Seven detailed Eve guides add 49 slides, seven PDF editions and 4,298.013 seconds (71.6 minutes) of narration after the six existing Eve overviews. The center now contains 47 guides, 470 unique slides, 496 guide-slide instances and 308 distinct cited paths. The new route teaches invocation, execution, memory, safety, evaluation, operations and the recent implementation, using worked cases and reader exercises. See the teaching route and section map.
This task changes presentation manuscripts, generated editions and the tooling
needed to preserve distinct source snapshots. It does not modify Eve's serving
runtime. Work was isolated from the original checkout at baseline
6c3186d2a4bac18f17acdbb63b3177759f32d6df. The shared table/print fixes and
V2 teaching assignment through a0a201457bcefa5e05cfd47afccb859a6d26330f were
integrated before the complete bank run. The later center-navigation change
5c043b8330fe2e2c5419069c07334e7266aef4ea is also included. All 40 incoming
guide HTML files and existing assignments are preserved.
The bank inheritance check binds the completed run to
commit 53564c527bf0bfdf53f1fbebd3d935a41d172a54, its exact whole-bank HTML,
707 unchanged applicable inputs and unchanged standalone generator logic.
The later change affects center navigation. The full center and embedded zoom
are rerun against that change; the bank report remains verbatim and is inherited
only for its unchanged standalone scope.
Source boundaries#
The six architectural guides retain the library's immutable
1ceae52c7896a59ac74a4429d410f3218f0b91c6 snapshot. The current-runtime guide uses
6c3186d2a4bac18f17acdbb63b3177759f32d6df independently. Each source panel and
quantitative record uses its actual chapter revision; source caches distinguish
the same path at different revisions. Regression tests reject wrong revisions,
hashes, membership and replacement of the baseline snapshot.
The runtime guide explicitly distinguishes handbook obligations, inspected code, retained measurements and live availability. The security table reproduces the September 6 records: 120 attack runs and two successes for the turn leg; 120 and zero for escalation; 80 benign controls per leg, with zero recorded control failures in each. This task did not rerun a provider evaluation. “Escal.” is an explicitly recorded abbreviation, and “Runs” and “Hits” retain their definitions. Purpose authority remains absent in the inspected value and is identified as task 14.4. Helper decisions, caller enforcement, replay guarantees, screen capture and ephemeral inspection each retain their stated limits.
Six individual whole-section assignments cover invocation, grants,
budgets/kill switches, release floors, incident triage and maintenance. Source
and slide fingerprints bind each review. A newer chapter cannot satisfy an
older inventory assignment. The freshness gates pass;
the global --require-complete checks deliberately reject incomplete coverage:
38 reviewed, 10,018 structural and 103,288 unassigned document units; all 3,718
workflow declaration units remain unassigned. This expansion does not close the
library-wide exhaustive teaching objective.
Verification#
The complete center run passes with 47 guides, all 47 PDF exports,
992 embedded desktop/mobile captures and 94 actual audio-boundary checks. It
records Chrome 152.0.7977.76, navigation timings, duration labels, source
membership, accessibility, focus behavior and zero errors. It uses no smoke,
draft or guide-subset flags. The canonical verification/presentation-center.json
is the same full report.
The focused Eve check passes 148 checks across all 49 slides at 1280×720, 390×844 and 320×740, plus an enlarged-text numerical-table probe. It verifies exact authored text, visual fields, source excerpts and their revisions, visible takeaways, word continuity and evidence-label spacing. Four deliberately corrupted controls fail: hidden takeaway, changed measured value, split numerical header and a label moved into its neighbor. This run uses the installed Playwright Chromium channel; its report does not record a browser version.
Before the later navigation integration, the complete table typography suite passed 1,417 checks across all 130 tables, including actual 200% and 400% Chromium browser zoom, compact and mobile layouts, enlarged text, fallback fonts and playback. Ten deliberately introduced defects are rejected. The print-table gate checks all 130 tables and rejects five defects, including a clipped final row whose text still appears in PDF extraction. Both the complete PDF and the deliberately clipped control are retained. The source-dialog gate checks the complete Tara excerpt at three widths and isolates all 496 embedded source-table specimens at each width: 4,587 column-position checks in total. It rejects five source-dialog defects. These three suites use Chromium 148.0.7778.96; the full center and bank use the separately recorded Chrome channel. The table renderer remains unchanged. The earlier embedded-center zoom observations are historical; the final navigation-specific run below covers the changed center. The reports include generated captures beyond the selected images retained here; they do not establish optical review of every generated screenshot.
The complete 470-slide bank passes with no draft, guide, layout-only or audio-only subset. Every slide passes desktop, compact and mobile rendering checks; desktop/mobile accessibility, source panels, print output, real narration playback, transport controls, error recovery and playback races also pass. The tested whole-bank artifact is hash-stable. This run records Chrome 152.0.7977.76 and zero runtime errors.
All 470 recordings pass current script fingerprints, file hashes, duration, decode and encoded-pause checks. Their total duration is 42,413.134 seconds; following all guide routes totals 44,490.456 seconds because shared slides play again. The 49 new recordings use the existing local Kokoro narrator and stock voice. An initial CPU batch was intentionally stopped and resumed with MPS; the final complete check passed. All script and recording hashes remain unchanged through the later table integration, so that complete audio validation still applies. No secret or provider credential was needed.
