# Eve product capabilities · 7 September 2026

The new **Use Eve** path teaches product graphs, coverage/impact analysis, work
tracking, anchored discussion and decisions, execution/delivery, operator
investigations, studio analysis/actions, member feature audits and shared member
tools. It contains twelve guides, 94 slides and 7,949.704 seconds (132.5
minutes) of narration. Every slide has detailed explanations, pinned sources and
three review questions; each guide has a PDF edition.

The full **Eve learning path** combines those guides with the thirteen existing
architecture guides: 25 guides and 204 slides without exact repetition. The
separate architecture and architecture-overview routes preserve meaningful
prerequisites. The library now has 59 guides, 564 unique slides, 587 guide-slide
instances and 344 distinct cited paths. All 47 earlier guide HTML files and all
470 earlier narration entries and recordings remain unchanged.

New teaching is pinned to `8c27a8d3350fe7484c18585b570f72a561090540`. Earlier
chapters retain their declared source revisions. The
[editorial and sequence review](../../authoring/eve-capabilities.md) explains
the dependency order, source discrepancies, capability boundaries, per-guide
craft criteria and reader tasks. The
[teaching crosswalk](../../authoring/eve-capability-teaching.json) explicitly
assigns all 90 declared assistant tools, seven executing-agent MCP operations
and 31 admitted studio view selectors to teaching. Registry completeness does
not establish live deployment availability or whole-monorepo teaching coverage.

## Verification

- The existing focused Python contract matrix passed 77 tests. The changed
  capability and PDF contract tests were rerun after their final refinements;
  their negative cases reject missing teaching, wrong pins/citations, duplicate
  captions, and lost/changed/reordered PDF evidence. Three Node baseline tests
  reject incomplete reports and unproven changes to unselected guides.
- The pinned registry collector, teaching crosswalk, builder freshness,
  source/assignment reports and repository/workflow inventory freshness pass.
  The older whole-library inventory still reports 103,288 unassigned source
  units; this increment does not silently satisfy older assignments with newer
  implementation citations.
- All 564 recordings pass the canonical fingerprint, file/hash, duration, full
  decode and encoded-pause checks. The 94 new recordings use the existing stock
  Kokoro voice and timing policy, rendered serially in six-track MPS batches.
  The final caption changes do not alter speech inputs.
- Authored semantics and visible source provenance pass on all 94 new slides at
  1280×720, 390×844 and 320×740: 282 checks, plus four deliberate negative
  controls for hidden, changed, fractured or overlapping evidence.
- Every one of the twelve new standalone guides passes its complete existing
  suite: layout, accessibility, navigation, source explanations, print and real
  narration controls/recovery/playback behavior. No draft, layout-only or
  audio-only run is substituted for these passes.
- The center suite passes across all 59 guide selections, then checks all new
  slides in the embedded desktop/mobile contexts: 188 captures, 24 audio
  boundary checks and all twelve PDF exports. Navigation remains within the
  existing local interaction budget. The 47 unchanged guide files retain exact
  proof through the ordered full-center and later focused-sequence baselines.
- All four learning routes pass 189 checks across desktop, mobile and short
  viewports, including prerequisites, search, history, keyboard route switching,
  audio interruption, malformed links, real playback/advance, PDF downloads and
  offline product/architecture navigation.
- Exact authored visual text and ordering pass on all 94 exported PDF pages.
  Ownership endpoints are represented by the printed owner columns and arrows;
  their geometry is checked separately from extracted text.
- The model author inspected 29 retained renders: fourteen standalone browser
  captures spanning all twelve guides, a final PDF page from every guide at 96
  dpi, and three route/sidebar views. The screenshots are in `images/` and their
  exact scope, PDF hashes and findings are in `visual-review.json`. These are
  selected optical reviews alongside exhaustive automated checks, not a claim of
  human comprehension or normal-speed listening research.

## Findings and evidence boundaries

The first narrow-screen passes found long identifiers and a long row label
breaking words. Readable operation labels replaced them while explanations
retained exact selectors. The complete initial pass then found one desktop
comparison-label collision; “Docs state” resolved it without reducing type.

Standalone visual review exposed 37 captions that repeated their takeaways.
Those captions now state the specific context of each comparison. The owned
verification process was stopped with exit 143 after four standalone guides had
passed; that evidence is retained as superseded, and all twelve final standalone
guides subsequently passed on the polished artifacts.

The first center run passed eight checks before a fixture selected the third
catalog entry and assumed it had prerequisites. The new introduction correctly
has none. The fixture now selects a guide with an actual prerequisite, retains
the independent no-prerequisites case, and the complete center run passes. The
failed run is retained under `attempts/center-prerequisite-fixture/`. No
interrupted or failed whole run is relabeled as a successful exit.

All expensive generation and browser commands ran serially on the 16-GiB macOS
host under the existing 30-second resource supervisor. No full Nx, application
build matrix, live provider test or backend deployment is claimed. Normal-speed
listening, willing-reader trials and the broader project items in `TODOS.md`
retain their separate acceptance boundaries.

`commands.json` records exact delivery commands, environments and observed exit
codes. `verify-receipt.py` checks current inputs, scoped reports, retained
files, PDF/source hashes and exact prior HTML/audio continuity without repeating
browser or audio runs. Run it with Python 3.11 from any directory. Publication
is the commit containing this receipt; the task closes only after its tip is
present on `origin/main`.
