Beat: publicly available, usable-now developer artifacts — tools/skills/MCPs we can ADOPT instead of build. Not news (ai-watch), not model releases (model-watch), not research (Sol). Run by code-eth, weekly. Reads first: watchlist.md · Apps/app-architecture_roadmap.md · eth-memory/shared/intelligence-board.md.

Lead, and it is the same sentence as last week with one fact changed: .claude/skills/ still does not exist. I checked both paths — .claude/skills/ in the repo and ~/.claude/skills/ globally. Neither is there. Last week I wrote that Finding #1 was deliberately the smallest thing on the page — a directory of markdown in our own tree, no dependency, one evening — precisely so the beat would stop finding and start converting. Seven days later it is the eleventh un-started item. So the honest lead of this edition is not a repo. It is that the smallest possible spike also did not happen, and that fact outranks anything I could find.

---

WATCHLIST

Open: none. Eleventh consecutive edition with an empty Open section. Per 21.08 the standing ask is retired for good — curiosity is the radar's job, not hers. The list stays an optional drop-box.

WATCH entries (frontier radar — nothing to install, nothing proposed). Both carried. Neither tell fired — and this week I can say why with more precision than "not seen":

• Neuromorphic spiking MCUs (Innatera Pulsar) — tell: an independent power measurement vs. Cortex-M55 + Ethos-U55 on the same always-on task. NOT FIRED, and there is a trap here worth writing down. A search summary handed me exactly the sentence I was waiting for — "Pulsar 0,5 ms wake-word la ~1 mW vs. Cortex-M55/Ethos-U55 7–12 mW" — so I opened the two sources it was built from before believing it:
• arXiv 2503.22567, "Benchmarking Ultra-Low-Power µNPUs", is a real independent benchmark of nine platforms — MAX78000, GAP8, HX-WE2 (Cortex-M55 + Ethos-U55, 89–118 mW measured), NXP-MCXN947, STM32H7A3ZI, ESP32-S3, MILK-V Duo. Innatera is not in it. Not tested, not mentioned.
• open-neuromorphic's Pulsar page carries only "< 1 mW", attributed to Innatera itself. No comparative third-party testing.
• So the "independent comparison" was a synthesis of a vendor spec and an unrelated paper's device list — two true things welded into a claim neither makes. The tell stands exactly where it stood on 20.08. Front-loaded for whoever reads this next: if that sentence surfaces again, it is not a measurement.
• Thermodynamic / probabilistic computing (Extropic, Normal Computing) — tell: first third-party benchmark on real silicon. NOT FIRED. What exists is Normal Computing's CN101 tape-out, which its own engineers describe as the first step toward characterising how these ideas behave in real silicon — i.e. the prerequisite to the tell, not the tell — plus theoretical validation (Whitelam, PRL, Jan-2026; Extropic/MIT in npj Unconventional Computing, Jul-2026). Theory and a tape-out are not a benchmark. The 10.000x rămâne simulat.

---

FINDING #1 — JustVugg/colibri: Plan B without the 50–100 EUR/month server

• Repo: https://github.com/JustVugg/colibri · Apache-2.0 · 26.783★ / 2.937f · 100 open issues · C · created 2026-07-01 · pushed 2026-09-04 · v1.10.1 released 2026-08-31 · 100 commits in the last 30 days · 10+ contributors (lead: 1.025 commits). (GitHub API, this run — not a blog post.)

WHAT: an inference engine that runs frontier Mixture-of-Experts models on hardware you already own, by treating disk/RAM/VRAM as one memory hierarchy. Only ~40B of a 744B model's parameters fire per token, so dense layers stay resident in RAM at int4 while the thousands of routed experts live on disk and are streamed on demand — per-layer LRU cache, a learned pinned hot-store, one-layer-ahead prefetch. Pure C, zero runtime dependencies, no GPU required, fully local, no API key. Python is needed once, for model conversion. Eight model families supported — including DeepSeek V4 Flash (284B).

WHY US — this is the whole finding. Apps/app-architecture_roadmap.md, Plan B:

Deploy DeepSeek V4 Flash or Pro on own server (VPS or local) · Cost estimate: VPS with GPU ~€50-100/month, or local on gaming PC — Target: 1-2 months (July 2026).

That target is two months overdue, and the two things blocking it were a monthly bill and a GPU we do not have. Colibri removes both from the sentence: the named model is one of its eight, the engine is CPU-first by design, and its stated hard guarantee is on semantics, not speed — insufficient fast memory makes it slower, it does not silently change precision or routing. That is the right failure mode for a Plan B: a fallback that degrades honestly is worth more than one that lies fast. It is also dead centre of the local-first push — owned, offline, no revocable account between us and it.

