Skip to content

QA ledger — AgentFabric overview

What was run, what it said. Append a dated entry per regeneration; do not edit past entries. An entry without a command and its output is not evidence.

2026-08-30 — initial AgentFabric edition

Package created by adapting the Literate-AI docs/presentations/ authoring package to AgentFabric, then replacing the deck, the narrative, and all specifications. Built against jordanhubbard/mac at 2976182.

Build

docs/presentations/agentfabric-overview/regenerate.sh
  • build_deck.py — built 20 slides -> docs/presentations/agentfabric-overview/agentfabric-overview.pptx
  • build_narrative.py — built narrative -> docs/presentations/agentfabric-overview/agentfabric-overview.docx (67 headings)
  • render_slides.py — 20 PNGs plus contact sheet, no text-frame overlaps detected
  • verify_pair.py — pair verified: 20 slides, notes on all of them, 67 narrative headings

Gates

Gate Result
Slide count matches specification (20) pass
Speaker notes present on every slide (20/20) pass
Text-frame overlap detection pass — none detected
Narrative heading spine, no skipped levels pass — 67 headings, 0 skips
Manifest claims no unauthorized publication pass — both members publication_authorized: false, published_location: null
Stale-source references (literate, litai, component://, case-insensitive) pass — the only remaining matches are the provenance notes in this ledger
Native diagrams only, no assets/ directory pass — package contains no images
No invented ROI / productivity / throughput figures pass — reviewed slide-by-slide against source-notes.md
make lint (ruff check + format --check) pass for this package (14 files, check clean, format clean). The repository-wide run also reports 15 pre-existing unformatted files under tests/, none of them touched here
scripts/run-contract-tests.sh documentation contract and environment registry pass; 11,136 tests pass, 44 fail. All 44 are the merge-queue / git-publication families the runner itself warns about on a host with git 2.34.1 (git merge-tree requires ≥ 2.38); unrelated to this package, which adds no runtime code paths
scripts/test-docs.py fence contract pass — console fences only under docs/; a bash fence in SKILL.md was rejected and changed
scripts/generate-docs-reference.py --write regenerated; docs/reference/documentation-inventory.md now lists the nine package documents

Defects found and fixed during this pass

  1. Slide 13 overlap. The terminal/held chip row collided with the ALSO TERMINAL OR HELD: label. Chips were shifted right and narrowed; the re-render reports no overlaps.
  2. Linux font fallback missing. render_slides.py only searched macOS font paths and silently fell back to PIL's bitmap default, producing an unreadable contact sheet. Liberation Sans and DejaVu Sans paths were added.
  3. Slide 18 counts were inherited, not measured. The copied deck's figures (219 modules, 430 routes, 125 CLI verbs) did not match the tree. Re-measured to 221 modules, 435 routes, 458 CLI leaf commands, 12 task states; commands recorded in source-notes.md.
  4. Metering was overstated. The deck claimed every model call is metered at the router and listed router-side metering as implemented. ADR 0017 is Proposed and quantifies a 29.5 per cent unmetered fraction over the seven days to 2026-08-19. Slides 4, 16, and 19 now say route events are recorded and priced at read time, with enforcement listed as proposed and the coverage gap quoted with its window.
  5. Named gates were overstated. Slide 9's claim now carries a speaker note distinguishing the implemented review gate (ADR 0011, Accepted) from the general contract (ADR 0022, Proposed).
  6. Publisher pointed at another project's deck. publish_google_workspace.py defaulted to the Literate-AI Slides ID and named the Literate-AI narrative. The default was removed; --slides-id and --authorized-by are now required.
  7. Regeneration referenced a missing verifier. Both regenerate scripts called scripts/verify_document_pair.py and a components/literate-ai-overview/ component file, neither of which exists in this repository. Replaced with the package-local verify_pair.py; the unused Codex build_deck.mjs fallback path was removed.

Visual review

Contact sheet inspected at _build/agentfabric-overview/contact-sheet.png. Confirmed: marketing-first opening, NVIDIA and OSS re-use inventories, unique mechanisms, trust boundary, actor roles, task lifecycle, AgentBus coordination, heterogeneous fleet, route ladder, evidence flow, honest status columns, adoption close. No slide is bullets-only except the two deliberate project inventories.

Publication

Not published, not authorized. No Google Slides or Docs resource exists for this package.