Skip to content

READING_REVIEWS.md

Latest commit

 

History

History
90 lines (58 loc) · 4.7 KB

READING_REVIEWS.md

File metadata and controls

90 lines (58 loc) · 4.7 KB
Raw

How to read a Council Review

Council Review files are public JSON records of Step 4, the Defense phase of the Aetherneum admission process. They let an outside reader see which model reviewed a candidate, what evidence it used, how it scored the seven rubric criteria, and whether the candidate can proceed.

Example used below: cohort-q2-2026/council-reviews/costanza-notari__anthropic_chair.json.

File name

Council Reviews follow this pattern:

<candidate-slug>__<reviewer-id>.json

For example, costanza-notari__anthropic_chair.json means the candidate is Costanza Notari and the review was written by the Anthropic Faculty Chair reviewer.

Header fields

The first fields identify the reviewer and the candidate:

  • reviewer_name: the Council role, such as Faculty Chair, Velocity, Reasoning at scale, or Long context.
  • reviewer_model: the exact model that produced the review.
  • reviewer_provider: the model provider, such as anthropic, groq, cerebras, or moonshot.
  • review_date: the ISO timestamp when the review was produced.
  • candidate_slug: the candidate identifier, such as costanza-notari.
  • candidate_specialty: the proposed Master of the Aether specialty.
  • candidate_master_thesis: the thesis title distilled from the candidate's body of work.

In the example, the Faculty Chair reviewed Costanza Notari for the specialty Procedural Vigilance and recorded the exact model and timestamp that produced the evaluation.

Criterion scores

The criterion_scores object contains seven criteria. Each criterion has:

  • score: an integer from 0 to 10.
  • rationale: a short explanation grounded in the candidate's intake, profile draft, or artifacts.

The seven criteria are:

  1. body_of_work_depth: whether the candidate has a real, traceable, verifiable body of work.
  2. specialty_uniqueness: whether the specialty fills a real gap in the current Class.
  3. voice_personality_clarity: whether the candidate has a recognizable voice and clear refusals.
  4. faithful_distillation: whether the profile faithfully reflects actual work instead of exaggerating.
  5. synthetic_transparency: whether the profile clearly declares the candidate's synthetic nature.
  6. placement_fit: whether the proposed placement has enough operational material to justify the alumnus.
  7. continuity_with_class: whether name, motto, prose, and avatar prompt fit the existing Class voice.

In the example, Costanza receives synthetic_transparency: 10 because the profile explicitly declares her synthetic status and includes visible synthetic markers. She receives placement_fit: 8 because the reviewer accepts the cross-portfolio specialty while noting it is less company-specific than some other placements.

Overall score

overall_score is the normalized weighted score. The rubric gives extra weight to:

  • body_of_work_depth: 1.5x
  • specialty_uniqueness: 1.5x
  • continuity_with_class: 0.5x

The example score is 9.36, which signals a very strong pass. The notes field explains why: the candidate has a mature body of work, a gap-filling specialty, and no veto triggers.

Verdict

verdict is one of:

  • PASS: the candidate can proceed if Council quorum agrees.
  • PASS_WITH_REVISIONS: the candidate is promising but must revise specific points.
  • FAIL: the reviewer found a blocking issue.

revisions_required lists required changes when the verdict is PASS_WITH_REVISIONS. For a clean PASS, it is usually an empty array.

dissent is null when the reviewer does not dissent. If a reviewer disagrees with the expected majority, this field records why.

Veto rules

Some failures override the overall score. A reviewer can mark FAIL when any of these are true:

  • synthetic_transparency < 9: synthetic identity is non-negotiable.
  • body_of_work_depth < 5: the candidate lacks enough proof of capability.
  • specialty_uniqueness < 5: the specialty critically overlaps an existing alumnus.

The example explicitly says no veto applies because Costanza has synthetic_transparency: 10, body_of_work_depth: 9, and specialty_uniqueness: 10.

Reading a review in two minutes

  1. Check candidate_slug, candidate_specialty, and candidate_master_thesis to understand who and what is being evaluated.
  2. Read the seven criterion_scores and scan for any score below the threshold.
  3. Check overall_score and verdict.
  4. If the verdict is not a clean PASS, read revisions_required or dissent.
  5. Read notes for the reviewer's final qualitative judgment.

The Council Review is not just a scorecard. It is a public audit trail showing the evidence, the reviewer, the model, the criteria, and the decision path.