Skip to content

Latest commit

 

History

History
158 lines (126 loc) · 8.26 KB

File metadata and controls

158 lines (126 loc) · 8.26 KB

Does patina rewrite reduce AI-likeness? — Study 1 (main study) results

Companion to 2026-rewrite-efficacy-prereg.md ("Study 1" section, registered 2026-07-10 before any Study 1 data) and to the Study 0 pilot (2026-rewrite-efficacy.md). Decision rules were fixed up front; nothing below moved after the numbers arrived.

  • Run: 2026-07-10. Rewriter: claude-cli (unchanged from the pilot).
  • Judge panel (fixed pre-run): kimi (moonshot, kimi-for-coding/k2.5) + codex (gpt-5.5) + grok (xai, grok-4.5) — all cross-family to the rewriter; 2-of-3 quorum. gemini was excluded before the first call (monthly spend cap; recorded in the pre-registration).
  • Corpus generation models (Arm D AI side, family rotation): gpt-5.5 / claude-sonnet-5 / kimi-k2.5 / grok-4.5.
  • Raw texts and judgments stay gitignored under artifacts/rewrite-efficacy-study1/; only this synthesis is committed.
  • Final matrix: 576/576 panel ratings present, 0 rewrite failures, 0 unparseable cells (61 ratings arrived via the top-up pass — see Execution notes).

Design recap

Two document-length arms (snippet arms were measured in the pilot and not re-run):

Arm Language Corpus n (AI+human)
A1 en HAP-E paired on prompt_id, disjoint from pilot items 25+25 (21+21 after spok exclusion)
D ko new: 27 human documents (1.2–4.0k chars, 5 registers, 39 vetted public sources) + 27 topic-paired AI documents (public title + register + length band only — the human text never entered a prompt) 27+27

Arm D length-band enforcement: 5/27 AI documents finished slightly under their paired band after 3 attempts (worst −20%; models chronically undershoot Korean char targets) and 1 slightly over (+2.8%); all recorded (band_met:false), none excluded.

RQ1 — construct validity: PASS both arms

arm units Krippendorff α mean pairwise ρ mean abs gap
A1 (en doc) 84 0.751 0.845 14.1
D (ko doc) 108 0.526 0.568 18.4

Both clear the pre-registered α ≥ 0.4 gate. Korean document agreement (0.526) is markedly lower than English (0.751) — "AI-likeness" is a noisier construct for Korean prose even at document length, consistent with the pilot's length-penalty finding but now visible within document-length stimuli.

RQ2 — perceptual efficacy: supported in both arms, with a 4× language gap

AI-likeness 0–100; negative Δ = reads less AI-like after rewrite.

arm class n before → after Δ (95% CI) Cliff's δ "AI"-call
A1 AI 21 82.9 → 59.4 −23.4 [−30.8, −16.5] −0.79 92% → 65%
A1 human 21 17.2 → 20.5 +3.3 [−0.0, +7.0] +0.17 2% → 5%
D AI 27 83.6 → 77.6 −6.0 [−10.0, −3.4] −0.49 95% → 93%
D human 27 45.4 → 41.9 −3.5 [−8.1, +0.8] −0.13 25% → 22%
  • H2a: SUPPORTED in both arms — every AI-text CI lies below zero.
  • H2b (AI-call rate halved): FAILS in both arms, decisively in Korean (95% → 93%).
  • The headline is the language gap. On English documents the rewrite removes ~23 points of perceived AI-likeness and drops a third of the AI verdicts. On Korean documents it removes six — judges still call 93% of rewritten Korean AI documents "AI". At document length, patina's Korean humanization is currently marginal.

Anti-circularity — no gaming signature, but watch Korean

Internal-score drop vs independent-judge drop: A1 −16.7 vs −23.4 (judges move more than the internal score — the healthy direction); D −13.4 vs −6.0 (the internal score falls 2.2× further than judges do). The pre-registered trigger (internal falls, judges flat) does not fire, but the Korean gap is the direction it fires in; if ko pattern work optimizes the internal score without moving judges, this becomes the detector-gaming finding.

