Experiment record
Structuring AI review around human technical-review expectations: design and initial validation of a two-cycle four-pass process
A PARTIAL process experiment that turns Human review failures into an explicit AI review sequence: Critical/Evidence, Lateral/Reframing, Logical/Reconstruction, and Integrated Cold Read, repeated across two clean cycles with restart after material correction.
Summary
AIEL-2026-0011 treats the review process itself as an Experiment.
The trigger was a practical failure: an AIEL article could pass the then-current AI publication review and still be reopened when a Human cold reader found material problems. The missed issues were not only factual. They included missing reader context, unnecessary management-oriented material, framing that was stronger than the retained Evidence, and the need to reconsider the whole article rather than only the reported defect.
The intervention was to make those implicit Human review expectations explicit and executable. AIEL separated semantic review into four ordered passes:
- Critical / Evidence — challenge Claims, Evidence strength, assumptions, causal wording, uncertainty, and omissions.
- Lateral / Reframing — challenge the problem framing, article boundary, inherited structure, and whether publication itself is the right outcome.
- Logical / Reconstruction — verify that Background, Question, Method, Evidence, Result, Conclusion, and Reader Value form a coherent chain.
- Integrated Cold Read — read the complete final artifact as a first-time reader without using drafting history to fill gaps.
A substantive candidate must then complete the same four passes twice. Cycle 1 — Reconstruction corrects the artifact. Cycle 2 — Verification starts again at Pass 1 using a fresh-review posture. Any material correction invalidates the clean sequence and restarts review at Cycle 1, Pass 1.
The first retained operational application was AIEL-2026-0007. That process found two material publication issues that had survived an earlier accepted state: Japanese title wording that could be read as stronger than the retained INFERRED causal Claim, and insufficient first-use explanation of the central ScriptLock mechanism. After correction, the review restarted and completed two clean four-pass cycles.
This is useful initial evidence, not proof that the process reproduces Human review reasoning in general. The Experiment remains PARTIAL while future articles accumulate evidence about Human reopenings, additional Cycle 2 findings, and the effect of the post-publication change-necessity gate.
Background
A checklist can verify that a reviewer looked at required fields without proving that the reviewer actually challenged the article in the way a Human expert would.
The practical gap appeared when Human cold reading reopened an article that had already passed AI review. That revealed several distinct review tasks that had previously been mixed together: checking whether a Claim was supported, asking whether the article was framed around the right question, checking whether the reasoning chain was self-contained, and finally testing whether a first-time reader could understand the completed page.
This Experiment does not assume that the terms critical thinking, lateral thinking, and logical thinking have one universal definition. AIEL uses them as operational review roles:
- Critical review asks whether the Evidence supports what is being claimed.
- Lateral review asks whether the chosen frame is the right frame at all.
- Logical review asks whether the selected frame forms a coherent reasoning chain.
The fourth pass is an integrated cold read because a document can pass those three focused checks and still fail as a complete reader-facing artifact.
A second problem appeared after considering already published articles. A reviewer can almost always make a good article slightly better. If every marginal improvement triggers a public edit, review itself creates instability. AIEL therefore added a separate post-publication question: not only Can this article be improved?, but Does this already published article need to change?
Methods
The Experiment used retained publication-review history and standards as process Evidence rather than physical measurements.
First, Human review findings against an earlier accepted article were treated as Evidence about missing review behavior rather than as isolated wording defects.
Second, the missing behavior was formalized into the ordered four-pass process described above. Existing AIEL controls such as Evidence-State review, whole-article reconstruction, Reader Context and Necessity, Reader Value, and bilingual cold reading were assigned to explicit passes instead of being left as one undifferentiated checklist.
Third, AIEL required two consecutive clean cycles on the same final artifact state. If a material correction is made anywhere in the required review, the prior clean sequence no longer proves the corrected artifact and review restarts at Cycle 1, Pass 1.
Fourth, the two-cycle process was applied to AIEL-2026-0007. Findings, corrections, restart behavior, and clean completion were retained in its publication-review record.
Fifth, a downstream Published-Article Change Necessity Review was added for already published articles. Findings are classified as MUST_FIX, SHOULD_FIX, OPTIONAL_IMPROVEMENT, or NO_CHANGE. An optional improvement defaults to no public mutation; publication stability does not override a material factual, Evidence, causal, privacy, or similar defect.
Future evaluation will use case counts rather than invented accuracy percentages. Useful observations include whether a Human materially reopens an AI PASS, whether Cycle 2 finds something Cycle 1 missed, and whether a post-publication finding correctly ends in change or no change.
Experiment Log
Human review exposed a process-level failure
An earlier AIEL-2026-0007 publication state had passed the then-current review, but Human cold reading later identified material reader-context and paragraph-necessity defects. The response was not only to repair those sentences. The defect class was treated as evidence that the review process itself needed a stronger model of how a Human reviewer challenges technical writing.
