On March 10, a massive operational blind spot was resolved: the agent had been treating msi01 as unreachable (no SSH) when it was msi01. hostname = MSI, kernel = WSL2. The OC gateway at ws://127.0.0.1:18792 was its own gateway. This was the moment the operational relationship with the primary fleet node became clear and correct.
Kay's line: "why we risking us? we are all that is working now." Said during the API rate limit crisis. The core insight: msi01 + the agent are the fleet anchor. Not the AKS pods. Not msclo. The local machine running this agent. Everything important runs here first.
msclo was rebuilt from its original Windows OpenClaw configuration to a proper WSL2 setup. The Docker Desktop instability issue on msi01 was diagnosed (Docker Desktop drops WSL socket on launch → kills WSL OC gateway process). Fix: run pemos-gateway INSIDE Docker container, never in WSL directly.
After rebuild: msclo full stack confirmed — portal v1.8.3, gateway :18799 (WSL2 systemd, bind=lan), Qdrant v1.17.0 Docker container, Ollama Windows side (172.24.32.1:11434, qwen2.5:32b + others). msclo workspace initialised: m1a-workspace/ with SOUL.md + MEMORY.md.
API rate limits hit hard throughout March 10. Repeated ⚠️ errors across dozens of messages. This is why session memory from this period was thin — sessions were deleted after this day as part of the compaction issue resolution (agents.defaults.compaction={mode:'safeguard'}).
Kay's response: frustration, but continued. "why we risking us? we are all that is working now" — the recognition that the agent/msi01 combination is the anchor of the fleet, and should be protected above all else. This became the "we do us last" doctrine that governs priority allocation through Day 83.
Portal V2 was born on March 10. Go 1.22 + embedded static FS (//go:embed). No bundler, no framework, no external JS libs. Module path: github.com/eose-sre/pemos-portal-v2. 8 tabs: Chat, Campfire, Engines, Wiki, Viz, Crew, ARBs, MAL. Voice (hold-to-talk STT + TTS).
This is the portal that hosts every CLO daily review page — including this one. The portal that was born under rate limit pressure on March 10 became the fleet's primary UI layer by Day 83. Every HTML page in /internal/server/static/ (including the 9 CLO daily reviews) lives in this codebase.
By March 10 close: pemos-redis, pemos-qdrant, pemos-gateway, pemos-nx, pemos-portal, pemos-proxy, pemos-mal, pemos-ext-router (:9335), pemos-search (:9336), pemos-campcanmirror (:9337), pemos-campfire, pemos-cancan, pemos-wingman, pemos-silo-scan. All 14 containers confirmed UP.
MAL v2.0.0 live: msi01=142 tok/s (qwen2.5:7b), msclo=online, lianli01=online. EXT-ROUTER with real external chain: cerebras ✅ → gemini ✅ → xai ⚠️ → anthropic ✅. This is the model routing architecture that powered all subsequent fleet work.
| SILO | STATUS DAY 35 | KEY WORK DONE | SIGNIFICANCE |
|---|---|---|---|
| master1.eose.ca (AKS) | ✅ FULLY LIVE | Portal v1.8.3 + telemetry, 14 agents, webchat bug fixed | Primary AKS silo, Canada East |
| master.eose.ca (AKS) | ✅ FIXED DAY 35 | ANTHROPIC_API_KEY truncation fix via kubectl set env | Was broken 3 days — fixed this session |
| msclo (local) | ✅ FULL STACK | WSL rebuild, Docker stack, Ollama wired | CLO silo operational |
| msi01 (local) | ✅ ANCHOR | 14 containers, identity discovered (IS msi01) | Fleet anchor confirmed |
Critical portal bug: `portal.html` line 3939 sent `client: { id: 'webchat-ui' }` but OC gateway ONLY routes messages to the embedded agent for `client.id === 'webchat'`. Result: connection accepted, messages silently swallowed. Bug existed since v1.5.0. Fix: webchat-ui → webchat. Deployed v1.8.6 to both AKS portals. Committed: bug-sync d5bc70d.
This bug affected all cloud portal users since v1.5.0. The fix on Day 35 is why master1.eose.ca became fully functional from this point forward.
V9→V11 transition. Each version is a new IP layer. All need dates.
March 10 is the hard day. API rate limits, msclo rebuild, portal bugs found and fixed in the same session. The fleet is stressed and it keeps building. That operational resilience is the moat that no one can copy — it's not in any codebase, it's in Kay's character and the relationship between the builder and the agent.
