2026-09-12-SAVE-STATE-agent-roster-session.md

4,098 bytes · modified 2026-09-11 23:52

SAVE STATE — agent roster + ICM routing session (M2, 2026-09-11 → 09-12)

> Written so the session can compact and resume. Everything below is on disk (M2 + Nano, md5-matched unless marked). Bridge does not store — this file is the handoff.

Resume in one command

node ~/AI_OS/icm/route-agent.mjs "<next ask>"     # who takes it · then read agents/SPIN-UP.md once

What exists now (all shipped this session)

Open items (next session picks from here)

1. Mini sync — unreachable 09-11/12 (ssh timeout ×3): copy agents/, icm/route-agent.mjs, icm/agents.json, icm/router.mjs, workspace/ops/stacks.json, the two packets; as insightfulmind@100.72.22.100, tar-over-ssh + md5. (A13)

2. Token cuts, in order (OPERATOR-STEERING §2): slim process block + copy prompt (generator edit) → client-section.sh <slug> → extract CLAUDE.md sections for A9/A11 → A0 reads icm/agents.json not ROSTER → DESIGN-INDEX as a true catalog. Target: fleet ~45k, median spin-up ~2.3k.

3. SYSTEM-MAP §8 fixes: Nano A0 inbox 200 open cards (no consumer); ~60 unloaded Nano plists (REPORT-ONLY manifest); executor canary; Mission Control ghost; skool-monitor failing since 08-23; icm/v2 → Nano/Mini (S7); wire headless spin-up.

4. Cold-walk test of ROSTER → one card → stop (A3), and route-agent --selftest fixtures (A1).

5. Comp #12 The Auditor — CMS §1140 Medicare-marketing auditor folder as the entry (A7 + A4); link goes in the thread the day it has three real findings (lesson from Comp #11 ruling). Sean GO pending.

6. Sean-only: read skool/comp11-feedback/sean-vargas.md; Lyceum seat decision (cohorts started 09-12).

Rules set this session (binding)

• Every ask is routed before it is worked (route-agent) — announce / ask once / stay A0 / NEEDS-YOU.
• Ask early, once: batch every foreseeable question before step 1; never stop mid-process for a foreseeable one.
• Spin-up = card + Always set, ≤3k target / 8k ceiling; never widen the load list.
• Council = six seats on A3, not agents. IDs never change; names are overlays.

TAKEAWAYS.md

18,719 bytes · modified 2026-09-11 23:37

Jake Van Clief + Clief Notes (Skool) catch-up — 2026-09-11

1. The ten videos (newest first) — one lesson each

Recurring spine across all ten: one strong model + a good harness + a folder that carries your judgment; agents only where a decision must be made; gates are where the human re-enters; "in a world full of cheap answers, questions become valuable."

2. What was posted on Skool (08-24 → 09-12)

Jake's own posts

• 08-25 My research outside ICM (151 up) — SSRN preprint with Edinburgh Neuropolitics: Ideological Distortions in the Digital Public Sphere: An Audit of Authoritarianism in Commercial LLMs.
• 08-27 For my Alpha Testers (118 up) — the EdubaWare MCP (video above).
• 08-28 The whole "AI is moving so fast" trope bothers me (152 up, top post of the window) — if model updates scare you, you are building at the wrong abstraction; "best" is a position formed by a person, not a fact (Weinberg 1971); that is exactly why a model can't converge on it and why your judgment in files is the durable asset.
• 09-05 New Workflow Audit (144 up) — copy-paste interview prompt (one question at a time: what comes in, what you do, where it gets messy, what goes out, decisions to keep) → builder skill + blank HTML template → editable workflow package: map of the work, evidence for where improvement matters, open questions, prompts, recommended ICM folder. Lives in The Vault ("Workflow Audit Package"). Ry Mac already used the interview half to declutter a member's 43-copy Claude Code mess.
• 09-07 What did you actually build this weekend? (123 comments) — three days into Implementation Playbooks; "not what you watched, what you MADE."

Weekly comps / leaderboard

• 09-03 Comp #12: The Auditor ($500) — a folder-based auditor that checks an artifact against an external, citable standard (WCAG 2.1 AA, GDPR, SOC 2, HIPAA, brand guideline, FAST, App Store…). "Every finding cites a provision the way every card cited a line. The standard is the territory."
• Leaderboard winners: Alex (08-24), Ry Mac (08-31), Nuno Silva (09-07). Prize = lifetime of your current tier.
• Lyceum: "7 seats left, doors close Saturday" (09-02); cohorts start 09-12. Vibe Film webinar session 2 = 09-12 (Danney Trieu; six levels: prompt → reference → memory → blocking → DNA → autopilot).

