Miadi Review

From External Reference to Author's Desk Prototype Blueprint

Oct 15, 2026 · model: manual-import · channel: JGWill

Watch source

In the chronicle

From External Reference to Author's Desk Prototype Blueprint

Source: screenwalk-261011-jev-gemini-authors-desk-prototype-handoff — JGWill https://youtu.be/ACQBJnSrtk8

Summary

In this screenwalk from the Miadi Factory, William reviews an external video about the Jev decision engine with Gemini voice and gets an architectural alignment report for Episode 120's Author's Desk interface. (TypeSafe's own documentation calls Jev a "System One model": it returns typed decisions with probabilities, not text.) He transfers the report into ChatGPT, which assesses Gemini's four recommendations and keeps the human author's judgment ahead of autonomous generation. The session ends by moving the Langfuse prompt specification from version 6 to 7 and logging observation 20 to the Episode 120 trace. This shows one pass of a pattern for bringing outside research into an agentic workflow before any code is run. One pass does not show that the pattern repeats. The figures in this review (four recommendations, v6 to v7, observation 20) and its timecodes are as summarized from the video, not checked.

Academic Fields

Human–AI Interaction (Human–Computer Interaction)

Studies how people and AI share tasks and control. Agents here surface discoveries as questions or sparks and the author decides, in line with mixed-initiative principles such as resolving uncertainty through dialog. Limit: the guidelines were tested on existing products and the principles shown on one scheduling system, so neither shows this desk keeps authorship.

Human Oversight in Software Process (Software Engineering)

Studies how people review, approve and gate software work. The session separates design, cloud prototype and local integration, and stops before running code. Limit: reviewers find fewer defects than expected, and people over-trust automated advice, so a gate works only if reviewers keep attention and understanding.

Provenance and Research Data Management (Information Science)

Studies how records of origin, versions and handling keep data reusable and answerable. The prompt versions v6 to v7 and the report logged to the Episode 120 trace are provenance records. Limit: a record shows what was captured, not that the content is correct or complete.

Narratology and Transmedia Storytelling (Literary and Media Studies)

Studies how stories are built and spread across media. The Author's Desk keeps authorial intent in charge of a story pipeline spanning video, prompt and trace. Interactive-narrative research treats authorial intent as a trade-off against system freedom and lists encoding it as open. Limit: transmedia theory states an ideal and tests no authoring tool.

Multi-Agent Systems (Computer Science)

Studies how several agents coordinate and check each other's work. Gemini produces a report and ChatGPT critiques it, with William relaying between them, so a person carries the messages. Limit: model judges show position, verbosity and self-enhancement biases, and multi-agent LLM systems fail through misalignment and weak verification.

Probabilistic Decision Support and Calibration (Machine Learning)

Studies models that return probabilities for decisions and whether those probabilities match outcomes. Jev returns typed decisions with probabilities, the kind of output a gating policy would act on. Limit: calibration is measured across groups of predictions and does not guarantee that a single answer is correct, as TypeSafe's documentation also states.

Aspects

Timecodes below are as summarized from the video, not checked. Some ranges overlap by seconds, and the review does not cover 05:25-15:10 or 17:35-22:00.

External Reference Ingestion via Gemini (00:00-05:25)

William uses Gemini voice to analyze an external video about Jev, mapping ideas onto Episode 120's Author's Desk architecture. Gemini outputs four recommendations (as summarized from the video, not checked): policy gating, privacy boundaries, dynamic mapping, and observability. [00:00-05:25]

Critical Evaluation and Authorial Sovereignty (15:10-17:35)

In ChatGPT, Gemini's report is assessed against existing history. The conclusion is to reject pure automated decision gating: AI agents should surface discoveries as questions or sparks at the Author's Desk, and the human author retains final authority. The review calls Gemini's recommendation "policy gating" and the rejected one "decision gating" and does not say whether they are the same thing. [15:10-17:35]

Separating Prototype Design from Integration (27:40-33:39)

The workflow describes a three-stage boundary: designing a feature, producing its prototype in cloud environments, and integrating it locally in the repository. The session deliberately stops short of running code, which keeps the conceptual specification apart from system deployment. [27:40-33:39]

End-results and Core Aspect

Core Aspect — Managed Prompt Specification Evolution (22:00-24:35)

Moving the Langfuse prompt specification (miadiPrototype-authors-desk-storytelling-pipeline-261010) from v6 to v7, adding Gemini's architectural points, such as evidence verification and privacy boundaries, without turning speculative recommendations into mandatory automation obligations. Evidence verification is not among the four recommendations named above, and the review does not say which one it belongs to. [22:00-24:35]

Results

  • Langfuse Blueprint V7 Saved (24:30-27:10): Prompt version v7 was saved in Langfuse, incorporating refined evidence verification, private research protection behind a public storytelling doorway, and mapping Episode 120 traces to suggested writing cards. Saving a version does not deploy it: Langfuse documents that a version is deployed by assigning it a label such as production, and the review does not say which label v7 carries. [24:30-27:10]
  • Episode 120 Trace Provenance Logging (27:08-29:15): Logged the Gemini report and the ChatGPT critique to the Episode 120 Langfuse trace as observation 20, keeping a record of the research chain across voice dialogues and prompt specifications. A log records what was captured. Langfuse identifies an observation by its id and groups observations under a trace id, and the review does not say whether "observation 20" is an id, a position in the trace, or the team's own count. [27:08-29:15]
  • Outside-Perspective Research Cycle (29:50-32:40): Wrote down a research cycle meant for reuse: reference outside material, capture a short agent consultation, assimilate it within an existing story trace, and produce structured screenwalk documentation without changing code. The video shows the cycle run once, so its repeatability is still to be tested. [29:50-32:40]

Sources