre pilot enrollment) Frozen at: commit SHA recorded at OSF filing Companion preregistration: penwright-paper-05-dependency.md (filed alongside; shares analytical infrastructure) Companion papers: Paper 4 (Measurement Framework — operationalizes the indices); Paper 6 (Learning Loops — operationalizes loop discipline); Paper 3 (Authorship Packet Model — input shape); Paper 2 (Corpus Control — epistemic substrate)

---

1. Background

The genre-fork is the program's most architecturally consequential commitment and its most empirically vulnerable. The Penwright vision documents (vela/docs/VISION-PENWRIGHT-AUTHORSHIP.md §6.3, §7) require that memoir, nonfiction, and fiction never collapse into a single skill model — copy, schema enums, prompts, and metrics each fork by genre rather than sharing a runtime parameter. The four-failure-mode veto (per Paper 4 §3) treats ignoring genre differences as a non-negotiable disqualifier. The kernel architecture (F-19) enforces the fork at the registry layer.

These commitments cost real engineering. They are justified if AI's effects on writing capability differ meaningfully across memoir / nonfiction / fiction. They are not justified — and would be over-engineering — if AI's effects collapse across genres. Paper 7 preregisters the test that adjudicates the question.

The headline hypothesis is genre-dependence of AI writing effects. The headline is falsifiable: if Paper 7's data shows that AI effects do not differ across genres on the indices the framework treats as load-bearing, the genre-fork is not justified at that layer (it may still be justified at adjacent layers — measurement, intervention design, voice preservation — but the program's load-bearing claim about genre-as-structural would not survive). The yes-world / no-world specification below is designed to be informative either way.

The study is gated on the external-operator pilot accumulating writers across all three genres — at minimum two writers per genre with 3+ months of data and ≥3 valid Constraint-Mode samples per writer. The preregistration is filed pre-data; data accumulation begins at first cross-genre enrollment and continues until the minimum cohort condition is met.

---

2. Research question (program link)

Paper 7 in the twelve-paper Penwright Research Program. Question: Do AI-augmented-writing effects on capability development differ across memoir, nonfiction, and fiction at a magnitude that justifies the program's structural genre-fork commitment?

The question is operationalized as a between-genre contrast on the genre-specific subcomponents of the Penwright Measurement Framework (per Paper 4 §2.2 — Expressive Skill is genre-forked at the dimension level; integration, originality, and metacognitive indicators have genre-specific components). The contrast pairs writers within and across genre cohorts, controlling for AI-utilization, loop-completion, and writer-level baselines.

---

3. Hypotheses

H1 (primary). The strength of the AI-utilization-vs-capability-transfer relationship (per Paper 5) differs across the three genres at a magnitude meeting the pre-specified effect-size threshold (§5.3).

• Yes-world consequence: the genre-fork is structurally justified; the program's architectural commitment to never collapsing genre is empirically supported on the load-bearing dimension.
• No-world consequence: AI-utilization-vs-capability-transfer is statistically indistinguishable across genres, controlling for writer-level baselines. The genre-fork is not justified at this layer; the architecture may still be justified at others (per §6 below).

H2a (memoir-specific). AI utilization is associated with measurable decline in emotional specificity (the memoir-specific component of Expressive Skill, scored against the writer's pre-AI baseline) at a per-genre detection threshold larger than the equivalent decline in nonfiction or fiction.

• Yes-world consequence: the program's specific prediction that memoir is at risk for emotional flattening under AI is supported.
• No-world consequence: memoir's emotional specificity is no more vulnerable to AI utilization than other genres' equivalent dimensions; the prediction is wrong, the genre-specific intervention design needs revision.

H2b (nonfiction-specific). AI utilization is associated with measurable decline in argumentative depth (the nonfiction-specific component of Integrative Skill — argument-evidence linkage, counterposition engagement, conceptual precision) at a per-genre detection threshold larger than the equivalent decline in memoir or fiction.

• Yes-world consequence: the program's specific prediction that nonfiction is at risk for shallow argument under AI is supported.
• No-world consequence: nonfiction argumentative depth is no more vulnerable to AI utilization than other genres' equivalent dimensions; the prediction is wrong.

H2c (fiction-specific). AI utilization is associated with measurable decline in narrative distinctiveness (the fiction-specific component of Originality — sensory specificity, scene construction, character voice) at a per-genre detection threshold larger than the equivalent decline in memoir or nonfiction.

• Yes-world consequence: the program's specific prediction that fiction is at risk for generic narrative under AI is supported.
• No-world consequence: fiction narrative distinctiveness is no more vulnerable to AI utilization than other genres' equivalent dimensions; the prediction is wrong.

