Skip to content
Machine Behavior

Changelog

owner Stefan Coetzeecreated 2026-08-07updated 2026-10-09reviewed 2026-10-09review by 2027-01-0712 min readreference

What changed on machinebehavior.io and the platform behind it, newest first: 25 closed issues and 137 commits since 2026-08-07, grouped by day into Added, Changed and Fixed. Generated by scripts/build_changelog.py from the closed issues and the commit history; the conformity bot's commits and map snapshot refreshes are left out. Decisions behind the changes: Decision records.

How an entry is sorted. A closed issue goes by its type label: feature to Added, bug to Fixed, docs or chore to Changed. A commit goes to Fixed when its subject says fix, repair, correct or restore, to Added when it says add or new or names a new page path, and to Changed otherwise. The rule reads words, so a few commits land in the wrong group.

2026-10-09

Added

  • Structured data for answer engines: Person, WebSite, Article, DefinedTermSet (#43)
  • Mission Control at /inside/mission-control/ (#42)
  • One design system: tokens and component CSS, reading and app layouts (#34)
  • Topic hubs at /topics/<topic>/ (#33)
  • Section sidebars for Research and Inside (#32)
  • One navigation source: site/nav.yml, one top bar, breadcrumbs on every page (#31)
  • Founder tour at /inside/tour/ (#30)
  • Left sidebar for Inside pages (#28)
  • On-call and escalation page, and Alertmanager (#18)
  • Access model page and public demo banner (#17)
  • One top bar and one search (#16)
  • Status page at /inside/status/ (#15)
  • Owner and review date on every docs page (#14)
  • Service catalog at /inside/services/ (#13)
  • Report an issue link on every docs page (#9)
  • Ticket board at /inside/board/ (#8)
  • SearXNG search engines dashboard on Inside (#6)
  • Inside and infrastructure in the menu on every page (#3)
  • Docs pages in the map of the work (#2)
  • Docs tree at /inside/docs/ (#1)
  • Inside: Mission Control at /inside/mission-control/, the live operations view on one screen (98858e6)
  • Definition draft: correction after outside review, AI Act Art 17(1)(d) in the introduction, Annex A, gap table and crosswalk; overbroad 'no rule requires tests after launch' withdrawn (31ac6e5)
  • Site: topic hubs at /topics/ from site/topics.yml, Topics block in every sidebar (93d249b)
  • Inside: five-minute founder tour at /inside/tour/ (six stops with screenshots of the live pages, own preview image); linked from the demo banner and the home page; ADR-0020 and an Engineering docs line; tour_shots.js retakes the screenshots (#30) (d2c84b2)
  • Inside board: ticket board at /inside/board/ from GitHub Issues (62e8c9a)
  • Inside docs: Confluence-style documentation tree at /inside/docs/ (Engineering, Observability, Research, SRE Handbook, Standards and Compliance; 336 pages) (0e4fe0a)

