Published July 10, 2026
Last post was about the stabilization weeks — the phase where the building slows down and the sharpening begins. There’s a trap hidden in that phase: it’s always the same pair of eyes doing the sharpening. Mine, and the model I build with. We’ve both looked at every part of this system so many times that we’ve stopped seeing some of it.
So we brought in a different pair of eyes. We handed EIDARA end to end to Fable — a new Claude model from Anthropic that hadn’t lived through a single one of this year’s small decisions. The ask was deliberately simple: look at the whole thing and tell us what would make it better. Structural, not cosmetic. Ranked by impact. And don’t be charitable.
It came back with a long list. This post is about the ones we’ve implemented.
We gave it about 24KB of prompt: the architecture, the constitution, the enablers, the compile pipeline, the maintenance stages, the way any AI reads and writes through MCP, the vault layout, the queues. Enough to reason about the system as a whole instead of one file at a time.
What came back wasn’t a critique. It was a ranked list, each item with a diagnosis and a concrete proposal. Seven of them turned into real changes. Here they are, heaviest first.
The vault is now split into four zones — personal, business, shared, and system. Every neuron belongs to exactly one. Every session is scoped to a set of zones, and the compiled brain each assistant reads respects the same scope.
In practice: a work session helping me with a client now sees only the business and shared zones. A personal session sees personal, business and shared. The system zone stays out of reach unless I ask for it explicitly. Today that’s 325 business neurons, 187 system, 8 personal, 5 shared — sorted by a three-layer classifier that leans on keyword rules first, then similarity to each zone’s prototype, and only falls back to the local model for whatever is left. More than 90% of it gets sorted without sending the whole vault through a model at all. A weekly pass keeps the keyword layer alive as the language drifts.
This is my favorite of the seven. Until now, when a frontier model contributed a fact, it wrote the same way the small local curator does — raw. Fable proposed making the division of labor explicit: five rules that get prepended to the context every session opens with.
The write itself also grew three optional hints: which neurons a fact probably belongs to, which zone it probably belongs to, and a flag to refuse the fact outright if it looks like noise. The frontier model does the semantic heavy lifting upstream; the local model verifies and integrates downstream. Better inputs, less work for the small model.
The senses are the exception. The listener, the keyboard reader and the screen reader capture raw audio, typing and on-screen text — they can’t reasonably propose where an unfiltered stream belongs. So anything tagged as a sensor turns the noise-refusal off and lets the fact through. Senses stay raw; the curator sorts them out.
It’s the first time the two models aren’t competing for the same job. They’re doing different ones.
Here’s a concrete one, because it almost got past us. The compiled brain — the file the assistants actually read — was rendering only 27 of the vault’s 531 neurons. The cause: eleven separate modules were reading neurons with a flat file lookup instead of a recursive one, and the neurons had since moved into per-zone subfolders. Every module passed its own tests. The system as a whole was showing the assistant under 5% of what it thought it was showing.
Fable caught it by cross-referencing counts between modules.
The real upgrade isn’t the fix — it’s the defense. We repaired the eleven read paths, and then added a check to the daily audit: if the compiled brain ever renders less than 80% of the active corpus, the daily report goes red on its own. That check has since caught two smaller regressions I introduced afterward. Both would have slipped past.
The numbers inside the system’s own summary notes — how many neurons, how many facts, the latest build — used to be maintained by hand, which meant they were quietly always a little wrong. Now they sit between two markers that get rewritten from the filesystem and git on every compile. The prose around them stays exactly as a human wrote it; only the numbers refresh. Same idea as the drift defense, one size smaller: don’t let a summary go stale over a number nobody remembered to update.
A new maintenance pass runs every 28 days. It scans three buckets — names that look like noise, duplicate lines in the compiled brain, and orphan neurons with no cross-references older than a month — and if any bucket crosses its threshold, it raises a task to the architect’s queue. It doesn’t act on its own. It just raises a hand.
We also ran the same logic backwards once, to clear what had piled up before the monthly pass existed. It archived 68 neurons: 55 noise-pattern names and 13 placeholders that said “consolidated into X” and never got cleaned up. The vault went from 646 to 578 neurons in an afternoon — the shape it should have had all along. Now it keeps that shape without me remembering to.
The rule that spins up a new “home” neuron when a fact has nowhere to go used to fire at three matching facts. Fable argued that was too eager, and it was right — it’s now eight. There’s also a new filter that catches subjects that don’t look like real subjects — session logs, audit timestamps, stray fragments — and sends them to review instead of letting them become a home. The vault grows slower and with more purpose now. When a new neuron appears, it’s because eight distinct facts converged on something worth naming, not because three fragments happened to share a word.
Forward planning used to live in two places: a summary neuron and a separate planning file. Two sources of truth is one too many. The neuron now opens with an explicit line ceding the role: the living plan is the planning file; this note is history. One file, one direction of travel.
The interesting thing wasn’t that Fable was smarter. It was that it had no working assumptions to defend. It hadn’t lived through the year of small decisions that produced the current shape of the system, so it asked blunter questions — and it cross-referenced counts across parts of the code that I had only ever looked at one at a time.
The split-brain is the clearest example. I had read each of the eleven modules that were reading the vault wrong, many times, over months. So had the model I build with. Neither of us saw it, because seeing it meant noticing a mismatch across the eleven, not inside any one of them. Fable did.
That’s a role worth structuring into any project that has been running a while. Not a replacement for the model you build with — a different participant, invited on purpose, asked to be uncharitable, and given enough context to cross-reference.
None of these upgrades are new capabilities. Each one is a small structural refusal: to let one assistant see everything, to let a big model write like a small one, to let the brain quietly render a fraction of itself, to let a summary go stale, to let cleanup pile up, to let the vault grow on a shared word, to let direction live in two places at once.
We want to keep this pattern going: external audit passes on a regular cadence, from a model that didn’t build the thing. Not because the system fails when we skip them. Because our own tolerance for its weirdness quietly grows when we skip them. We are still working on this. Every week it’s a little closer to the thing I want it to be.
— Javier
EIDARA v2 is free. SUPER DARA is what comes next.