dendrite.research.md is six thousand words long and it is magnificent. It proposes rebuilding MKUltra — my agentic OS — on the computational principles of biological dendrites. Reduced compartmental models via Dendrify. Calcium action potentials as logic gates. A holographic memory tier storing knowledge in interference patterns and retrieving it by Fourier transform. Organic electrochemical neurons at 40 picojoules a spike. FPGA-accelerated dendritic logic. Backpropagating credit assignment through an orchestrator shaped like a dendritic arbor.

It has equations. It has comparison tables where my system wins. It describes the whole thing in the calm present tense of software that exists.

Almost none of it existed.

I have a rule now, earned the hard way and written up last week: the roadmap goes in the future tense. A document that describes your intentions as your capabilities isn’t a design doc, it’s a press release you’re issuing to yourself. So this post is the audit. What was real neuroscience, what was extractable, what actually shipped, and what is still — a year of enthusiasm later — prose.

The ratio turned out to be one for nine. I think that’s a good ratio. I want to explain why.

First, the part that isn’t hype, because the underlying biology is genuinely one of the best results in modern neuroscience.

The textbook neuron is a summing junction. Weighted inputs arrive, get added, and if the total clears a threshold the thing fires. Leaky integrate-and-fire. That abstraction is the foundation of essentially every artificial network ever built, and on the strength of it we proved a famous limitation: a single neuron cannot compute XOR. Exclusive-or isn’t linearly separable, so you need a hidden layer. That proof nearly killed the field in the seventies.

Then somebody patched into human Layer 2/3 pyramidal neurons and found dCaAPs — calcium-mediated dendritic action potentials. And they don’t behave like axonal spikes at all. Axonal spikes are all-or-none: cross the threshold, get a full spike. dCaAPs are graded and, crucially, biphasic — amplitude is maximal right at threshold and then decreases as the stimulus gets stronger.

Sit with that. Push harder, get less.

That single property is an XOR gate. One input: strong response. Both inputs: the drive exceeds the optimum and the response is suppressed. A single human neuron, in a single dendritic branch, computing the function we spent decades insisting required a whole network.

The measurements are unglamorous and specific in the way real findings are: dCaAP widths around 4.4 ± 1.4 ms, mean firing rates near 4.6 Hz. The biphasic profile is sensitive to sodium and potassium conductances — turn up gNa and the XOR behavior degrades; gK is what stabilizes it.

So: real, measured, load-bearing. The question was never whether the biology is good. It’s what survives the trip from a pipette in a human cortical slice to a Python service in my basement.

Reading a document like this, the temptation is to try to build all of it, get four levels deep into compartmental ODEs, and quietly abandon the branch six weeks later. I’ve done that. It’s how you end up with a README full of things that don’t exist.

So I went through it looking for a much narrower thing: ideas that map onto a gap I already have. Not ideas that are interesting. Ideas that are load-bearing for a problem currently costing me something.

Three survived:

• #1 — Multiplicative coincidence gating in PoonGram’s focus-point selection. Inputs landing close together on a branch should multiply, not add.
• #2 — The dCaAP-XOR kernel as a merge rule: agreement amplifies, conflict suppresses.
• #3 — Activation plateaus as a working buffer. Dendritic plateau potentials last ~100 ms, an order of magnitude longer than a spike, which makes them a natural jitter buffer for asynchronous inputs arriving at different speeds.

Everything else — the Fourier holography, the neuromorphic hardware, the arbor-shaped orchestrator — failed the test. Not because it’s wrong. Because it answers questions I don’t have yet.

Idea #2 got built, because it landed exactly on a hole I’d been staring at for weeks.

The hole: my memory bus fans a question out to four tiers — episodic, semantic RAG, and two biological lobes — and then merged the results by concatenation. Stack them up, dump them in the prompt, equal weight, no referee.

Which is indefensible when you say it out loud. Four systems with four different ways of being wrong, and “the emotional lobe is fixated on this” carried exactly as much authority as “two independent systems both surfaced this.”

The dCaAP kernel is the missing rule, and it’s three words: agree → amplify, conflict → suppress.