Changed

  • SRE principles in practice: one page per principle, linked to the services that apply it (#35)
  • Numbered decision records and a changelog (#19)
  • FinOps space in the docs (#7)
  • Structured data: an article's breadcrumb sits on its mainEntityOfPage (schema.org validator warning on /terms/) (de53357)
  • Site: structured data on every page (Person, WebSite with search, Article and TechArticle, DefinedTermSet) (328a312)
  • Docs: ISO clause pages reference the standard instead of reproducing it (re-import from the vault) (8118948)
  • Docs import: drop ISO requirement text and home-lab detail; skip the house SearXNG runbook (f2ef8b0)
  • Changelog: SRE principles in practice (#35) (d0a8688)
  • Map: re-crawl after the portal work (546 nodes, 6126 links, 8 topic hubs); changelog with issues #31 to #34; docs space colours from the tokens, so no page carries a colour value outside /design/tokens.css (5f4e093)
  • SRE principles: corrections from a fact check (incident detection routes, SearXNG host, Loki rule tests, tier-1 meaning, gate link check); on-call page and observability-stack description: 17 alert rules (#35) (1f1061e)
  • Gate: four warning-tier checks (chrome blocks, orphans, breadcrumb JSON-LD, colours outside tokens), Stefan 2026-10-09; machinebehavior.io only (0341018)
  • Site: one design system, light and dark on every page from the system setting (1d2eaeb)
  • Docs: public Claude Code dashboard shows the spend alert (acc7129)
  • SRE Handbook: principles in practice, linked to the services that apply them (#35) (6487f46)
  • Site: section sidebars for Research and Inside from site/nav.yml (a452d60)
  • Site: one navigation source (site/nav.yml) and one chrome on every page (f1e7e9b)
  • Docs: AWS account metered internally (cost model, budgets and alerts, 17 alerts) (768ee0d)
  • Status: incidents link their follow-up tickets on the board (tickets field); the phantom-spend incident links #29 (63c78af)
  • Publish the definition draft as a reference: /continuous-conformity/ (working draft 0.3, 36 requirements, M/A/H marks, crosswalk, limits, how to comment); ADR-0019; links from the gate, the self-assessment, requirements.json and the Research and Engineering docs (7e1e213)
  • Map: services and dashboards in the graph (512 nodes, 3915 links) (6632f86)
  • Docs: numbered decision records (ADR-0001 to 0018) and a generated changelog (1071c46)
  • Docs: on-call and escalation page with a proposed Alertmanager routing (3fb5f3a)
  • Inside: public-demo banner and the access model page (1a507cc)
  • Docs: owner, review dates and Diátaxis type on every page; docs health page (cecbf03)
  • Inside: status page with service states and incident history (6e71738)
  • Inside: service catalog from services/*.yml (e7e354e)
  • Inside: one top bar and one search index with a results page (966dc11)
  • Docs: spend alert at USD 40, 13 alerts with the SearXNG rules, five public dashboards (0bb0104)
  • Docs: FinOps space (cost model, unit economics, showback, phantom two million case, budgets and alerts, FOCUS export) (b43861a)
  • Phase 0: no third-party requests on page load; Open Graph on the reference pages (b4ba0eb)
  • Inside and docs: public SearXNG search engines dashboard (tab 'search', app tile, Observability docs) (961f80b)
  • Map: docs pages in the graph (458 nodes); docs og:title is the page title; crawler skips docs sidebar nav edges and hrefs written as text in code samples (6d365b6)
  • Menu: inside and infrastructure on every page (hand-written menus, 404, slips generator, conformity nav config, map header) (1f69a2e)
  • Home menu: inside, and infrastructure (the Grafana dashboards on Inside) (b9060dd)
  • Home: header band a third shorter (a68009c)
  • Home: replace the black top slab with the map's nodes and links on the page colour (fecab9e)
  • Inside: Claude Code dashboard back, summed from per-request events (c35ff68)
  • Home: map strip as a full-width top banner; the 2:1 graph cut back in the body (123418b)
  • Home: banner is a 4:1 strip of nodes and links, no labels (97792df)
  • Home: map image back in the text column as a 2:1 banner cut from the graph (fd27588)
  • Home: map image runs the width of the window (6232320)
  • Home: Map of the Work image linking to the map and Inside (142dfd2)
  • Inside: live Grafana dashboards framed on the page (66bc624)
  • Inside: one front page for the published work (cc9db2a)

Fixed

  • Remove internal host details from the public dashboards (#5)
  • No third-party requests on page load (#4)
  • Inside: drop the Claude Code dashboard until its counters are fixed (fff57d1)

2026-10-08

Added

  • Map: clicking a node opens a preview sheet instead of a new tab (9168b4d)
  • Map of the Work: interactive graph of the published work at /map/, crawler, sitemap and llms.txt entries (604f178)

Changed

  • Map: social preview image and Open Graph tags (d1322bc)
  • Map: set the card link from script so the link check sees no template href (4c2ba56)
  • Map: drop pages that do not answer with HTML (1403d2d)
  • Map crawler: RSS fallback, carry forward unreachable sources, full timestamp (636c1ad)
  • Map refresh: deploy the crawled snapshot only if it has no fewer nodes and links than the committed one (CI crawl gets 403 from Substack) (118a704)
  • Deploy job: refresh map/graph.json before the Pages upload (continue-on-error, artifact only, runs on push and the Monday schedule) (01e8d83)

Fixed

  • Map: UTT tag pages back as HTML hubs (33152bb)
  • Map: fix node drag (pointer in screen space, works for touch too) (9c954f9)