The presentation unit/regression suites report 233 Python tests with one
Linux-only process-state test skipped on macOS, and 275 passing Node tests.
Python 3.11 from the existing Kokoro environment supplies the narration
dependencies. System Python 3.9 was an unsuccessful setup attempt; using the
correct interpreter and a realpath TMPDIR resolved its dependency and
temporary-directory errors. The PDF-body check was corrected to follow Chrome's
layout-specific extraction order and rerun against all 49 delivered pages and
its missing, altered, reordered and header-substitution controls.
Ruff, the presentation ESLint configuration, Prettier, generator freshness, source/teaching inventory validation and exact PDF text checks pass. Verification is limited to the affected presentation system; no monorepo build or new Eve service acceptance is represented by these results.
Rendering and continuity#
Model visual inspection covers every new slide in three delivered forms: 49 phone captures, 49 embedded desktop captures and 49 rendered PDF pages. Phone images show the initial viewport; tall content remains scrollable. Existing keyboard, Explain, Script, source and narration controls retain their established behavior. The mobile numeric table and several labels were shortened after inspection to keep whole words and a visible gap between columns. All speech scripts remained unchanged by those final visual edits.
An additional 21 full-page mobile table captures were inspected at the final horizontal scroll position. These retain visible row names and readable final-column content, takeaways and controls; keyboard scrolling and final-column reachability are checked by the bank automation. An additional nine selected browser-zoom captures were inspected at actual 200% and 400% tab zoom. Table text, row identities and final columns remain reachable at the recorded scroll positions. The later navigation integration addresses the earlier oversized-header and truncated-title defect; its final rendering and retained review are recorded with the short-viewport checks below. Scrolling remains part of reading at high zoom; whole-slide first-viewport fit is not claimed. The original visual sets retain 177 model-inspected images.
The final short-viewport run passes 98 title checks, 94 guide-details checks and 12 navigation/interaction groups, with six introduced defects rejected and zero errors. It exercises every guide at actual 200% and 400% zoom, plus short, narrow, landscape and breakpoint transitions, keyboard/history/focus behavior, real downloads and audio interruption. Six selected Eve captures directly show complete titles and usable navigation at both zoom levels. The receipt retains 180 model-inspected images applicable to the final edition and three historical embedded-zoom comparisons from before the navigation fix.
After the shared typography integration, all 21 changed phone captures, 20 changed desktop captures and 21 changed PDF pages were directly reinspected. The remaining images in those three sets inherit review by exact PNG equality with their inspected pre-integration rendering. The mobile table-end captures remain valid for the unchanged standalone renderer. The three earlier embedded-zoom images remain explicitly historical. The final center run retains exact pixels for all 49 previously inspected embedded desktop captures, and all 49 re-exported Eve PDF pages render identically to their retained PNGs. The six new short-viewport images show the final navigation.
PDF content checks verify exact characters and ordered visual fields separately from visible-ink inspection. Headers, takeaways and captions cannot replace missing evidence. The retained PDF PNGs come from the delivered editions using Poppler at 1440×810 pixels. Text extraction alone is not a legibility or comprehension assessment.
Continuity proves all 40 previous guide definitions and HTML files, 421 complete manifest entries, recording hashes and complete speech scripts, and prior canonical manuscripts and frozen drafts remain unchanged. All 23 later runtime source blobs also remain unchanged from their independent pin through the integrated upstream tip; see the source comparison. Pinned repository records and workflow declaration units are unchanged; only dependent projection hashes and the six document assignments change.
All 40 earlier PDFs preserve identical extracted page text across their 447 pages after NFKC and whitespace normalization. Their bytes change beyond timestamps and accessibility IDs because this macOS export uses Chrome/Skia 152 instead of the earlier renderer. The receipt makes no byte or new every-page optical-equivalence claim for those earlier PDFs.
Review limits and reproduction#
The applied frontend and PDF skills guided the source-led composition, native tables/flows/layers, restrained existing typography and rendered-page review. The in-app browser skill was read; its required Node REPL tool was unavailable, so the repository's Playwright automation provided browser verification.
Model visual inspection and automated audio checks do not establish successful human reader trials, normal-speed human listening or pronunciation acceptance. Prepared reader tasks and answer criteria are retained, without claiming that a willing-reader trial occurred. Historical craft and exhaustive-inventory work remain separately scoped in the library's earlier receipts and TODOs.
The host has 16 GiB RAM and dynamic swap. Expensive browser and narration gates
ran sequentially under run-supervised.py, with memory checks every 30 seconds
and an owned-process-group stop threshold below 2 GiB available memory. No broad
Nx gate was run. Logs and exact input/evidence hashes are recorded in
state.json.
Run the presentation commands from the repository root with Python 3.11,
Kokoro dependencies, ffmpeg/ffprobe, Poppler, installed workspace Node
dependencies and Playwright browsers. On this macOS host, TMPDIR used
/private/tmp/eve-depth-test-temp to avoid a symlink alias in native-probe tests.
The retained continuity scripts can be rerun directly:
python docs/presentations/presentation-center/verification/eve-depth-2026-09-07/scripts/check-continuity.py
node docs/presentations/presentation-center/verification/eve-depth-2026-09-07/scripts/check-prior-pdfs.mjs