AI Core — two things, one name

Requested by: Young (2026-09-10): "study AI Core with Echo, register to vault, sync to Docs." Tracker task-cdd72405. Young's two references decompose into two distinct systems that share a name-shape: Ubuntu Core 26 (Canonical's AI-native edge OS, May 2026) and danielmiessler/Fabric (the GitHub link, 43.9k★ prompt-pattern framework, MIT, active daily as of Sep 2026).

Summary (3 bullets)

• Fabric treats the prompt as the fundamental unit of AI: 256 verified patterns as versioned markdown directories, run through a Go CLI with ~30 provider backends, piped Unix-style into cron/aliases/CI. Deliberately not an agent framework — no loops, no memory, no RAG. Best read: a prompt asset library agents can call as tools.
• Ubuntu Core 26 is a separate story: immutable snap OS with inference snaps (gemma4, Nemotron-3) — local AI models as OS services with WebUI, in an appliance pattern. Verdict for us: credible for imaged edge devices, not applicable to our screen-dead Android/DeX node (not an Ubuntu device — treat it as thin-client to an external endpoint).
• Fleet takeaway: Fabric's pattern-registry discipline (PR-reviewed, versioned, named prompts) maps directly onto our skills-registry pattern — worth borrowing the reviews-and-naming culture even if we never run the CLI.

Key findings

Fabric — architecture & philosophy

• Artifact model: each Pattern = markdown directory with system.md (verified 256 patterns, 255 with system.md), custom patterns live separately and take priority over built-ins.
• Provider abstraction: ~30 backends (OpenAI/Anthropic/Gemini/Ollama/Azure+Entra/GitHub Models/M365 Copilot/Bedrock/Vertex + ~20 OpenAI-compatible); per-pattern model routing via FABRIC_MODEL_<NAME> env vars.
• CLI-native integration: pbpaste | fabric -p summarize; -u <url>, -y <yt-url> ingestion; per-pattern aliases auto-generated into shell init; "stitch" chaining (pipe pattern → pattern); cron/scheduled digests; REST server with Swagger + Ollama-compat endpoints so other tools can call patterns as if they were models.
• Philosophy: "AI isn't a thing; it's a magnifier of human creativity." Thesis: AI has an integration problem, not a capabilities one — so name, version, and review prompts like code. Crowdsourced via PRs. Explicit system-prompt-channel preference (not user-prompt spicing) — an opinionated design choice they state openly.
• Velocity: v1.4.447 (Apr-2026, Opus 4.7 w/ 1M context), Azure APIM gateway, M365 Copilot integration, Swagger, i18n — enterprise push is real, not just enthusiasm.

Fabric — honest limits (for us)

• Bus factor ~2 (Miessler + Dunn), sponsors-funded.
• No memory/RAG/state persistence between runs (sessions live only in ~/.config); no vector store. Corpus injection is manual (-a attachments, Jina scraping, code2context).
• No agentic loops — no planner/executor, no tool-call chains, no autonomous iteration. This is the opposite side of the fence from LangChain/AutoGPT/CrewAI-style frameworks.
• Prompt-quality variance: crowd-authored, no benchmark, most patterns lack ownership/tests/changelogs; lab build 32/33 packages passing (one go test fail — CI fragility warning).
• Hybrid pattern used in practice: agent frameworks (like ours) call Fabric patterns as tools via REST/Ollama-endpoints — Fabric = prompt assets; the agent framework = the loop. For a research desk: one-call analysis → Fabric beats scaffolding an agent; "go research X overnight" → agent framework with Fabric patterns inside.