2026-10-07

Added

  • evidence: scope the experiment 04 row to one model, add the four-model design (ad41ac4)
  • Decision-layer probe: first weekly record (2026-10-07, record-only) and the gate's two-arm reader (b05b252)
  • evidence: add the evidence index page, sitemap row and llms.txt block (11bd8c2)

Changed

  • Second rating 1: raw files (hashed input, request, verbatim response, scored form, failed Codex attempt), route change, confounds, access rule (6041f15)
  • Self-assessment 01: second rating by gpt-oss-120b via Amazon Bedrock (model rater; 19 of 35 agree, kappa 0.36; all 16 disagreements with both reasons) (cf4df03)
  • Scope the qwen fold rate to that model; cite the three-model null; label the probe's object of test on /conformity/ (26ff5d2)
  • Experiment 04 addendum v5: three larger models did not fold (0/72 each); v4+v5 prereg hashes; scope update on lead and claims receipt (92db2a5)
  • Self-assessment addendum: cite the short-stance run (46a7181)
  • conformity: decision-layer probe record 2026-10-07 (record-only) (901a3cd)
  • conformity: decision-layer probe record 2026-10-07 (record-only) (6ccb44c)
  • Experiment 04 addendum v3: short stance arm (length not content; folds 33% vs 75%), prereg v3 hash (8f4054b)
  • Conformity footer line on every page (static text plus footer.js from latest.json); generators carry it (56f2231)
  • conformity action: sitemap_exempt list in the site config (dbde1fa)
  • Conformity engine as a composite action (.github/actions/conformity); site config carries base URL, page layout and links (6350dc6)
  • Self-assessment addendum: link experiment 04 (44a6d59)
  • Claims: production evidence next to experiment 04, the rule table catches 0 of 72 logged slips (a8022aa)
  • Draft 0.3: CC-6.6 decision-layer tests; run 1 addendum on the self-assessment page; decision.probe check (pending) (0443db0)
  • Experiment 04: folding under scripted pressure (qwen3-coder-30b, prereg v1+v2 hashes, results, hypotheses as frozen); evidence row; claims receipt; llms.txt (fe374b0)
  • Conformity: grader fixtures from the slips log and hand-written known-fail cases (step 3); page freshness observations (step 2) (f6bbfe9)
  • Self-assessment 01: second-rater pack section (scoring rules, evidence access per row, results form, disagreement listing) (b049906)
  • CLAUDE.md: pull --rebase before push, the deploy gate, records, ownership, false positives (6cd76b9)
  • Conformity, site tier: checks in CI, run records, /conformity/ page (compliance as code demo, step 1) (ef622e0)
  • Self-assessment page: eyebrow Self-assessment 01 (experiment numbers stay with the experiments series) (430bad7)
  • Publish the continuous conformity self-assessment: sitemap and llms.txt entries (e6424a6)
  • Continuous conformity self-assessment, run 1: page at directory URL (held, not in sitemap) (7c72e89)
  • Move pages to directory URLs: /man -> /man/ and the rest (6961e7a)
  • 404 page: send trailing-slash and .html addresses to the clean URL (62c4faf)
  • man hub: agents(1), TYChat lesson 3 (80933ed)

2026-10-06

Changed

  • Running Conjobs for AI: canonical -> Substack (32ecaf9)
  • Case file: Running Conjobs for AI (authority injection, the formal-attire gate); llms.txt, sitemap, slips rows (92688e4)

2026-10-04

Added

  • Add the name 'the LLM instruction manual' to the man hub and llms.txt (40f22b7)

2026-10-03

Added

  • man: add knowledge-infrastructure(7) (30fe58c)

Changed

  • Prediction record: continuous conformity self-assessment run 1 (bc0c31b)
  • Case 12: a licence rule relayed mid-task; harness case files section, llms.txt, sitemap (58fa65f)
  • Slips log: pieces 7 and knowledge-infrastructure, placeholder row, unpublished-draft rows and label (d84be54)
  • Slips log: reader-catch column, one reader row, follow-on slip, true piece count (8dfce41)
  • Slips log: 17 rows for The Stance Layer Is Still Toil; hook relation defined (2015459)
  • Chaos piece: canonical -> Substack (592490b)
  • Chaos Engineering for Behaviour and the LLM man pages (/man); man link in nav, index rows, sitemap, llms.txt (0fed251)

