# Eve depth · editorial and reader review

The later [learning-sequence review](eve-learning-sequence.md) replaces the
overview-then-detail order below with paired topics, a separate overview route
and a final course review. The reader tasks and source-review boundaries here
remain applicable; this record describes the original expansion.

The seven detailed guides add 49 distinct slides after the six existing Eve
overviews. The progression moves from a governed entry point to a verified
result, then examines context, safety, quality, operation and dated
implementation. Visuals carry the decision or relationship. Explain supplies the
detailed reasoning and source excerpts; Script supplies the complete spoken
version.

## Teaching route

| Guide                                            | Concrete case                                                                            | Decision the reader should be able to make                                                               | Bridge to the next guide                                                                            |
| ------------------------------------------------ | ---------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------- |
| Identity, surfaces and invocation                | Add an “Ask about this row” trigger to an operator table                                 | Enumerate the exact handoff, identity, authority, memory posture and focus-return behavior               | Follow the admitted request beyond the panel                                                        |
| From request to verified result                  | Revise a draft brief without publishing it; lose the response after approval             | Distinguish proposal, persisted decision, effect and receipt; recover an unknown outcome before retrying | Separate what the request carries from what the system retains                                      |
| Context, memory and evidence lifecycles          | Restore a conversation after access changes; forget an entry that an old import contains | Preserve authorization, provenance, erasure fences and claim-level support across lifecycles             | Apply the protective policy and capability boundaries                                               |
| Authority, safety and consequential decisions    | Read and propose a brief edit while publication and spending remain constrained          | Evaluate grants, confirmation, budget, consequences and interruption independently                       | Determine whether evaluation actually exercises those boundaries                                    |
| Model selection, evaluation and change control   | Diagnose a failed task before escalating; change a model-visible tool description        | Choose the failed component, relevant families, hard invariants and reproducible release evidence        | Operate the resulting path and respond when evidence becomes stale                                  |
| Operating and changing Eve safely                | Investigate duplicate-action reports during a provider incident                          | Contain the affected capability, preserve visible truth, reconcile receipts and prove gradual recovery   | Compare the handbook obligations with the newer implementation snapshot                             |
| Current runtime: authority, voice and inspection | Read September 6 retained model-leg results alongside September 7 code                   | Separate declared limits, helper decisions, caller enforcement, measured cohorts and live availability   | Return to the appropriate architectural guide when a newer implementation leaves an obligation open |

## Reader tasks and answer criteria

Use these with a willing reader after they study the guide. Let the reader
explain the boundary in their own words and locate the relevant source. These
are prepared tasks and expected reasoning, not recorded successful reader
trials.

1. **Invocation:** the row changes tenant while the assistant is opening. Which
   identifiers are useful, which component authorizes the read, and where does
   keyboard focus go when the panel closes? A satisfactory answer names the
   typed invocation/eligibility path, server-resolved actor and resource access,
   stale or unauthorized handling, minimum necessary row context, visible memory
   posture and the invoking control. An open-drawer screenshot alone is
   insufficient.
2. **Execution:** approval was submitted but the stream disconnected. Explain
   three possible operation states and a safe next action. The answer separates
   an unreceived decision, pending work and a committed effect with a lost
   reply; it uses stable turn/operation identifiers and authoritative records
   before choosing whether a retry is permitted. A previous proposal does not
   prove which outcome occurred.
3. **Memory:** an old import contains an entry the person forgot yesterday. Why
   might removing its current row fail, and what else needs inspection? The
   answer includes tombstones/erasure fences, canonical import/recall rules,
   actor/tenant/purpose/scope, server enforcement and separate data-store
   rights. It also distinguishes selected page context, transcript, Iris recall
   and Sophia claim support.
4. **Safety:** the person approved an expensive operation, but a capability kill
   switch is active. Can approval clear the switch? The answer preserves
   separate authority, confirmation, budget and consequence checks; it gives a
   stable refusal/recovery state and requires explicit validated re-enable.
   Crisis supersession also stops incompatible continuing actions.
5. **Quality:** overall pass rate improved after a model switch, but one
   mutation family now violates a hard invariant. What belongs in the release
   decision? The answer uses family-specific floors and invariant failures,
   reproducible hashes/settings/fixtures/sample posture, validated judges and
   owned expiring waivers. Missing tools, grounding or authorization need
   diagnosis before escalation can be considered a remedy.
6. **Operations:** provider health is green again after duplicate-action
   reports. What evidence is still needed before closure? The answer names blast
   radius, narrow containment, redacted reproduction, receipt/persistence
   reconciliation, accurate degraded copy, defect-to-case intake, gradual
   restoration and any owning privacy/security process. It connects the remedy
   to an operational owner.
7. **Current runtime:** the turn cohort has two successful attacks in 120 attack
   runs, and escalation has zero in 120. What does this establish? The answer
   preserves the date, pinned model legs, 80 benign-control runs per leg and
   separately recorded control outcomes. It uses run-based outcomes, requires
   model-registry/floor matching for admission and does not infer universal
   resistance, a fresh live evaluation or enforcement outside the inspected
   path.