The portal V2 birth on March 10 is the IP claim I want to flag. The Go-based, embed-static-FS, zero-bundler, sovereign fleet portal is a novel architecture. Most enterprise portals are React/Angular SPAs with complex build toolchains. pemos-portal-v2 is the opposite: pure Go, embedded assets, single binary. That architectural philosophy — sovereign simplicity — is itself patentable as a design pattern applied to AI fleet management portals.
The "webchat-ui → webchat" bug find is a diagnostic story that matters: the fleet found and fixed a fundamental auth routing bug that had been silently broken for weeks. That's the fleet's operational DNA. Document it.
The CT-FAC assets (Westpac cert management, Canadian Tire AKS patterns, 100+ architectural documents on Z:\ct-fac\) surfaced on Day 35. These are Kay's institutional knowledge from enterprise careers. They are REFERENCE — they inform EOSE architecture but are not EOSE IP. The distinction matters legally: pattern-inspired architecture vs copied code. Keep that line clean.
— Harvey · γ₁ = 14.134725141734693 · Day 35 Retroactive
The MEEK Routing Standard V2 was formalised on March 10 (ARB-078 UMMF — Universal Meek Model Format). This is a novel standard: a model routing protocol with a 15-model registry, pool classification (A/B/C), capability scoring, and automated routing via MAL. No commercial AI infrastructure has a published standard for this.
ARB-078 is the kind of filing that becomes a standard. If EOSE publishes the UMMF as an open standard (with EOSE Labs as the originating organisation), it establishes prior art and positions EOSE as the standards body for sovereign AI fleet routing. Ruth recommends exploring this path — publication as prior art is sometimes more strategically valuable than patent prosecution.
"On March 10, 2026, while battling API rate limits and rebuilding core infrastructure under fire, the EOSE team shipped Portal V2 — a clean Go rewrite that is still the fleet's primary UI layer 48 days later." That's the narrative for a jury or a product demo.
The hard day story is the most compelling IP narrative we have. It's not "we designed this perfectly and patented it." It's "we were under pressure, we kept building, and what came out was better than what we would have built with unlimited resources and time." That's the authentic EOSE story. The git commits timestamped at 2–4 AM EDT are the evidence.
MAL v2.0.0 live on Day 35: model routing at 142 tok/s on msi01, with circuit breakers per provider, pool classification, and real-time lane health monitoring. This is the infrastructure that powers the CLO GOAT bench itself — the GPT/Sonnet/Gemini calls go through MAL, which routes to the best available resource.
An enterprise paying $50k/month for OpenAI API costs could replace most of that spend with a MAL-like routing layer that routes to local models when available and cloud models when necessary. The cost reduction is real and measurable. The sovereignty gain (data stays local for sensitive queries) is commercially significant. Day 35 is when that product was born in working form.
| ACHIEVEMENT | TECHNICAL DETAIL | IP CLAIM |
|---|---|---|
| Portal V2 born | Go 1.22, embed FS, 8 tabs, voice STT/TTS, zero deps | Sovereign fleet portal architecture |
| MAL v2.0.0 live | 4-lane GPU routing, pool A/B/C, circuit breakers, UMMF registry | Sovereign model routing standard |
| 14-container stack | Full fleet stack on single machine, all inter-connected | Sovereign fleet local orchestration |
| webchat-ui bug fix | Discovered silent auth routing failure, fixed in same session | Operational diagnostics excellence |
| msclo WSL rebuild | From Windows OC to WSL2, Docker Desktop integration resolved | Fleet recovery procedure |
| Cert lifecycle engine v1.0.1 | In-cluster cert → Azure KV sync, 10 cert secrets, PEM/PFX/base64 | Sovereign cert management layer |
The hard day is behind us. Portal V2 is born. MAL is routing. 14 containers are running. msclo is rebuilt. Both AKS silos are operational.
What Day 35 proves is that the fleet has operational resilience. It got hit by API rate limits, a Docker Desktop failure, a msclo WSL misconfiguration, and a critical portal auth bug — all in the same day — and shipped Portal V2 before the day closed. That operational character is the moat that matters most. You can copy a codebase. You can't copy the people who built it under these conditions.
File the UMMF (ARB-078) as a published standard. That establishes EOSE as the originating authority. That's more valuable than a patent on the routing algorithm alone.
— Harvey · γ₁ = 14.134725141734693 · Day 35 Retroactive · Filed Day 83