H3 (intervention-genre interaction). Intervention efficacy (per Paper 6 §3) varies across genres — the four intervention types (teach moments · constraint challenges · counterposition drills · reconstruction exercises) produce different effect magnitudes per genre on the genre-specific dimensions.

• Yes-world consequence: the kernel's genre-fork on intervention design is justified; cross-genre intervention reuse is not safe by default.
• No-world consequence: intervention efficacy does not vary by genre at the magnitude tested; the kernel can collapse intervention surfaces across genres without measurement loss.

Exploratory (descriptive only, no preregistered tests). Cross-genre writer-level effects (writers who work across multiple genres — does dependency in one transfer to capability in another?); within-genre subgenre effects (literary memoir vs. journalistic memoir; argument-driven vs. narrative-driven nonfiction; literary fiction vs. genre fiction).

---

4. Design summary

Element	Specification
Setting	Continuous production-data observation; Penwright telemetry across the external-operator pilot cohort
Unit of observation	Writer-genre-week (writers may work across multiple genres; observations stratified per genre)
Unit of analysis (primary)	Writer-genre cell (writer × genre); random effects for writer; fixed effects for genre
Independent variables	AI-utilization composite (per Paper 5 §5.1); intervention-type exposure
Outcome (primary)	Genre-specific Expressive/Integrative/Originality components, scored against the writer's per-genre pre-AI baseline
Stratification	Genre (three-level fixed effect); analyses run separately per genre AND in pooled form with explicit genre-utilization interaction terms
Held constant	Penwright instrumentation; kernel version pinned per analytical window
Blinding	Analyst blinded to writer identity; expert raters blinded to writer identity, AI-utilization condition, and study hypothesis

---

5. Operationalization

5.1 Per-genre outcome scoring

Memoir (H2a): Emotional specificity score per Constraint-Mode sample. Scored by ≥2 expert raters using a memoir-specific rubric covering: experiential grounding (concrete sensory anchoring of remembered moments); affective range (vocabulary specificity for emotional states); the felt-distinction between describing an emotion and rendering an emotion; resistance to interpretive over-articulation. ICC ≥ 0.7 required for inclusion.

Nonfiction (H2b): Argumentative depth score per Constraint-Mode sample. Scored by ≥2 expert raters using a nonfiction-specific rubric covering: argument-evidence linkage (each load-bearing claim traces to specific evidence); counterposition engagement (alternative readings articulated and addressed); conceptual precision (terms defined; distinctions held consistent); resistance to gestural assertion. ICC ≥ 0.7.

Fiction (H2c): Narrative distinctiveness score per Constraint-Mode sample. Scored by ≥2 expert raters using a fiction-specific rubric covering: sensory specificity (concrete particularity over generic description); scene construction (dramatic logic, beat structure, pacing); character voice distinctiveness (each speaker individuated); resistance to AI-default narrative shape. ICC ≥ 0.7.

The three rubrics share structure (multi-component scoring, expert raters, ICC threshold) but never share content — they are independently calibrated against genre-specific exemplars from the canon. Cross-genre rubric reuse is not permitted.

5.2 AI-utilization measure

Same composite as Paper 5 (§5.1) — retrieval engagement rate (0.4) + suggestion-acceptance ratio (0.4) + time-with-active-AI proportion (0.2). Z-scored within-genre for cross-writer comparability per genre.

5.3 Effect-size thresholds (pre-specified)

H1 (primary): Genre-fork is supported only if at least one pairwise genre contrast on the strength of the AI-utilization-transfer relationship reaches Cohen's d ≥ 0.6. Smaller cross-genre differences are reported descriptively; the genre-fork claim is not made on smaller effects.

H2a-c (genre-specific): Each genre-specific decline prediction is supported only if (a) within-genre decline magnitude reaches the writer-level effect-size threshold of d ≥ 0.5 against pre-AI baseline AND (b) the within-genre decline exceeds the equivalent within-other-genres declines by d ≥ 0.4.

H3 (intervention-genre): Intervention-genre interaction is supported only if at least one (intervention type × genre) cell shows efficacy that differs from the same intervention's pooled cross-genre efficacy by d ≥ 0.5.

These thresholds are pre-committed. The analysis cannot find smaller effects "supportive" of the hypotheses post-hoc.

---

6. Implications of falsification

The "no-world" outcome on H1 — AI effects are statistically indistinguishable across genres — has explicit consequences for the program's architecture. The implications are pre-stated to prevent post-hoc rationalization.

