AI REVIEW PERSPECTIVES — METHODS AND DATA NOTES Published with Joachim Stokke's post, 7 November 2026. https://blog.stokke.dev/posts/ai-review-perspectives/ WHAT WAS RUN The completed round-two batch contained 15 independent architecture reviews: five procedures, repeated in three randomized blocks. All reviews used the same frozen Megamind source snapshot, the same requested/observed reviewer model (gpt-6.1-sol), provider (openai-codex), high reasoning setting and Pi 1.4.0 environment. The source commit was a8d1415c1ad6427b6dc0187ae59011946de7207d. This describes a historical snapshot, not the current state of the project. The base task required source inspection, tracing data and decisions across component boundaries, evidence, a concrete failure scenario, recommendations and verification criteria. Reviewers had read-only repository tools. They did not execute tests. A separate evaluator read reports with neutral filenames, with condition labels withheld, and inspected source to adjudicate claims. This was one model evaluator, not an independently reproduced bug benchmark. CONDITIONS Plain review: shared architecture-review brief. Structured exploration: base plus explicit exploration of alternative framings before returning to evidence and selecting justified findings. Passive narrative: base plus an unrelated story, without explicit analogy transfer instructions. Narrative-assisted exploration: base, exploration, unrelated story, and an explicit instruction to consider structural analogies and validate them. Journey-assisted exploration: base, exploration, and concrete user requests and follow-ups traced through actual data, state and decision paths. The two narrative procedures received the same story within each block. Stories were generated in separate task-blind sessions without repository access. Three stories and one reviewer model family cannot establish an optimal story type or whether same-family authorship helps. Journeys were an investigation procedure. The default coding-assistant framing remained; demographic or professional identity replacement was not tested. The journey prompt already asked for passes within one review. The proposed future experiment adds completed first-stage checkpoints and separately prompted continuations from retained native sessions. WHAT THE CSV COUNTS The evaluator classified 20 deduplicated mechanisms: nine defects established by source inspection, ten evidence-supported risks, and one speculative opportunity. The CSV counts only defect entries credited as fully diagnosed. Partial diagnoses, risks and opportunities are excluded. Reference defect IDs: F1, F2, F5, F9, F11, F12, F16, F17, F18. These identifiers are local to round two and are not interchangeable with the first pilot's F numbers. One numbered report finding can cover several mechanisms; the evaluator sometimes split them during adjudication. FULL DEFECT COUNTS BY BLOCK Procedure Block 1 Block 2 Block 3 Mean Union Plain review 1 2 3 2.00 3 Structured exploration 2 2 2 2.00 2 Passive narrative 2 2 1 1.67 3 Narrative-assisted exploration 3 2 2 2.33 5 Journey-assisted exploration 3 2 3 2.67 6 Union means distinct full defect diagnoses across THREE reports. It is not the number expected from one review. The chart plots the actual three counts and their mean, without confidence intervals or a significance claim. The combined six narrative-assisted and journey-assisted reports covered all nine defects admitted to this assessment. This does not mean all defects in the repository. Earlier reviews found phone-action problems that none of these 15 reviews reported, despite the same source snapshot. RESOURCE FIELDS duration_seconds is recorded active duration for each review. reported_model_cost_usd is the provider/catalog cost estimate in the exported usage records. It is not an invoice, subscription charge or quota measure. Per-review values exclude story generation, assessment and synthesis. The 15 reviews total approximately $8.76 in recorded model-cost estimates; story generation about $0.03; the separate assessment about $5.23. Equal high reasoning settings do not establish equal computation. Actual context, cache and tool usage differed. LIMITS One codebase, one reviewer model, three repetitions per condition, one evaluator. Each report could contain at most five findings, so discovery and selection are entangled. Source-established behavior is distinct from an executed reproduction, deployed incident or measured model failure rate. Some incidental report claims were qualified by the evaluator. A full-match label does not certify every sentence in the original report. The narrative-assisted condition adds both a story and explicit analogy transfer. Its effect cannot be attributed to the story alone. The first pilot used different prompts and had toolchain problems; its measurements were not pooled into this comparison. Individual failed lookups in the completed round-two reviews were retained and can affect coverage. FUTURE TEST — NOT EXECUTED AS PART OF THESE RESULTS Run a fresh methodical review M and save its native session. Compare: A. M + methodical continuation from its checkpoint. B. M + journey continuation from that same checkpoint. C. M + a fresh independent journey review, externally deduplicated. Ordinary continuation controls for extra effort. Native context preservation tests reuse of understanding; a report passed into a fresh session is a different intervention. Remove the five-finding cap from all new conditions. Validate, deduplicate and freeze a withheld inventory of known defects; score partial diagnoses and new verified findings separately. Known-defect recall is not total defect recall. Record retained findings, serious misses, precision, actual resource use and synthesis cost. Charge full M cost to each workflow even when experimental checkpoint reuse reduces the pilot's bill. PROVENANCE Recalculated from the author's exported round-two records: evaluation/index.json, evaluation/key.md, prepared.json and usage.csv. The source repository and original reports can contain private implementation details and are not reproduced in this public data extract.