Eve library

Eve models and quality · whole-source review

Guide

Revision 1

On this page

This map records a complete editorial read of the pinned 155-line V1/docs/eve/models-quality-and-evaluation.md. It adds three whole-unit assignments to an existing detailed guide and keeps seven other meaningful units open. It does not establish a live evaluation, serving-model state, measured economics, reader comprehension, listening or product acceptance.

Reviewed assignments#

The earlier eve-depth-boundaries.json assignment already binds the complete Floors and release gates unit. The new eve-model-quality-boundaries.json file binds:

Source unit Guide and slide Complete teaching boundary
707f44fdd23a0e806f9e5802 · Escalation, lines 97–107 eve-model-detail / eve-detail-escalation-judge Four observed triggers, capped/logged retry, family-policy preservation, stronger/different model, repair, deterministic fallback, human review and high-risk refusal.
dddaaf98d9058c6fb14eddfa · Judge validation, lines 108–118 eve-model-detail / eve-detail-escalation-judge Human-label calibration, behavior-specific false accepts/rejects, pinned judge prompt/version, disputes and all rubric-dependent evidence inputs.
878366914d7d57c11f28b22b · Cost and performance, lines 132–142 eve-model-detail / eve-detail-outcome-economics Every cost dimension, quality-before-optimization, retries/cleanup, every latency stage and the sentence-safe-streaming limitation.

The slides cite the exact headings at inventory revision 1ceae52c7896a59ac74a4429d410f3218f0b91c6. The accepted-outcome cohort is explicitly illustrative; its assignment explains the documented measurement contract and does not convert the example into an observed benchmark.

Units that remain unassigned#

The source contains eleven meaningful units. Four now have assignments across the two files above. Seven remain open:

  • Eve Models, Quality, and Evaluation, lines 1–27: existing slides teach the per-family selection idea but do not preserve the full request-to-default, drift, escalation, diagnosis and repair flow as one explanation.
  • Small-model doctrine, lines 28–43: the slides distribute these principles across several guides and do not teach the linked design's role plus all nine principles in one complete unit.
  • Runtime quality components, lines 44–56: the diagnostic slides summarize stages but do not retain every table row's responsibility and principal failure for all eight components.
  • Evaluation deck, lines 57–77: the overview slide covers most case families and the blocked-environment rule, but omits the source's budget-limit and copy/string-quality coverage from its complete enumeration.
  • Prompt and skill change control, lines 119–131: model-visible byte and family-floor behavior is taught, but the slide does not identify both linked builder inventory and standing-maintenance documents.
  • Standing maintenance loop, lines 143–149: the outcome-economics slide covers every loop action, but does not name the required Eve SMX scorecard destination for the outcomes.
  • Related, lines 150–155: no slide explains the purpose and relationship of all four linked documents. Citation and navigation are not complete teaching.

No assignment is inferred from lexical overlap or from a slide citing a source. A future assignment must teach the whole remaining unit and pass the same exact unit, revision, slide, citation, guide and assembled-fingerprint checks.