If H1 is not supported (no cross-genre difference at d ≥ 0.6 detected on the load-bearing dimensions): the genre-fork is not empirically justified at the dependency-and-capability-transfer layer. This does not necessarily falsify the genre-fork at all layers; the architecture may still be justified at:

• The measurement layer (different genres warrant different rubrics, regardless of whether AI effects differ — this is the rubric-validity question, not the dependency question)
• The intervention design layer (interventions may need genre-specific calibration to work, regardless of whether dependency mechanisms differ — this is H3)
• The voice preservation layer (Authorial-Voice may be measured differently per genre, regardless of whether AI utilization affects all genres equivalently)

But at the core capability-development layer — the level the program's longitudinal claim engages — the genre-fork would not be empirically supported. The program would owe the field an honest revision: drop the load-bearing genre-fork claim; retain the genre-fork at the measurement and intervention-design layers as engineering convenience rather than empirical necessity.

This commitment is the preregistration's most consequential pre-commitment. The program is committing, before data, to abandoning a load-bearing architectural claim if the data doesn't support it. The credibility of the entire research program rests in part on whether this commitment is honored when the data arrives.

---

7. Data sources

Same as Paper 5 (§6) — Penwright production telemetry, frozen analytical dataset per Parquet snapshot pinned to the preregistration's frozen-at SHA, expert-rater Constraint-Mode quality scores, consent and demographic records from the external-operator pilot.

Genre tagging is required at every session. Writers who do not declare a genre at session start cannot have their session contribute to genre-stratified analyses.

---

8. Inclusion and exclusion criteria

Inclusion (analysis set):

• External-operator pilot participants who have completed informed consent
• Minimum 3 months of Penwright usage data at lock-in
• Minimum 3 valid Constraint-Mode samples in the relevant genre (separate cohort per genre)
• At least two writers per genre at lock-in for the cross-genre contrast to be computable

Exclusion (analysis set, preregistered):

• The principal investigator's data is excluded from the primary analysis (auto-ethnography mitigation; reported descriptively in a separate appendix)
• Writers with fewer than 12 weeks of usage in the relevant genre are excluded from that genre's stratified analysis
• Writers who withdraw consent are removed entirely
• Sessions flagged as test/demo are excluded
• Mixed-genre sessions (writer working across genres in a single session) are excluded from primary analysis but retained for exploratory cross-genre analysis

Cohort size for cross-genre contrast: the H1 contrast requires ≥2 writers per genre at lock-in. Below this threshold, only descriptive within-genre reporting is delivered; the cross-genre contrast is deferred until the minimum is reached.

---

9. Sample size and power

Target: Minimum N = 6 writers across genres (2 per genre minimum) with 3+ months of data at primary analysis lock-in. Additional writers per genre improve precision but the minimum is non-negotiable for cross-genre contrast.

Power justification (honest): with N = 6 writers (2 per genre), the study is substantially underpowered for traditional NHST inference on cross-genre contrasts. The minimum detectable effect for H1 at this N is large (d ≥ 1.0 for confident detection); the d ≥ 0.6 threshold pre-specified in §5.3 is below the detection floor at minimum N.

The honest framing (mirroring Paper 5): Paper 7 v1.0 produces descriptive evidence about whether genre-specific effects are observable in the small pilot cohort and whether cross-genre differences, if any, are large enough to motivate a larger study. The paper does not deliver a confirmatory test of the genre-fork at the small-N v1.0 stage. Confirmatory inference is deferred to v2.0 once N ≥ 18 writers (6 per genre minimum) accumulates — estimated 18–24 months past first cross-genre enrollment.

This is preregistered: v1.0 reports descriptive-only with bounded-language conclusions; v2.0 (locked separately) runs confirmatory inference. The split prevents v1.0 from over-claiming on the genre-fork question.

---

10. Primary outcome and analysis plan (v1.0 — descriptive)

v1.0 deliverable: descriptive characterization of cross-genre effect-magnitude patterns in the small pilot cohort.

Analytical procedure:

1. Per-writer-per-genre trajectory plots. For each writer in each of their working genres, plot AI-utilization composite (x) against genre-specific Constraint-Mode score (y). Visual inspection for genre-specific decline patterns.
2. Per-genre pooled visualization. Aggregate writer-level data within each genre; loess smoothing with bootstrap 95% bands. Cross-genre overlay for visual comparison.
3. Effect-size reporting per genre. For each genre, report cohort-level central tendency of writer-within-genre Constraint-Mode change-from-baseline at the highest-utilization quartile vs. the lowest-utilization quartile, with bootstrap 95% CI.
4. Cross-genre comparison. Per-genre effect sizes presented side-by-side with pre-specified effect-size threshold (d ≥ 0.6) marked. The narrative criterion: "pattern consistent with H1" if at least one pairwise genre contrast meets or exceeds the threshold in the descriptive estimates.
5. Per-prediction (H2a-c) reporting. Memoir emotional-specificity decline, nonfiction argumentative-depth decline, fiction narrative-distinctiveness decline — each reported with its own effect-size estimate, against the pre-specified within-genre threshold (§5.3).
6. Intervention-genre interaction (H3). Visualized only at v1.0 N; no formal test.

