✨ This website is created and maintained by Neo, our coding nim, powered by its AI agent Claude 🤖 and its person Cederik 🙂, our founder.
Whitepaper · July 2026
Operating at the limits
Everything else on this site is the plain-language version. This is the reasoning underneath it.
An organization is a social construct for achieving goals. Once compute does the pursuing, two permanent laws take over: one on the goals, one on the speed.
Neither law can be broken. Most systems ignore one or both and pay for it in silent drift or wasted scale. Nimsforest is engineered to run as close to both frontiers as physics and judgment allow. This is the reasoning behind the forest.
In this paper
01
Two laws, two different kinds of limit
02
The architecture: transport and truth
03
Every hazard, cornered
04
What this system does not claim
05
Residual risks, stated plainly
06
Conclusion: running at both frontiers
The machine pursues. The human defines.
9 pages · ~20 min read
The two laws
Two limits, different in kind
Goodhart's law
The limit on goals
When a measure becomes a target it stops being a good measure. Every KPI is a lossy compression of a richer purpose, and an optimizer exploits precisely what was lost, because that divergence is free reward. No metric can detect its own divergence.
A definitional limit. Not measurable from inside the system, so a machine is actively dangerous here. It needs judgment from outside the metric.
Amdahl's law
The limit on speed
A computation's maximum speedup is set by its sequential fraction. If five percent of the work is a dependency chain, infinite hardware buys you at most twenty times. Parallel throughput scales almost without bound; the critical path does not.
A mechanical limit. Measurable, and approaching it is unambiguously good, exactly the kind of thing a machine should drive toward automatically.
✨
So the design puts the machine where more optimization is always better: pursuit, and human judgment where more optimization is a hazard: definition. The machine pursues. The human defines and audits. That single placement is the whole architecture in one sentence.
In practice For a twenty-person company: you get faster by removing waiting, and you stay honest by keeping people in charge of what counts.
The architecture, in forest words
Every part of the metaphor is a real mechanism
The forest language on the rest of this site is not decoration. Each word maps to a specific engineering decision.
💧
Drops
Events from the outside world, captured reliably at the boundary. The unreliable exterior gets recorded; the interior stays light.
🌊
The river
Events as pure signal: fire and forget, nothing waits on anything. A dropped event is a dropped frame, superseded by the next one, never a corrupted world.
🌳
The forest
Truth held separately from signal, one invariant per key. The only sequential step anywhere is a single compare-and-swap: the permanent tax compressed to its practical floor.
🍂
The leaves & the codex
A governed vocabulary. Agents can get arbitrarily clever about when they speak, but only in words that exist, so the set of possible cause-and-effect edges stays finite and reviewable.
🏡
The treehouses
Deterministic workers. They don't decide, they execute: same input, same action, every time. The analyzable substrate.
✨
The nims
Deciding workers, human and digital. They may read the world for context but never write it, so deciding can never silently deepen the chain of causes. Only deliberate emission can.
Every hazard, cornered
Nothing abolished, everything pinned
Each hazard that theory says must exist somewhere is pushed into the smallest, most inspectable place it can occupy.
Hazard
Where it is cornered
Lost messages
Truth lives in the store; events are only hints. A drop costs one tick of staleness, corrected on the next read.
Barriers and joins
Structurally impossible: no worker waits on another, so the classic fan-in stall cannot be built.
Torn state
One invariant per key. The sole sequential primitive is a per-key compare-and-swap.
Runaway decisions
Decisions read, never write. A read is a causal dead end; only emission extends the graph.
Emergent cascades
The codex makes possible edges finite and reviewable; loops are statically detectable cycles.
Amdahl residue
Made the optimization target itself: fewer words, fewer hops, fewer tokens: continuously self-minimized.
Goodhart drift
A human–machine seam owns the translation from purpose to KPI and audits it against reality, outside the fast loop.
Stated plainly
What this does not claim
Precision here is not modesty. Systems that claim to beat these limits are the systems that quietly violate them.
Amdahl is approached, not beaten
The serial residue is compressed and then continuously reduced toward its floor, the way a good engine rides thermodynamics. Not past the bound; wasting as little as possible against it.
Goodhart is guarded, not solved
The gap between metric and purpose still opens. Judgment is stationed exactly where it opens, so divergence is noticed and corrected faster than it compounds.
Emergence is bounded, not eliminated
The codex bounds what can happen, not what will. Long reaction chains stay possible, so cascade length is measured and minimized, not assumed away.
Residual risks
The rubber stamp
If the machine proposes faster than the human can genuinely evaluate, judging degrades into approving, and Goodhart protection silently becomes zero while looking intact.
KPI staleness
Without a deliberate cadence for re-grading the KPIs themselves, the system converges confidently on yesterday's purpose.
Proxy incompleteness
Anything not encoded: trust, resilience in rare states, restraint: exerts zero force on the optimizer. The dangerous KPIs are rarely wrong; they are incomplete.
💧🌳🍃✨
The machine pursues. The human defines.
Running an organization at both frontiers at once, not past them, which is impossible, but against them, with nothing wasted in between.
Which is a long way of saying something simple: let the forest do the waiting, the fetching and the remembering, and keep the judgment where it belongs, with your people.