Talk-to-your-data was a UI, not an architecture
A chat box is an input method. It says nothing about what is on the other side of it, which is why two systems with the same interface can differ by an order of magnitude in what they are worth.
Ask what a chat box tells you about the system behind it and the answer is: nothing. It tells you how a question arrives. It says nothing about whether anything on the other side holds a model of your business, whether that model changes when a contract is renegotiated, or whether anybody is told when it does. Two products with identical interfaces can differ by an order of magnitude in what they are worth, and the interface is the part every vendor shows you.
The category was named after its input method, which is why it stalled. A team that has bought “talk to your data” has bought a way of asking. It still has the original problem: the expensive things happen while nobody is asking.
The alternative is not a better retriever. It is a structure that holds the entities the business runs on, knows which system of record owns each one, carries expectations about how they normally behave, and raises its voice when they stop behaving that way. That is an architecture, and it is judged on different evidence: not demo quality, but its own prediction error over time.
Why your RAG index should be smaller next quarter
Cortex reaches 2–3x adult spine density and then throws away 50–60% of it. Retrieval systems are the only structures we build that are expected to only ever grow.
Human cortex overproduces synapses to two or three times adult density and then eliminates 50 to 60 percent of them; in prefrontal cortex the process runs into the third decade of life (Huttenlocher 1979; Petanjek et al. 2011). Roughly 70 percent of the axons that cross between hemispheres early in life are gone later. None of this is damage. It is how the structure becomes usable: what survives is what carried signal.
Retrieval systems are, as far as we can tell, the only structures we build with the expectation that they will only ever grow. Every quarter adds sources. Nothing is ever withdrawn, because withdrawal feels like loss and nobody is measuring what each source contributes. The result is a corpus where the median document has never been retrieved and the answer quality is set by whatever noise happens to be nearest the query.
So we instrument retrieval per source, per path, per entity, and we delete. Our audits usually cut more sources than they add, and our design target is the cortical figure: minus 58 percent from peak. The index being smaller next quarter is not a regression to explain away. It is the deliverable working.
sigma = 1: the alert-fatigue constant
Every notification system has a branching ratio, whether or not anyone measured it. Below 1 the signal dies in a silo. Above 1 people turn it off. There is exactly one setting that is neither.
Cortical activity comes in avalanches whose sizes follow a power law, P(s) ~ s^-3/2, with a branching ratio of about 1.0: on average each firing neuron recruits exactly one more (Beggs and Plenz 2003). That number is not an accident of measurement. At a branching ratio of one, information travels arbitrarily far without the network consuming itself; below one it dies out locally, and above one everything fires at once and nothing is distinguishable.
Every alerting system in every company has a branching ratio, whether or not anyone has measured it. Below one is the silo: the warehouse knows, and the account manager finds out from the customer. Above one is the Slack channel everybody muted in March. There is exactly one setting that is neither, and it is measurable, which means it can be tuned and reported rather than argued about.
We report it monthly. If it drifts, you see it on the instrument panel before we mention it.
Why the company is called LEAPNODE
A myelinated axon is insulated along almost its entire length. The insulation is interrupted only at tiny bare gaps — the nodes of Ranvier, each 1–3 µm long, separated by internodes of 1–1.5 mm. The signal does not crawl along the membrane. It leaps from gap to gap, and at each gap a dense patch of sodium channels rebuilds it to full amplitude (Ranvier 1878; Huxley and Stämpfli 1949). Unmyelinated fibres conduct at 0.5–10 m/s. Myelinated ones reach up to 150 m/s.
Two orders of magnitude, and it costs less current per unit distance, not more. The speed comes from structure. And about a thousandth of the axon’s length performs all of the regeneration. That leap — and the rebuild at each node — is the idea behind the name LEAPNODE.
That is the consulting thesis stated inside the name. In most companies information loses amplitude at every hop: the warehouse knows on Monday, the planner hears a version of it on Wednesday, the customer hears nothing. We find the few places where the signal must be rebuilt to full strength, and we build only there. Everything else is insulation, and insulation is cheap.
Request a 2-week auditor read the references