Changelog

2026-08-23

Published. A research section (/research, nine pages) documenting the multi-card retrieval experiments: five rounds on public corpora between 2026-08-19 and 2026-08-22, with human relevance judgements where they exist, seeded and replicated on a second machine. The section carries the experiments, the adversarial review and the five headline reversals, the approach the evidence supports, and the reading list behind it. The site's visual design was aligned with the maintainer's marketing site (warm paper surfaces, serif display headings, an ember accent, a floating pill header and a full footer).

Publication hold released (D035). The maintainer decided not to pursue patent protection for the multi-card retrieval mechanisms. The held technique entry is published with a status note that the research pages supersede it, and the hold register stays in place, empty, as the build's mechanism.

Positions revised by research.

• Multi-view (multi-card) embeddings moved from "validated in one domain, transfer conditional" to a measured result: several vectors per document beat one decisively; purpose-specific views beat matched fixed-window chunks only where queries target one aspect (+0.188 nDCG@10 at ten aspects) and lose where queries concern whole documents (−0.032 to −0.042 on human-judged scientific abstracts). Recorded as CD-25 and in the glossary, the memory-pipeline chapter, and the R14 findings.
• The economics of the two-pass design are stated as a constant-factor saving of roughly two hundred times at comparable topic granularity. The earlier claim that the advantage widens with corpus size was a clustering artefact and is withdrawn.
• The relevance gate is described as a quarantine contributing a linear factor of one to two, not as the source of the saving.

Claims investigated and withdrawn, recorded so they are not repeated: that purpose alignment rather than embedding count explains the multi-vector benefit (a blind three-word window recovers 0.610 of the 0.674 gain on LIMIT); that the right amount of result diversity depends on whether a model or a person reads the results (the measured interaction vanishes under a cross-family judge); and that conditioning the view design on the objective, the business context, the schema, or a late selection step produces a better index (worse in every form tested, including real instruction-following data).

2026-08-19

The guide was researched, written and published on this date. Entries below record what landed and, where research changed a previously published position, what changed and why.

Published. Vision and target state; the Autonomy-Learning maturity model; the size-by-gravity archetype grid; light and heavy readiness assessments; all 14 layer research tracks; seven cross-layer synthesis chapters; the use-case portfolio framework and the roadmap generator; seven department and four vertical blueprints; the vendor hub with question bank, scorecard, coverage matrix, ten profiles and adoption pathways; the documentation site.

Positions revised by research, in the order they were revised.

• The definition of an agent was reframed. Learning moved from the definition to an orthogonal axis of the maturity model, because no mainstream definition requires it (D010).
• Adoption sequencing was split into two tracks, observed and recommended, because the adoption data contradicted the recommended order (D011).
• The deterministic boundary was sharpened to "models may inform, never decide" after evidence showed three of the four irreversible zones already run probabilistic signals internally (D013).
• Compliance posture became tiered rather than uniform after the AI Omnibus deferred Annex III obligations to 2 December 2027 (D014).
• Identity levels were reassigned by capability surface rather than by platform enrollment, once per-user licensing removed the cost brake on presence identity (D015).
• Rule promotion was re-gated on counterexample survival rather than frequency, after 2026 research found generation rather than promotion to be the bottleneck (D018).
• The accountable-human sponsor moved from presence identity to access identity, because the sponsor attribute already exists one level down (D019).
• An oversight-capacity gate was added to the higher autonomy levels, expressed as a burst rate, because no credible human-to-agent supervision ratio has ever been published (D022).
• The founding metaphor was corrected in the vision chapter. The residual-work claim stands and is sourced to 1983; the headcount extrapolation was removed, and the widely circulated lights-out crew figure is corrected in the text (D021).

Claims investigated and rejected, recorded so they are not repeated: a funding story about an observability vendor that traced to an AI content farm and was contradicted by the verified acquisition record; a single-versus-multi-agent token comparison that misread a benchmark paper's domain fingerprint; and the "128 robots, nine workers" automated-plant figure, where primary reporting says several dozen workers per shift.

---

Source: CHANGELOG.md in the evidence repository behind this site.