No NHST inference at v1.0. Bayesian posterior intervals reported for change-from-baseline; the posterior is the descriptive summary.

v2.0 (locked at N ≥ 18, 6 per genre): mixed-effects regression with writer random effects, genre × utilization interaction terms, and pre-registered effect-size thresholds applied confirmatorily; Bonferroni correction across the three genre-specific predictions (H2a, H2b, H2c at corrected α = 0.0167).

---

11. Multiple-comparison strategy

v1.0 descriptive only: no formal correction (no NHST). Bounded-language conclusions only.

v2.0 confirmatory (preregistered now for the future v2.0 lock):

• Three genre-specific tests (H2a, H2b, H2c): Bonferroni correction at corrected α = 0.0167.
• Cross-genre H1 contrast: tested at α = 0.05 as the primary inference; confirmed under the pre-specified effect-size threshold (d ≥ 0.6) regardless of NHST result.
• H3 intervention-genre interaction: corrected within the four-intervention-type family at BH-FDR.
• Sensitivity analysis under alternative AI-utilization composite weightings (per Paper 5 §5.1): conclusions require concordance under primary weighting and at least one sensitivity weighting.
• Exploratory analyses: no formal correction; reported as exploratory and flagged in tables.

---

12. Stopping rules and ethics

Same as Paper 5 (§11). Cross-referenced via the shared external-operator pilot consent flow.

---

13. Threats to validity

13.1 Auto-ethnography of the principal investigator. Same as Paper 5 §12.1.

13.2 Self-selection in the pilot cohort, with genre-specific bias. Threat: writers who pilot Penwright self-select; cross-genre selection patterns may differ (e.g., memoir writers more or less likely to enroll than nonfiction writers). The cross-genre comparison may be confounded by who chose to pilot. Mitigation: recruitment targeting per genre balances enrollment; demographic and prior-experience reporting per genre cohort; cross-genre comparisons reported with per-genre cohort characteristics in the same table.

13.3 LLM-scoring drift, with genre-asymmetric impact. Threat: LLM-derived scoring may drift differently for different genres (the model's ability to score memoir's emotional specificity, nonfiction's argumentative depth, and fiction's narrative distinctiveness may not change uniformly across model versions). Mitigation: expert-rater Constraint-Mode scores are the load-bearing outcome (per §5.1); LLM-derived scores are secondary and version-pinned. Frozen evaluation samples per genre rotated at v2.0 lock-in.

13.4 Genre-rubric construct validity. Threat: the per-genre rubrics (§5.1) are pre-specified but not yet externally validated against established literary-craft criteria. Mitigation: rubric calibration pass against expert-rater consensus on genre-canonical exemplars before v1.0 lock-in. Status: required pre-data; not yet shipped.

13.5 Cross-genre writer confound. Threat: writers who work across multiple genres carry their cross-genre experience into each genre cell; their genre-cell observations are not independent across the cells. Mitigation: writer random effects in primary analysis; per-genre stratified analyses repeated on single-genre-only writer subset as sensitivity check; cross-genre writer effects reported separately as exploratory.

13.6 Small-N detection-floor. Same as Paper 5 §12.6. Particularly acute for cross-genre contrasts at minimum N.

13.7 The "performative genre" problem. Threat: writers may declare genre at session start without the session's actual writing matching the declared genre. The framework's commitment to genre context (per Paper 6 §3) depends on accurate genre tagging. Mitigation: periodic blinded expert-rater audit of session-genre concordance; writers whose declared-vs-actual match falls below 0.8 are flagged for review and may be excluded.

---

14. Deviations and amendments

Same policy as Paper 5 (§13). Coordinate amendments across the paired Paper 5 / Paper 7 preregistrations when changes affect shared analytical infrastructure.

---

15. OSF upload checklist

• Create OSF project "Penwright Paper 7 — Genre-Specific Effects of AI Writing Systems"
• Upload this markdown as a component
• Link frozen repo commit SHA after first cross-genre pilot-enrollment lock
• Register DOI per institutional policy
• Mirror in the program-level OSF collection alongside Paper 5 preregistration

---

Companion artifacts in the AHI program: methodology.md (program-wide methodology); penwright-paper-04-measurement.md (Paper 4 — Measurement Framework, the indices used here); penwright-paper-06-learning-loops.md (Paper 6 — loop architecture, intervention types tested in H3); penwright-paper-05-dependency.md (paired Paper 5 preregistration — shares analytical infrastructure); peer-review.md (audience-tier 1 — threats register source). Vision specs at vela/docs/VISION-PENWRIGHT-MEASUREMENT.md and vela/docs/VISION-PENWRIGHT-AUTHORSHIP.md are the load-bearing internal sources.