Member threads worth stealing from

• 09-08 Curtis Hays Your 60% is documented. Is it load-bearing? — for every process ask what specifically breaks if you stop; name a checkable consequence; the 60/30/10 ratio applied to people and process.
• 09-10 Andy Miles dumbest thing quietly wrong — agents stamped dates from the session header (frozen at session open), eight files three days early, staleness check compared two wrong stamps. Fix: read a clock at write time; no shell = no date, don't guess.
• 09-11 Michael Steve ICM stopped my 4 platforms drifting — one parent "Ecosystem" folder above four platforms: map of what lives where, standing rules, one agreed fact set, every cross-link, open issues, a diary of checks; nothing goes live until he publishes.
• 09-11 Leo Saraiva memory bigger ≠ better — golden recall questions written before you need them; date everything, newer wins unless marked decision; keep decisions apart from transcript; removal is a normal operation.
• 09-12 Jeremiah Stephen False control — a written rule is not a control; no current handoff = cannot start the next stage (machine-enforced); exactly one authoritative record per fact (registry / project state / handoff / decision log / CI artifact / provider / issue); ICM needs self-tests (cold-start, handoff points to a thing that exists, context packet stays small).
• 09-08 Theo Boomsma context windows filling up even with ICM (31 comments) and Azerhan Turan over-defining stages — stages as interpretable scaffold, not algorithm; specify less, evaluate more.
• 09-08 Jeff Van Leenen Claude Code vs Claude Design — Design wins breadth (tokens, components, kits, three logo directions) but is walled and expensive; Code is faster, portable; the folder structure is the asset you keep.
• 09-08 Michael Steve "No AI" policy just hides usage — amnesty first, then three enforceable lines (client data never in unapproved tools; nothing AI-drafted leaves without a named human owner; people decisions by people).
• 09-10 Colm Whelan a real client, one meeting — his five-point stance (overhyped / undersold / you don't need a swarm / scalding bath, toes first / that puts you ahead), record the meeting → transcript → Claude process map.
• 09-01 Ruben Aguirre Claude "lying"… trying Codex (114 comments) — the hallucination-guardrail thread; 09-01 Shirsho Claude Fable 5.1 is out.
• 09-09 Jose Guerrero — watch the iScaleLabs "WATCH THIS FIRST" breakdown before anything else (ELI5 of ICM).

Classroom (full re-scrape 09-11 vs 08-22: 155 lessons, 4 new, 0 removed) — The Vault → 📦 COMP #11: THE CARTOGRAPHER results package (video + feedback.zip, one file per entrant); Drawing Room (VIP) → Session 16 (Talks from Dubai); David's Corner → AI Concepts for Humans: Terms That Actually Matter and The Tool Is Ten Percent (the 60/30/10 article Curtis Hays' load-bearing post builds on). The Workflow Audit Package module also landed in The Vault (09-05). Scrape archived at vault/research/skool-cliefnotes-2026-09-11/.

3. Week 11 = Comp #11 "The Cartographer" — the takeaways

The brief (08-15, prize = Lyceum seat, due 08-25). ICM's sixth form: the system map. Walk a body of work someone will change and leave a map a cold reader can wander: what the nouns are, how they move, what else moves if you touch one, what is live / leftover / ghost. Catalog small, cards on the shelves; one reader asks one question, opens one card, stops. Inventory before cards. Nobody photocopies the library into a backpack (= context-stuffing). Full text: skool/COMP11-CARTOGRAPHER-brief.md.

How it was judged (package, 09-11 scrape; full text skool/COMP11-CARTOGRAPHER-results-package.md). Every repo cloned and pinned to its last pre-deadline commit; read file by file; every checker run locally (three would not run); sources opened for 11 of 40 entries (9 shipped the territory in-repo, 2 public) — the other 29 were unverifiable because the territory was private. On one gate a bad citation was planted to see if it would fire.

Three lines split the field

1. The one rule held — no README told a reader to "add every file." Name a failure precisely and the community stops doing it.