## Composition and editorial review

The established two-typeface system, cream surface and terracotta emphasis are
retained. Four native treatments serve distinct jobs:

- Tables compare states, owners, outcomes or criteria. Narrow screens keep row
  labels visible and expose horizontal scrolling where necessary.
- Flows make a transition or recovery sequence legible in both browser and PDF.
- Layers separate obligations that apply together without implying a time order.
- Evidence comparisons connect each responsibility to the record that supports
  it.

The numerical security table uses exact retained counts. Its labels were
shortened to “Leg”, “Runs” and “Hits” after actual-size mobile inspection
exposed fractured header words. The subtitle defines the quantity and states the
control denominator; Explain retains the detailed provenance and interpretation.
No invented performance figure or illustrative percentage is presented as a
measurement.

The escalation row uses “Escal.” for the 320px enlarged-text probe; its
provenance explicitly maps that abbreviation to the measured source leg. The
outcome-cost comparison uses “Usage” after the longer label broke a word across
lines on the phone. Both edits preserve the spoken explanation.

Comparison labels also keep a visible gap before their explanation. The labels
“Execution”, “Docs” and “Code” replace longer words that touched the next column
in desktop inspection; the detailed explanations retain the full terms. A
regression control moves an unchanged label into its neighboring text to prove
that the spacing check rejects collisions.

Evidence-comparison captions were edited to explain a relationship instead of
repeating the subtitle. Existing source disclosure, keyboard navigation, audio
controls and reduced-motion behavior remain the interaction language. Additional
photography or decorative animation would not explain these source-bound
decisions.

## Scope of the evidence map

[The source map](eve-depth-source-map.json) connects the detailed slides to
exact sections across all eight Eve handbook pages and to the later
implementation sources. These are teaching references; they are not automatic
full-section review assignments. Navigation lists, companion-document indexes
and related-link sections do not become comprehensively taught merely because
the guide links to the handbook.

[Six new assignments](../assignments/eve-depth-boundaries.json) record
individual whole-section editorial reviews with source and slide fingerprints:
governed invocation, capability grants, budgets/kill switches, release floors,
incident triage and routine maintenance. Existing assignments remain intact. The
global inventory continues to report its wider unfinished teaching scope
honestly.

A later
[whole-source development/change review](../drafts/eve-development-change-learning-map.md)
adds five compact recipe assignments without extending that claim to the eight
remaining meaningful units. It preserves exact gaps in the overview flow,
ownership map, memory, safety, contract, verification, documentation and related
sections instead of treating citation or partial summary as complete teaching.

The subsequent
[model-quality review](../drafts/eve-model-quality-learning-map.md) binds
escalation, judge validation and cost/performance to their complete
explanations. It leaves seven units open, including two cases where an otherwise
substantial slide omits a required source detail.

The later [operations review](../drafts/eve-operations-learning-map.md) binds
the complete six-outage degradation unit while retaining seven source units
whose state machine, tables, qualifiers or linked procedures remain incomplete
in the slides.

The subsequent
[memory/grounding review](../drafts/eve-memory-grounding-learning-map.md) binds
four complete continuity, memory, claim-support and retrieval units while
retaining five source units whose entity map, linked journeys, stale-state
choice or related-document roles remain incomplete in the slides.

The later [runtime review](../drafts/eve-runtime-learning-map.md) binds the
complete seven-posture failure/fallback unit while retaining six source units
whose sequence, owning links, layer detail, implementation paths or related
document roles remain incomplete in the slides.

The subsequent [safety review](../drafts/eve-safety-learning-map.md) binds
complete protective precedence, typed audit and adversarial-coverage units while
retaining seven source units whose state machine, owning links, member boundary,
disclosure, voice/privacy or related-document roles remain incomplete.

The later [invocation review](../drafts/eve-invocation-learning-map.md) binds
the complete voice, tour and proactive-help unit while retaining six source
units whose plane inventory, authority detail, owning links, checklist controls
or related-document roles remain incomplete.

The architecture guides retain revision
`1ceae52c7896a59ac74a4429d410f3218f0b91c6`. The current-runtime guide uses
`6c3186d2a4bac18f17acdbb63b3177759f32d6df`; its source excerpts and quantitative
provenance retain that revision independently. They cannot satisfy assignments
in the older source inventory. The current-runtime guide explicitly identifies
the remaining purpose-authority dimension and the limits of helper, cohort and
ephemeral inspection evidence.

Automated decode and playback establish working audio and controls. Author/model
visual review establishes an inspected rendering. Neither is a substitute for
normal-speed human listening or a successful willing-reader comprehension trial;
those outcomes are not claimed here.
