Eve library

Opening Eve route: source-review backfill

Guide

Revision 1

On this page

Source revision: 1ceae52c7896a59ac74a4429d410f3218f0b91c6. Review date: 2026-09-06. This is a teaching-review record, not a live-runtime acceptance report or a declaration of complete Eve coverage.

Reviewed explanations#

Read the complete pinned Eve handbook, runtime/tools/actions, memory/grounding/continuity and surfaces/invocation pages. Compare meaningful sections with the assembled presentation content, including explanatory notes; do not count a source citation alone as teaching.

Source unit Guide / slide Defect and correction
Handbook, Status vocabulary, lines 124–138 orientation / how-to-read-evidence Separate built-but-unwired code from proposed work; teach all four definitions, every gating category and dated completion evidence. The API/adapter/interface example applies the distinction without claiming deployment.
Runtime/tools/actions, Task-family routing, lines 62–79 eve-foundations / eve-task-families Add deterministic high-confidence versus model-judged ambiguous routing and explicit escalation. Compare read, edit and hypothetical-change intents; require router tests and affected evaluation decks for verbs, nouns, examples and new families.
Memory/grounding/continuity, Four context layers, lines 28–40 eve-context / eve-context-layers Teach every purpose, lifetime and user-control expectation in the table, including memory tombstones and opt-out. Use a selected-paragraph/preference example to explain non-substitution, including operator session versus durable memory.

The existing execution guide also completely explains two reviewed units without requiring content changes:

  • Runtime/tools/actions, Confirmation bridge, lines 95–114: eve-confirmation-record teaches persisted database truth, all seven confirmation field groups, approve/decline outcomes and store/receipt-based completion. Its work-item/brief, stale-record and replay cases explain why the model's summary cannot be the approved payload or completion proof.
  • Runtime/tools/actions, Client actions and page control, lines 115–124: eve-client-action-safety covers command/selector registration, current authorization and visibility, all six structured outcomes and honest translation. Missing-target, partial-action and responsive-layout examples make the contract concrete without claiming a live execution.

These five units have reviewed assignments in eve-opening-boundaries.json. Each binds the exact source-unit hash and fully assembled slide hash, including inherited chapter and release. Later source or slide changes invalidate the corresponding review. The validator checks those bindings and citations; the editorial comparison above establishes what was actually reviewed.

Remaining source detail#

The opening route is orientation and six focused Eve guides: foundations, execution, context, trust, quality and operations. Making their existing audio current does not backfill every section or finish the handbook.

  • The identity unit distinguishes the member persona, the V1 Lilith policy substrate (persona, tone, crisis, voice and spiritual-domain policy), and V3 Lilith Metaverse. A naming overview without these responsibilities cannot stand in for the full unit.
  • The four-planes unit includes detailed ownership inventories: member feedback and fallback, operator-session credentials and workspace context, ledger/lease/MCP/report/verification, and family evals, prompt-byte ratchets, grounding/routing audits, judge validation, cost and maintenance. A four-box overview is not automatically a complete assignment.
  • Surface inventory separates docked/overlay, full-screen and voice/tour member experiences from operations, studio/workbench and leased agent entry points. Invocation includes eligibility guards, actor/tenant, selection/evidence, grants, memory, persona/release and focus restoration. The contextual-trigger checklist additionally demands event and accessibility/test evidence.
  • End-to-end requests retain policy before ordinary execution, bounded argument repair, actor/tenant reads, governed mutations, sentence-safe outcomes and separately owned persistence. Tool layers, workbench round trips, coding-agent leases and each failure/recovery case still need independent unit-level assignments. The two execution assignments above do not cover these adjacent sections.
  • Context handoff, durable conversation, Iris governance, Sophia support, retrieval, per-store data rights and degraded/conflict states extend beyond the four-layer introduction. In particular, a summary is not the transcript, a cited URL is not support, search does not replace workbench state, and a shared assistant-data label does not settle retention for all stores.

Continue these and the rest of the inventoried handbook against the same full unit standard. No additional assignment is inferred from reading a complete page, from these remaining-work notes, or from successful narration/browser checks. Comprehensive whole-monorepo, cross-cutting, Eve integration, Domains, UI/UX and end-to-end listening requirements remain in TODOS.md.