Pull the salient terms from each lobe. Anything two or more lobes independently surface is corroborated — promote it to a converged-signal banner at the top of the context. Any lobe agreeing with nobody is an outlier — flag it unverified and truncate it. In between, the context budget scales with the agreement score. That’s the biphasic curve, doing in a merge function what it does in a dendrite: fire on coincidence, go quiet on noise.

converged: [healthy, docker, services, running]
agreement: { assoc: 1.0, struct: 1.0, rag: 0.04, episodic: 0.15 }

Both biological lobes at 1.0 — total agreement, signal amplified. RAG at 0.04 and episodic at 0.15, contributing almost nothing to the consensus and correctly demoted.

Then the test that actually mattered: feed it a deliberate liar. One lobe spouting unrelated fantasy prose into an infrastructure question — which, if you read the lobotomy post, is a very specific failure mode I have. It scored 0.0, got stamped unverified, not corroborated by other lobes, and dropped to a footnote.

It ships as _merge_memory_lobes(), wired into both /query and /query/stream, surfacing a memory_merge object and an SSE event so you can watch the deliberation happen instead of taking my word for it. Unit test confirms outlier suppression.

Total: one function. Maybe an afternoon. The memory doesn’t just retrieve now — it decides who to believe.

Still ideas. Both listed in the build notes as candidates, neither implemented, and I’m putting them here rather than in a features list for exactly the reason this post exists.

#1, multiplicative coincidence gating in PoonGram. Right now focus-point selection sums activation. Biologically, inputs clustered on the same branch produce super-linear responses while distant inputs sum linearly. Same evidence, different answer depending on whether it arrived together. I believe this is right and I haven’t built it.

#3, the plateau buffer. Sub-agents and local models stream at wildly different tokens-per-second. A dendritic plateau — a sustained ~100 ms depolarization that outlives the spike that caused it — is precisely a resettable buffer for integrating inputs that don’t arrive together. My arbiter currently has no such thing, which means a slow lobe can miss its window and get treated as silence. That’s a real bug with a known fix that I have not applied.

Two good ideas, sitting in a file. That’s allowed. What’s not allowed is writing them into the architecture doc in the present tense and letting them count.

Now the honest column. Everything below is in dendrite.research.md, described as architecture, and does not exist in any form:

• Holographic associative memory. Storing knowledge in oscillatory interference patterns, retrieving by Fourier transform, with a “moving bubble of attention” over dendro-dendritic connections. Beautiful. Pribram-flavored. Not a line of code.
• Organic electrochemical neurons at 40 pJ/spike. This is real hardware research by other people. It is not in my basement.
• FPGA-accelerated dendritic logic. No.
• Dendritic Localized Learning replacing backprop. Genuinely interesting — local updates from pre/post activity and calcium signals, dodging the weight-transport problem. Not implemented.
• The arbor-shaped Conductor with backpropagating credit assignment. conductor-research.md is dated April and is, per its own header, research to inform design. There is no conductor.
• SOUL.md as a morphological skill arbor that grafts and prunes skills by structural plasticity. SOUL.md is a file. It does not grow branches.

And one specific number I want to correct, because it’s the kind of thing that propagates: the doc credits LiquidBrain with 18.4 million synapses in a comparison table against a competitor. That figure is PoonGram’s. LiquidBrain is a different brain with different numbers — 42.5 million neurons, 105.6 million synapses. Two lobes, two sets of stats, quietly merged into one impressive-sounding row pointed at a rival.

So was the document worth it, if 90% of it is still prose?

Yes, and the reason is narrow enough to be useful.

The biology paid for itself once, and it paid in the form of a rule I would not have arrived at otherwise. I knew concatenation was wrong. I did not have a principled alternative, and I’d have reached for the obvious ones — weight the tiers by hand, or ask a bigger model to arbitrate, which is how you turn a merge bug into a latency bug and a bill. “Agree amplifies, conflict suppresses, corroboration sets the budget” came straight out of a curve measured in human cortex, and it’s better than what I had.

That’s the return: one transferable rule. Not an architecture. Not a paradigm. A rule small enough to implement in an afternoon and specific enough to test.

The tax is what happens when you mistake the metaphor for a blueprint. Nothing in my system has a membrane, a calcium channel, or a millisecond that means anything. Writing an ODE for a compartment I’m not simulating would be cosplay — and cosplay is exactly what produces a comparison table claiming holographic retrieval against somebody else’s honest FAISS index.

The dendrite is worth studying because evolution solved a merge-arbitration problem under a brutal energy budget and left the answer lying around. It is not worth imitating, because I’m not energy-constrained, I don’t have ions, and my lobes talk over HTTP.

Take the rule. Leave the cell.