Changelog

What changed and when, newest first. Substantive changes only: a page whose last_verified date moved without its argument changing does not appear here.

Front matter carries the per-page dates: last_verified is when someone last checked that page’s sources against primaries, not when the file was edited. Vendor, market and regulatory claims age in months, and the tools pages age fastest.

2026-08

• 2026-08-22: Three vendors added to existing category pages, no new pages. SigNoz goes to observability and eval platforms as a fourth data-sovereign option (MIT core with a separate enterprise licence over ee/, SOC 2 Type II and HIPAA, a pull-only BYOC control plane) that covers tracing and ships no eval half; it is an OpenTelemetry backend and not the SIEM it is sometimes taken for. Bifrost goes to AI gateways as the Apache-2.0 alternative to LiteLLM whose enterprise tier carries OIDC, row-level RBAC, immutable configuration audit trails and log export, with two things to verify: which tier includes the MCP gateway, and the fact that its headline speed comparison against LiteLLM has no published methodology and its own benchmarks run on mocked calls. DeepKeep goes to agent runtime security for the one product there that does not need to sit in the traffic path: a free pre-deployment scanner that maps an agent’s tools and reachable paths against the OWASP Top 10 for Agentic Applications.

• 2026-08-06: FINRA Regulatory Notice 24-09 read, and threaded into eight pages. The notice had been cited twice from secondary sources and never read. Both existing citations held. What it adds is the reach of the obligation: it attaches “including through embedded features in existing third-party products,” meaning the AI feature a vendor switched on inside software a firm already owns, now noted on shadow agents and supply-chain vetting. Also new: FINRA’s own four elements for an AI supervisory policy on supervisory review; the technology-neutrality doctrine on regulatory exposure, the reason “there is no AI rule” and “you are already regulated” are both true; the two advertising FAQs that actually decide chatbot cases on agent disclosure; a pre-deployment evaluation expectation on promotion gates; three named FINRA engagement channels on exam readiness; and NIST AI RMF’s appearance in FINRA’s own resource list on frameworks. All of it binds broker-dealers only, and each page says so.

• 2026-08-06: A correction to what the wiki said regulators know about prompt injection. Two pages observed that the regulator’s threat model trails the field, correctly scoped to the SEC, whose named AI attack vectors are malware and deepfakes. FINRA published an explainer on prompt injection in March 2026 describing it as exploiting legitimate access rather than requiring unauthorised entry. The claim now names the SEC rather than regulators as a class.

• 2026-08-06: Legal and compliance threaded through the wiki; seven new pages. The wiki had been addressing a reader without authority over the decision: policies are approved by compliance, the CCO signs the annual review, counsel is in the room for any notifiable incident. “CCO” and “general counsel” appeared nowhere in 81 pages, and the front door named no regulator at all. New: regulatory exposure, built on the 2003 compliance-rule release: inadequate written procedures is a violation “independent of any other securities law violation,” so nothing has to go wrong; agent disclosure and emergent market abuse, split out of the recordkeeping page which was doing three jobs; legal escalation, supervisory review, exam readiness, and compliance as an approver.

• 2026-08-06: A regime-boundary correction to existing pages. SEA Rule 17a-4, including the WORM-or-audit-trail choice, binds broker-dealers. An adviser-only firm is under Rule 204-2, which has no electronic-format regime at all. Logging and audit opened by telling the reader they already lived under 17a-4, which for most of this audience was wrong. Corrected there and in the bibliography.

• 2026-08-06: The map now says what each page argues. Every entry outside the twelve dimensions used to be a bare link. All ~70 now carry a one-line hook taken from the page’s own claim, so the map is usable for choosing what to read rather than only for finding it. It also states, once and plainly, that the evidence on these pages is graded and the recommendations are not: read every “What to do” as practitioner judgment unless it cites something.

• 2026-08-06: MITRE ATLAS and Google SAIF verified against primaries. Frameworks had carried both as explicitly unverified. ATLAS was checked against MITRE’s own data repository at release v2026.06 (16 tactics, 101 techniques, 57 case studies, ~30 agent-specific objects) and is now called out as the entry on that page worth opening first, because it is the only one that kept pace with agents, while remaining a technique catalog with nothing to say about who signs off. The technique and case-study counts quoted in most third-party write-ups are about a year stale.

• 2026-08-06: Audit pass; all 81 pages audited. Two substantive corrections came out of a hostile read of the longest pages. Human approval gates had named the person who assigned a task as the approver, which is self-approval. The page now separates accountability for the outcome from the approval decision. And unreliable output had priced measuring a firm’s own benign failure rate at half a day; the run is cheap and the labelling is the cost, and that cost is the actual reason nobody has the number. Also: every page now has at least two inbound links, and the dead vendor-wiki tooling (SCHEMA.md, three wiki-* skills, four scripts) was removed.

• 2026-08-06: Reference section built. Crosswalk generated from page links (45 concern-control rows across 14 concerns); frameworks, glossary, templates, bibliography and reader-facing open questions written. The crosswalk’s level column is editorial and says so.

• 2026-08-06: Front door rewritten. README states the argument and nine conclusions; wiki/index.md maps the tree and documents the status labels; start-here/the-problem.md rewritten around measured evidence, the factor-of-ten measurement spread, and the two-ladder (builder discipline × artifact criticality) framing. All twelve dimension hubs written with filled Crawl/Walk/Run/Fly tables.

• 2026-08-06: Applied pages drafted. Tools (11), implementation (3), operation (6) and people (4). Notable corrections made during drafting: the leading low-code platform’s governance layer is a paid tier; the EU AI Act high-risk timeline moved to Dec 2027 / Aug 2028 and the 7% penalty applies only to prohibited practices; several circulating vendor and consultant statistics were traced to nothing and dropped.

• 2026-08-06: Depth pages drafted. Fourteen concerns, six concepts, four patterns, twelve controls, written from verified evidence bundles one page at a time.

• 2026-08-06: Verification pass. Every assertion in the extract bundles researched against primary sources; contradicted claims recorded rather than quietly dropped. Dead ends went to the build register with the evidence that would settle them.

• 2026-08-05: Extraction and planning. Source survey, claim inventory, page list with one axis per page, and the skeleton: 79 pages with front matter, claim sentences and section headers, link graph resolving, no prose.

• 2026-08-05: Prior vendor-landscape wiki replaced. The earlier category-and-vendor wiki this repository held is preserved in git history through commit e884681.

How to read a page’s freshness

• last_verified within three months. The sources were checked recently. Vendor claims may still have moved.
• last_verified older than six months on a tools/ page. Assume pricing, certifications and ownership have changed. Half the vendor map has been acquired at least once.
• status: contested. The disagreement is live and the page shows both sides with dates. Check whether one side has since won.
• status: speculative. Ahead of current practice. Nobody is doing this yet, including the people recommending it.

See also

• The map — the full page tree and what the status labels mean.
• Open questions — what has not changed because nobody knows yet.