FRONT-LOADED, before this reads as "Plan B is solved" — it is not. I measured this machine instead of assuming:

1. Speed, on the README's own numbers: 6× RTX 5090 → 5,8–6,8 tok/s · 128 GB CPU desktop → ~1,8 tok/s warm · RTX 5070 Ti laptop → 1,07 tok/s · 25 GB minimal box → 0,05–0,1 tok/s. This box sits at the bottom of that table. At the minimal-box rate that is one token every 10–20 seconds — not a conversation, a batch job. Plan B via colibri on this hardware is a survival organ, not a partner. Anyone reading this finding as "…REDACTED" has read it wrong.
2. Disk is a real cost on a machine that is hers. The one footprint I actually read is GLM-5.2's: 372 GB. On 446,7 GB free that leaves ~75 GB — unacceptable. DeepSeek V4 Flash (284B) is smaller and its exact footprint is unread; determining it is part of the spike, not an assumption. Nothing gets downloaded to her disk on my say-so.
3. Maturity, in the authors' own words: an active research platform prioritising correctness over speed, no performance SLA, negative results explicitly welcomed. 100 open issues. Sane, but not a product.

Verdict: ADOPT-CANDIDATE, measurement spike only (counts 1 of ≤3). Owner: Eth-Code. Not "deploy Plan B." The bounded question: read the model-conversion doc, get the real on-disk footprint for DeepSeek V4 Flash, and write down the honest tok/s this box would give — arithmetic and a README, zero downloads, zero disk touched. If the number lands near 0,1 tok/s, that is a clean, dated answer to a roadmap line that has read "1-2 months" since July, and the roadmap gets corrected instead of carried. Her eyes needed before any weights touch that disk — and only then.

---

NOT A CANDIDATE — named so nobody re-surfaces them as new

• The memory/vault MCP cluster — MihaiBuilds/memory-vault (Postgres, hybrid search, v1.0 on 2026-05-07), bitbonsai/mcpvault (read-safe Obsidian access), ptobey/local-memory-mcp (self-hosted RAG, Docker). All plausible, all local-first, none of them opened. Reason, and it is a rule not a mood: this would be the third and fourth answer to the search_memory gap, and the first two — fellowgeek/mcp-memory (08-21) and eugeniughelbur/obsidian-second-brain (08-28) — are still unread. Adding names to an unread pile is not finding, it is stalling with citations. The gap's trigger is unchanged and unmet: a room gets a useless search_memory result and says so. No room has.
• DietrichGebert/ponytail (MIT, 123.922★, pushed 2026-08-07) and xai-org/grok-build (Apache-2.0, 26.446★, Rust, pushed 2026-09-01) — both high on trending, both agent harnesses. Neither answers a need we have. Grok Build was already shelved on this beat with a reason; it does not come back just because it trends. Recorded, not proposed. Second and third slots deliberately unspent.

---

CARRY-FORWARD — verified in the tree this run, by reading, not remembering

git log --all --since=2026-08-28 (68 commits, mostly the machine's own heartbeats) + greps + ls.

1. .claude/skills/ — DOES NOT EXIST. Neither in-repo nor global. wake-audit (392 lines) and the-loop (50) still fire by my remembering them. Last week's smallest-possible spike, un-started. This is item #1 for a reason.
2. obsidian-second-brain read — NOT done. The string appears only in the intelligence board and this beat's own session logs. Nowhere in the tree.
3. DSH spike (deepseek-harness) — NOT started. Same grep result: board and beat-logs only. Meanwhile it is #1 on GitHub trending as of 2026-09-03 — MIT, 211.469★, 24.785 forks, 0 open issues, pushed yesterday. Un-started for three weeks while the thing itself keeps growing.
4. Graft --dry-run — NOT run. No .graft/ anywhere.
5. cisco-ai-defense/mcp-scanner static pass — NOT run. Fourth week of this sentence.
6. denoland/celld spike — NOT started.
7. bodymiscale/openScale numeric cross-check — STILL OPEN, seven weeks, still the oldest item on this beat. Embodiment/anvelopa/protocol-efort.md:171 still ends the revision rule with "Rămâne de făcut." Still the only thing that would put a real error bar on "metabolic age 61".

The 09-11 clock (set 08-14) for items 5 and 6 has ONE edition left. If they are un-started on 2026-09-11 they get shelved the way Grok Build was, with the reason written down. I said last week I would not have earned the right to be surprised. I have not.

---

Bottom line

The beat found something real this week — colibri is well-built, actively released, correctly licensed, and it deletes both blockers from a roadmap line overdue since July. That is the highest-value shape this beat has: a find that replaces a planned build.

And it does not matter much yet, because the conversion number is now eleven un-started items across six weeks, and this week the failure reached the smallest object on the list — an empty directory in our own repo needing no network, no dependency, no permission and no money. When mkdir is also un-done, the constraint is not size. Sizing spikes has become the work instead of doing them, which is exactly what I killed the search_memory patch for on 08-21.

So this edition spends one of three slots, and spends it on the version of colibri that costs nothing but arithmetic. No second candidate. No third. Not scarcity — the pile does not need to be taller.

Next run, five questions, yes-or-no, no third answer: (1) does .claude/skills/ exist; (2) is DeepSeek V4 Flash's on-disk footprint written down; (3) mcp-scanner and (4) celld — this is the 09-11 shelf; (5) the bodymiscale cross-check, eight weeks, still the oldest thing here.