Every saved run, scored out of 60.

Sonnet 5.5 completed three passes of every prompt at four effort levels. Xhigh scored 58/60 throughout, but individual answers still changed under classifier and decision-tree instructions. More instructions did not consistently add matches. See scores, changed answers, token use and API-rate costs.

The six fresh Codex comparisons also show unstable prompt gains. Neither added-instruction prompt beat the base task in all three passes for any of these six configurations. Classifier instructions switched between a gain and a loss in four configurations. GPT-6 Luna low ranged from 51 to 55 all-four matches with those instructions; its decision-tree prompt lost two or three matches against the base task in every pass. Compare all six three-pass results.

DeepSeek Flash scored lower with both added-instruction prompts in all three passes. With reasoning off, the base prompt matched all four labels on 48–49 of 60 reviews, classifier guidance on 45–47, and the decision-tree prompt on 43–46. Testimonial agreement improved, while sentiment agreement fell. The same prompt still produced changed classifications on two to four reviews across passes. See the field-level trade-offs and repeat variation.

Across the 17 Claude configurations reviewed on 28 September, neither added-instruction prompt improved all-four agreement over the base task in every pass. Seven configurations switched between a gain and a loss for classifier instructions; seven did so for decision-tree instructions. Those groups overlap. Compare all 17 configurations and the limits of this finding.

Gemma E2B's three fresh local passes scored 35/60 with the original prompt each time, while nine reviews changed a decision. Classifier instructions scored 33–34/60; adding a decision procedure scored 31–33/60 and changed 15 reviews. The earlier local runs are kept separate. Read the field-level differences and local timing limits.

Haiku's classifier-instruction score ranged from 53 to 59 matches out of 60 across three fresh passes. Its advantage over the base prompt changed from two fewer matches to two more. Read the Haiku repeat comparison.

Opus 5 high effort gained two matches from decision-tree instructions in its first two passes, then lost two in the third. See the Opus prompt comparisons.

Sonnet 5 medium gained two and one matches from classifier instructions in its first two passes, then lost four in the third. Opus 5.5 medium scored 58/60 with those instructions in every pass, while one review changed classification. Sonnet results and Opus 5.5 results.

Across 18 Codex configurations, classifier instructions improved agreement in every pass for two configurations; decision-tree instructions did so for none. Twelve configuration/condition groups kept the same score while individual answers changed. Compare all 18 configurations.

Qwen's decision-tree prompt gained a small number of matches in all three passes. On reviews valid in both conditions, P2 gained two, one and two all-field matches over P0. P1's advantage changed direction. Two provider failures required later continuations, so this is a descriptive comparison with interrupted dispatch. See all nine results, changed reviews and retained failures.

DeepSeek low now has results for all nine scheduled passes, including failed answers. Its first P0/P1/P2 scores were 58/57/53 out of 60. The interrupted P2 includes 56 valid answers, one invalid answer and three failures; no reviews remain unsent. In the final pass, P0/P1/P2 scored 58/57/58, with 59/59/60 valid answers. These totals retain failures and do not establish nine clean repeats. See the first-pass outcomes. Compare the final passes and changed answers. Earlier reports: first P0/P1 report and interruption report.

OpenJev generated-off and generated-on have each closed all nine development phases. Across three passes, generated-off P0 scored 48, 51 and 49 all-field matches out of 60; P1 scored 48, 50 and 49; P2 scored 51, 54 and 51. Generated-on P0 scored 54, 53 and 49; P1 scored 50, 46 and 49; P2 scored 46, 47 and 42. Among reviews valid in all three generated-on passes, 8/54 P0, 2/39 P1 and 8/35 P2 reviews changed a four-field answer. Invalid outputs remain in the fixed 60-review scores. These comparisons do not establish an instruction effect. See the separate generated findings.

The Gemini repeats show why both scores and individual answers matter: Gemini 3.7 medium kept the same P0 score while changing three reviews. Gemini 3.8 medium sometimes ran out of output tokens, leaving a batch of ten reviews invalid. Read the Gemini findings and limitations.