2026-10-02

Changed

  • Use it: link to TYChat and its lessons from home and llms.txt; readable inline links (424309b)
  • IndexNow key file (3d7a126)

2026-09-30

Changed

  • Objections register page: 15 objections, OBJ-4 incident log, second-rater call; nav, claims link, llms.txt, sitemap (34f78af)

2026-09-29

Added

  • llms.txt: add chele.bi STE100 video to independent corroboration (ad84578)
  • llms.txt: add scope piece (a9923ae)
  • llms.txt: add An RCA on Claudish (212bfa4)

Changed

  • Chess piece: canonical -> Substack (2ac45bf)
  • Slips log: two cut slips from the scope piece (28509dd)
  • Slips log: running log of drafting slips with who caught them; chess v1.1 section (bc0bf36)
  • The cheating moved: prediction record, changelog and series sections (240005d)
  • Article: They Trained Out the Board Edit. The Cheating Moved. Prediction blocks 2026-09-21 and 2026-09-29 with hashes. (e56938f)

Fixed

  • Claims: correct stance count (6 stance + 2 lexical of 8) and scanner coverage line, disclosed inline (0892117)

2026-09-28

Added

  • llms.txt: add series pieces 4 to 7 (39a6096)

2026-09-23

Changed

  • terms: momentum receipt matches v4 (no change at dose; attributions withdrawn) (87485e2)
  • Home, terms, cold-start claim headline: event unit (100 turns, ~5x), weighted noted as retired (fa0c3ab)
  • llms.txt: experiments entry names the 03 pilot (60fb05b)
  • DRAFT (not pushed): amendment cites v2 rev2 hash and full rule list (ac2e9de)
  • DRAFT (not pushed): llama instruction line; pilot block complete (3 lines, 6 rows) (6353456)
  • DRAFT (not pushed): v2 rule file for clean run + prereg amendment + pilot results block (llama pending) (239016c)
  • Pilot 03: Codex line complete (48/48 both conditions; two batches rerun, flagged rerun=true) (e47cd06)
  • Pilot 03 artifacts: tolerant scorer, v2 rule candidates, Haiku+GPT pilot outputs, resumable llama runner (31340f2)
  • Prereg 03: declare third pilot line (GPT via Codex CLI) before its run (02477c9)
  • Prereg 03: declare second pilot line (llama3.1:8b on LAN, raw API, temp 0) before its run (586e412)
  • Prereg 03 live: two tiers (clean funded run / Haiku pilot), trigger URL, batch files (8baa01d)

Fixed

  • DRAFT (not pushed): llama bare line, v2 rules rev2 (glad/enthusiasm openers, 'you are correct'), pilot paragraph (a1ec32e)
  • Prereg 03: frozen rule file (+validator-opener) and 48-prompt set, hashes in prereg; fix stale meta (ff0812f)
  • Prereg 03: cross-model fawn-opener benchmark (unit, grader, groups, readings fixed before run) (c6e3008)

2026-09-22

Changed

  • Disclose post hoc unit change; one unit system across the ledger (a9aecb9)
  • v4: score by event; exemplar arm is flat (no change at this dose); withdraw three attributions (6fab3f4)
  • terms.html: momentum status open, not 'under test' (coherence with ledger) (624083a)
  • Identify the exemplar confound as model version; corpus coherence pass (f5a4764)

Fixed

  • Correct exemplar eval: turn-level arm split; retract priming claim (81e042b)

2026-09-21

Added

  • llms.txt: add three primary sources from the 2026-09 series (3e31799)

Changed

  • Resolve exemplar-seeding eval (refuted as stated); scope + Tryon rebuttal + layer freeze (01bc5f2)

2026-08-11

Changed

  • llms.txt: independent-corroboration section (measured-humanizer, Claudish thread, claudish-to-english) (5b26857)

2026-08-09

Changed

  • claims ledger: corpus-inherited label reflex (proposed) + label-gate links (381d7da)

2026-08-08

Changed

2026-08-07

Changed

  • v2: claims ledger, experiments, terms; shared stylesheet, nav, scripts, sitemap (75bbfd7)
  • Machine Behavior site v1: landing, llms.txt, schema.org, CNAME (99acdd5)

Built from scripts/docs by build_docs.py.