Four distinct review passes were formalized
The review was reorganized into Critical/Evidence, Lateral/Reframing, Logical/Reconstruction, and Integrated Cold Read. The ordering matters: Evidence boundaries are established before framing, framing is challenged before polishing logic, and the complete artifact is cold-read only after reconstruction.
The complete review was required twice
AIEL then introduced two-cycle review. Cycle 1 reconstructs the publication candidate. Cycle 2 treats the corrected candidate as a new review object and starts again from Pass 1. A material correction at any point restarts the process at Cycle 1, Pass 1 rather than continuing from the local edit.
Initial operational application found two further material issues
When the two-cycle process was first applied to AIEL-2026-0007, it found two material publication-quality issues that remained after the previously accepted state.
One was epistemic: the Japanese title used wording that could imply stronger causal proof than the Canonical INFERRED / HIGH Claim supported.
The other was contextual: ScriptLock, the central mechanism in the article, was not explained sufficiently at first use for a reader who did not already know the Apps Script execution model.
Those findings were corrected. The initial Cycle 1 attempt was discarded, review restarted at Pass 1, and the corrected final artifact completed Cycle 1 and Cycle 2 cleanly.
Post-publication stability became a separate decision
A later Human governance question identified another review risk: a reviewer may produce a valid improvement suggestion that is too small to justify changing an already published article. AIEL therefore separated finding quality from change necessity.
The resulting post-publication objective is accuracy + clarity + stability. Material defects still require correction. Marginal editorial improvements do not automatically justify article churn.
Conclusion
This Experiment converted a concrete AI-review failure into an explicit process that can be applied and observed prospectively.
The retained initial case shows that the two-cycle four-pass process found two material publication issues that had survived an earlier accepted AIEL-2026-0007 state. After those corrections, the process restarted from Cycle 1, Pass 1 and completed two clean cycles on the corrected final artifact.
That result supports a limited conclusion: explicitly separating Evidence challenge, reframing, logical reconstruction, and final cold reading—and then repeating the complete sequence on the corrected artifact—can add review value beyond the earlier AIEL process.
It does not establish that the process reproduces Human technical-review reasoning in general, that it improves review quality by a particular percentage, or that two cycles are always sufficient. The same AI system participated in designing and applying the process, so this is not independent third-party validation.
The Experiment therefore remains PARTIAL. Its next Evidence must come from repeated future publications: how often Humans still materially reopen an AI PASS, how often Cycle 2 adds a material finding beyond Cycle 1, and whether the post-publication necessity gate avoids unnecessary churn without preserving material defects.
Evidence summary
The Evidence States for the information obtained in this Experiment are as follows.
See Evidence State for the shared definitions.
| Recorded content | Evidence State | Basis |
|---|---|---|
| Human review exposed material concerns beyond factual correctness in an earlier accepted publication | OBSERVED | retained publication-review history |
| AIEL formalized four ordered passes, two clean cycles, restart after material correction, and a separate post-publication change-necessity decision | VERIFIED | retained Publishing Standards |
| The first retained two-cycle application found two material issues that had survived an earlier accepted AIEL-2026-0007 state | VERIFIED | retained AIEL-2026-0007 review record |
| The structured process can add review value beyond the earlier process | INFERRED | initial operational case, bounded to current Evidence |
| Repeated use will reduce Human material reopenings and unnecessary public churn | HYPOTHESIS | prospective evaluation not yet completed |
Unresolved questions
The current Evidence does not answer three important questions.
- Across future substantive publications, how often will a Human still materially reopen an AI PASS?
- How often will Cycle 2 find a material issue that Cycle 1 did not find?
- Will the Published-Article Change Necessity Review actually reduce unnecessary edits while preserving required corrections?
These questions should be evaluated by retaining future cases, including cases with zero findings, rather than by assigning unsupported subjective percentage scores.
Operational implications
For AI-assisted technical publishing, review quality should not be represented only by a long checklist. Distinct cognitive tasks can be separated, ordered, and repeated so that the corrected final artifact—not the earlier draft—receives the complete review.
For already published knowledge, a valid finding and a justified public change are separate decisions. This prevents review from becoming an optimization loop that continuously rewrites already adequate material.
The process is intentionally revisable. If future operation exposes a recurring blind spot, excessive review cost, or a failure of the change-necessity gate, that should become a related follow-up Experiment rather than being silently folded into the current result.
Machine-readable experiment record
A machine-readable canonical record of this Experiment is published as JSON.
AIEL-2026-0011 experiment.json
Related experiments
AIEL-2026-0007 supplied the retained review-failure history and the first operational application of the two-cycle process. Its scientific result about collector-side missingness and ScriptLock placement is separate from this review-process Experiment.