Gemini 3.8 Flash low scored 57/60 with the original prompt in all three passes, yet changed one review's classification. The classifier prompt scored 56/57/56 and changed five reviews; the decision-procedure prompt scored 56/55/56 and changed three. Neither added prompt improved the score over the original in this series. See the changed reviews, costs and limits.

OpenJev completed all three native P0 triplets. Fixed scored 53/60 and adaptive 52/60 all-field matches per pass, with no within-mode classification flips. Thinking scored 53/60, 54/60 and 54/60; its three pairwise comparisons changed at least one decision on 9, 9 and 4 reviews. All nine passes had 60 valid outputs. Cross-mode scores do not isolate a causal effect. See the class-level findings and measurement limits.

Stable answers can still be wrong under the rubric. Three Laya native configurations returned unchanged classifications across all three passes, but each matched all four reference decisions on 0 of 60 reviews. Testimonial suitability matched on only 10–13 reviews. See field scores and class distributions.

SemIf's testimonial agreement hides missed positive reviews. All three native modes match 54/60 testimonial labels, but identify only four of the nine reference-positive testimonials in their completed passes. See the class-level findings and repeat coverage.

AnyJev raw repeated the same mistakes. All three native passes matched all four reference fields on 0 of 60 reviews. It called every review neutral and marked every review for follow-up and serious concern. No classifications changed across passes. This result concerns the raw readout; calibrated variants are separate. Read the field-level analysis.

AnyJev L0 repeated its answers, but matched all four reference fields on only 4/60 reviews. All three passes had 60 valid outputs and no changed classifications. For follow-up, it answered "insufficient information" 44 times against one such reference label. The raw and L0 methods differ in several ways, so the score gap does not tell us which change caused it. Read the class-level findings.

Alex 0.8B marked nearly every review as a potential testimonial. Each of three fresh native passes had 60 valid outputs but only 3/60 all-four reference matches. It predicted testimonial yes 59 times, including 49 reference-negative reviews. No classifications changed across the three passes. Read the field-level findings and measurement limits.

Alex 4B repeated the same strengths and misses. All three native passes matched all four labels on 39/60 reviews, with no changed classifications. It found 17 of 25 reference-positive serious concerns and missed the same eight each time. Stable answers can still disagree with the reference. See the class-level comparison and runtime measurements.

AnyJev L1 returned identical classifications in three calibrated passes, matching all four labels on 6/60 reviews each time. All 60 four-label predictions matched across passes, with zero pairwise flips and a three-pass range of 6–6/60. In each pass it returned "insufficient information" for follow-up on 41 reviews and missed 16 of 25 reference-positive serious concerns. Calibration used five held-out folds. Repeated agreement on this fixed set does not prove deterministic behavior elsewhere. Read the class-level findings and calibration method.

AnyJev L2 repeated the same classifications in three verified native passes, with 13/60 all-four matches each. All three pairwise comparisons have zero flips across 60 valid reviews. It identified 29 of 35 reference-positive follow-up cases, but also marked nine of 24 reference-negative cases for follow-up. These fitted-head results use held-out development folds; agreement here does not establish accuracy elsewhere. Read the class-level findings and timing limits.

A provider failure left one Gemma 26B comment without an answer. Its decision-procedure test stopped after six answers and one error. A separate continuation processed the 53 remaining comments. Together they produced 57 all-four matches out of 60, with two disagreements and one missing answer. This interrupted result is separate from the matched repeat study. Read the disagreements, recorded charges and timing limits.

SemIf sometimes returned the schema instead of a classification. Across three fresh passes per prompt, P0 scored 35/60 all-field matches each time, P1 scored 26/60 and P2 scored 43/60. The same eight, 22 and two reviews respectively returned the schema each time. No answers changed among reviews valid in all three passes: 0/52 for P0, 0/38 for P1 and 0/58 for P2. These repeated results describe the fixed review set. Read the generated-output findings.

AnyJev's generated JSON control completed all nine full phases. P0 and P1 returned 60 fenced, invalid responses and scored 0/60 in each of three passes; neither has valid classifications to compare. P2 returned the same 30 valid and 30 invalid reviews each time, scored 1/60, and had 0/30 observed four-field changes among reviews valid in all three passes. This control is separate from native AnyJev readouts. Read the source-bound findings.