RQ5 — meaning and collateral

  • RQ5a: met — safety-gate pass 95.8% (92/96) against the ≥95% target. All 4 failures are dropped-number cases, recorded with the emitted text kept.
  • RQ5b: one flag. Rewriting human English documents nudges them toward AI (+3.3, CI [−0.0, +7.0], AI-call 2% → 5%). The CI touches zero, so this is a borderline over-editing signal, not a verdict — but it is the same direction as RQ4 below and deserves a targeted look before any "safe on human text" claim. Korean human controls moved the other way (−3.5, n.s.).

H6 — structural-tell survival (pre-registered primary): SUPPORTED in both arms

Deterministic keyword rubric over judges' strongest_cue on rewritten AI text (rubric fixed in the harness before any data):

arm structure lexical specificity-absence other
A1 (63 cues) 40% 11% 11% 38%
D (81 cues) 75% 9% 0% 16%

Exploratory (not pre-registered) — restricted to judgments that still called the passage "ai": A1 49% structure; D 81% structure.

The Korean surviving tells are overwhelmingly architectural, and the verbatim cues name the exact shapes: "보고서식 문장", "균일한 설명 리듬과 체크리스트식 포괄성", "너무 매끈한 문제-위기-교훈 서사", "rigid three-axis scaffolding plus tidy triadic close", "uniform exhaustive tutorial cadence, no personal fingerprint". Meanwhile, cues that flipped judges to "human" cite specificity (lived edge cases, px values, Slack reactions, first-person asides). This is the evidence base for a ko structure pattern pack: break uniform paragraph rhythm, checklist exhaustiveness, and the tidy problem→crisis→lesson arc; inject document-level irregularity rather than word-level substitutions.

RQ4 — humanizer fingerprint: now significant in BOTH arms

arm rewrite cohesion human cohesion gap permutation p
A1 0.9323 (n=25) 0.8833 (n=25) +0.049 0.029
D 0.9621 (n=27) 0.9313 (n=27) +0.031 0.045

The pilot's suggestive p = 0.078 was real: at n≈25, patina's rewrites are measurably more alike than independently-authored human texts, in both languages. patina removes the generator's fingerprint and leaves its own — a second-order tell a per-text score cannot see, and a concrete work item (voice diversification / persona-conditioned variation), not just a caveat.

Sensitivity — spok restored (Arm A1)

α 0.751 → 0.747; Δ −23.4 → −25.7 [−32.4, −19.2]. Both verdicts survive; the headline does not rest on the exclusion.

Execution notes (transparency)

  • codex quota died mid-run again — headroom was verified on a fresh reset before the first call (as pre-registered) and the recorded run still exhausted it. Unlike the pilot's Deviation 3, the pre-registered 3-judge / 2-of-3 quorum absorbed it: every affected row kept a valid 2-judge quorum during the run, and all 33 missing gpt cells were repaired by the top-up pass after the next reset. Limitation: those gpt ratings were collected hours after their kimi/grok counterparts (judge × time confound on 33/576 cells, D arm only).
  • 13 Korean document rewrites exceeded the 300 s rewrite timeout in-run; all 13 succeeded in the top-up pass at 600 s. Document-length Korean rewrite latency runs 4–10 minutes through the live CLI path — itself a product observation.
  • Top-up total: 61/576 ratings (10.6%) repaired; 5 in-run retries; zero schema-drift keys (the pilot's Deviation-2 hardening held).

Verdict

Against the pre-registered rules: RQ1 passes in both arms; H2a is supported in both arms; H2b fails in both; anti-circularity does not trigger; H5a is met; H6 (structure-modal surviving cues) is supported in both arms; H4 (house style) is now significant in both arms.

Honestly stated: patina reliably makes English AI documents read much less AI-like (−23, a third of AI verdicts flip) while preserving meaning — but on Korean documents the perceptual gain is small (−6, AI verdicts nearly unmoved), the surviving tells are overwhelmingly structural, and the rewriter leaves a measurable house style in both languages. The product implications are ranked by this data: (1) a Korean structure pattern pack, (2) house- style diversification, (3) an over-editing guard for human English input.