SOURCE REGISTER · INSPECTED 28 SEPTEMBER 2026
Sources & evidence.
Each claim has a date and a boundary. Current implementation references, historical demonstrations, calibration fixtures and personal anecdotes are kept distinct.
Current product reference
This presentation follows Lintel 49f2d15260f096213086f02dfeb1fae6cbe62d45. The changelog marks 0.12.0 as unreleased. This is a source inspection, not a claim of a tagged release, production certification or complete live-host acceptance.
The technical reference includes the current 96 skills, 69 agent roles and 33 hook scripts with per-item links. The portable adapter exposes 15 core wrappers; the full skill library is not 96 installed native wrappers. Counts describe source inventory, not quality or observed execution.
The Universal integration report records its bounded checks and exclusions. Removed document-format acceptance was removed, not passed or waived. A helper, adapter or hook file does not establish execution in a particular client.
Kitchen protocol v2
Source revision: e31f80ac411b6a0a6bea6de2afb5bd1d7199f1ed from the separately maintained Service House harness, inspected 28 September. No completed controlled three-arm comparison or verified native v2 rehearsal is available in the inspected evidence. No model runs were started by this presentation update.
Public exports retain the exact implementation prompt, 32-requirement brief and v2 runtime contract. SHA-256 input hashes and status record their provenance. The runtime contract supersedes only the brief’s server-start and free-technology paragraphs.
| Available evidence | What it establishes | What it cannot establish |
|---|---|---|
| Protocol v2 and exact inputs | The intended treatments, timing, submission boundary and grading contract. | A completed native run or a winning treatment. |
| Earlier naked pilot | A timeout; subsequent grading rejected after source drift. | A valid comparable quality score or completion time. |
| Earlier external demonstrations | Uncontrolled demonstrations with incomplete final evidence. | A matched three-arm cohort. |
| Historical calibration, 24 September | Reference fixture 100/100; incomplete control 0/100 under the earlier harness. | Participant outcomes, v2 live acceptance or productivity gain. |
Unknown usage and invoice credits stay unknown. Raw trials, host identifiers, private session records and billing receipts are excluded. Read the current results status or present the protocol walkthrough.
Saved demonstrations keep their original date.
The saved helper explorer retains frozen source 28061e434be455ca02f135b73244eaf4f73f3a69. It shows saved helper outcomes, not a live agent or a rerun against current source.
Older technical excerpts at 275a35447c4ad271e05816ade43ac48f1acec24f remain in the historical source archive. Existing verification manifests retain their own pins. The earlier two-dashboard comparison remains in repository history/source and is excluded from the current site export; it is not the v2 kitchen comparison.
Current Review Method and MARS explanations describe source behavior. No live MARS panel was executed for this presentation, and no accuracy or productivity improvement is inferred from its design.
The swarming story: personal, not controlled.
“I stopped being the queue.”
The supplied before screenshot and after screenshot show 55 and 711 contributions on two selected days. These are visible activity counts, not matched workload, elapsed-time, defect or quality measurements. The connection to the new swarming workflow is the presenter’s account.
“We broke GitHub.” *
*The browser tab, technically. The supplied screenshot shows PR #93: 740 commits, 760 files, +147,756/−15,126 lines and a page-unresponsive dialog. The overnight work and forgotten authorization for incremental PRs are Johannes’s personal account.
The image does not prove a GitHub service outage, the cause of the freeze or a productivity gain. The takeaway is bounded work, explicit delivery authorization and reviewable integration.
Background for the context discussion
Chroma: Context Rot and Anthropic: Effective context engineering for AI agents inform the discussion of deliberate context. They are background sources, not Lintel benchmarks, current model rankings or proof that more process improves every task.
Assets and attribution
See asset attribution. Supplied screenshots are preserved without invented UI. The selected Folded L mark and Neon, Paper and Fluent themes remain part of the presentation’s visual identity.
Presentation validation — 2026-09-28: source audit and browser checks for this refresh; not agent-host certification.