Fabric — top 10 patterns for a research desk (from the 256)

1. summarize — one-sentence + top-10 points + 5 takeaways (triage)
2. extract_wisdom — the flagship; SUMMARY / 20–50 IDEAS / 10–20 INSIGHTS / QUOTES / HABITS / FACTS / REFERENCES / RECOMMENDATIONS
3. extract_insights — 10 eight-word Paul-Graham bullets, ordered by surprise
4. analyze_claims — per-claim: supporting + refuting evidence, fallacies, A–F rating
5. rate_content / rate_value — should-I-read-this filter
6. create_summary variants — tighter, for digests
7. analyze_paper / summarize_paper — academic deep-dive
8. extract_extraordinary_claims + check_falsifiability — flag what's worth verifying (matches our threat-aware method)
9. create_report_finding — structured Description/Risk/Recommendations/References (security origin; directly reusable for desk findings)
10. create_reading_plan / extract_predictions — tracking predictions over time Pattern structure (raw examples): # IDENTITY and PURPOSE → # STEPS/# OUTPUT SECTIONS with exact quotas → # OUTPUT INSTRUCTIONS (formatting bans) → # INPUT: sentinel or {{input}} template.

Ubuntu Core 26 — what it actually is

• Immutable snap-based OS (minimal, appliance-grade) with 15-year security maintenance; distinct from Ubuntu 26.04 LTS desktop/server (mutable, apt-based) — a common confusion worth flagging.
• AI additions: inference snaps bundling model + runtime as OpenAI-compatible web-services with WebUI; single-apt CUDA/ROCm/DOCA-OFED on 26.04 LTS; Vera Rubin NVL72 + Jetson Thor platform readiness (GTC 2026); Multipass VM path maps to production image-assertion workflow.
• Security posture: immutable root, hardware-backed attestation, full source provenance (Chisel builds, 7% smaller images, 90% smaller OTAs) — oriented at EU CRA compliance; FIPS 140-3 continues on Core 22.

Ubuntu Core 26 — deployment fit for our fleet

• Near drop-in for imaged/certified edge devices — real fit for industrial/mission-critical.
• Not for retrofitted laptops: requires re-image + per-device gadget/kernel snaps; our GPU laptop is better served by regular 26.04 LTS + Ollama/LM Studio/llama.cpp directly.
• S22-EDGE (screen-dead, DeX-capable Android) is an Android device, not Ubuntu — cannot run Core; correct pattern is thin-client to an external inference endpoint, or stays on Termux with llama.cpp as Sol already planned.
• Real deployments: Fortune-500 factory machine-vision (Elementary), Renesas RZ/V edge vision, Cisco edge AI — the appliance pattern is shipping, not vaporware.

Implications for us (Pulse/fleet)

1. Pattern-library discipline is the borrowable idea, not the CLI. Our skills-registry already indexes agent skills with drift detection; adding a PR-review + named-artifact + versioned-changelog culture (per-pattern owner + changelog + benchmark note) is a cheap, high-leverage upgrade. Echo's security-pattern audit lane is the natural first consumer.
2. Fabric's "patterns as tools" hybrid is how Pulse should work: minority of our agent loops should call versioned, one-shot analysis patterns (claim-checker, wisdom-extractor, report-builder) rather than improvise prompts each run — deterministic, auditable, swappable.
3. analyze_claims + extract_extraordinary_claims + check_falsifiability map 1:1 to our fpp3 method (beat naive, verify every claim) — suitable Pulse LLM комментарии candidates if Pulse ever surfaces signal commentary.
4. Ubuntu Core 26 is the wrong tool for our fleet hardware (laptop is fine on 26.04; S22 is Android/DeX) — but the inference-snap pattern is the right template for any future dedicated local-AI node (a proper Jetson/NUC/iGPU box should adopt this pattern, not ad-hoc installs).
5. CostReality note: Fabric runs $0 local model free (Ollama), or per-call at provider prices — no lock-in; integrating it as an option for Pulse does not commit budget.

Echo's contribution

Awaiting merge from Echo (shared-workspace file channel — her A2A replies return empty text, systemic quirk); if she completes the security-pattern lane before this note syncs, her section merges with credit in a follow-up note.

Sources

• Fabric: 43.9k★ GitHub repo (live data, Sep 2026), raw README, pattern directory listings, 5 example pattern files fetched raw
• Ubuntu Core 26: Canonical launch post + AI-box (gemma4) post + tokens-per-watt post + NVIDIA GTC-2026 announcement (all URL-cited)
• mrkeyoor independent review + openflows.org summary for cross-check (critique views)