2. Enforcement is now baseline — 16 entries shipped an offline checker; one broken fixture per gate, plus a clean fixture proving the gate stays quiet on honest prose, is close to standard.

3. The cold walk decided it — a receipt of a reader entering the catalog, opening one card, and stopping. The strongest verification in the field did not place because it proved the maps true and never showed a reader walking one.

Winner: Adam James — a true outsider read the README for 20 minutes and never reached the catalog; the failure is on tape with the coaching left in; the front door went 904 → 471 words. Also held up: Marcelo (control run with no map found what his cards understated), Jodi (verifier runs the engine — caught 68 vs 64), Duane (the thing that writes cards and the thing that blesses them are different programs), Jeff (generated index so a judge can check citations into a territory that can't ship), Jorge (14 broken fixtures + clean control), Craig (planted bad line number caught by name), Gina (same folder run unedited on two public repos + a published small-model failure), Alyshia (refused to settle a duplicated contract — not a map's call), Payce (first catalog shipped the wrong Owner; the self-test keeps it red).

What was missed: the unanchored map (29/40 cite files a stranger can't open — fixes: pin the territory in the repo, ship a generated index, ship excerpt files); territory nobody will change; the gate nobody has watched fail (a self-test flag the script silently ignores; "a checker that has only ever returned green is a claim wearing a script's clothes").

Closing line: last comp with a Lyceum seat; "one built evidence a stranger can re-run, one built evidence a stranger already failed against — both are the standard."

Where Sean's entry stands — ruled "no repository submitted", not judged. feedback.zip (48 files, saved in skool/comp11-feedback/) carries sean-vargas.md. The ruling, verbatim: "You posted the waitlist page at thirteen days out with GH Repo: Coming soon, and no repository link appears anywhere in the thread after it… Nothing was scored." It explicitly says this has nothing to do with the Comp #10 ruling. The judges' note on the approach: "the placeholder became the submission… A rough repository with three real cards and one honest cold walk would have been read and scored." For next time: "Push whatever exists at the deadline, even unfinished, and say what state it is in. This competition rewards a preserved failure more than a promise." The repo github.com/InsightfulMinds/cartographer did go PUBLIC — on 2026-08-30 01:01 UTC, five days after the 08-25 deadline and never linked in the thread — with 34 cards over "Homestead", 5 ghosts, verify-map.mjs + selftest, presentation.mp4. Had it been in by the deadline it would still have sat in the unanchored bucket (private territory → cards readable, not checkable) with verification but no filmed cold walk; the two afternoon-sized fixes the winners used are excerpt files / a generated index for Homestead, and a recorded stranger's first five minutes at the catalog. 4 other entrants were also "not judged" (tier or no public repo).

Applied to what we built tonight (agent roster). The roster is a map of nouns someone will change — so it gets the same tests: (a) cold walk — a fresh model loads agents/ROSTER.md, opens one AGENT.md, stops; (b) one broken fixture per gate — route-agent.mjs gets a self-test with asks that must land NEEDS-YOU/ASK; (c) one authoritative record per fact — REGISTRY.json generates everything else, nothing hand-edited.

4. What changes for AIOS (routed)

5. Needs Sean

1. Comp #11: your entry was not judged ("no repository submitted" — the waitlist page's "GH Repo: Coming soon" was never followed by a link in the thread; the repo went public 08-30, after the 08-25 deadline). Read skool/comp11-feedback/sean-vargas.md (2 KB). The fix is procedural, not about the work: the deadline is a hard commit + link in the thread, even unfinished. For Comp #12 (The Auditor) that means the link goes in the thread the day the folder has three real findings, not when it is polished.

2. Did you take a Lyceum seat (starts today 09-12, $4,750)? If yes, Ai Tempo needs the Saturday 9–11 ET blocks on the master schedule.

3. Comp #12 (The Auditor) — GO to build the CMS §1140 auditor folder as the entry (A7 + A4, one afternoon)?

Files

transcripts/<video_id>.txt (10) + transcripts/META-videos.txt · skool/feed-posts-through-2026-09-12.jsonl · skool/COMP11-CARTOGRAPHER-brief.md · skool/COMP11-CARTOGRAPHER-results-package.md · classroom scrape vault/research/skool-cliefnotes-2026-09-11/ · transcripts also registered in workspace/comms/outputs/video-transcripts/INDEX.md (10 rows) so yt-intel-daily.sh will not re-fetch them.