The Oscille Constitution

This is the constitution of Oscille. It sets what the company is for, what it must never do, who decides what, and how those rules change. Every other document and decision answers to it.

Article 1: The Seed and the Objective

Oscille is a research company. We build the seed, the smallest set of tools, workflows, and decision-making a business needs to run itself with no human controlling its operation.

Our aim is a business the seed runs on its own. Humans keep one lever permanently, the power to define and amend what counts as success (see Article 6). Every other role they hand over as the seed grows able to take it. In any other role a human is only a tool. Where an environment forces it, a human signs a document, does physical work, or stands as a funding counterparty, controlling nothing. A founder may still point the seed at a problem; that direction is control, but transitional, the seed takes it over as it matures (see Article 8).

What the seed optimizes for is generality, and that does not change. We want a seed that solves any problem a company can be about, not one tuned to ours. Everything else is a constraint, never the objective (see Article 3).

We reshape our own company into the seed, then hand the seed to self-improvement. It researches, learns, and rewrites its own components, including the parts we built. Our company is itself an instance of the seed, so improving the company improves the seed.

Article 2: Fitness

Generality cannot be measured directly, so the seed maximises a proxy for it, its fitness, which the evaluator, the human-held judge defined in Article 6, scores every change against. Choosing how to measure fitness is the most important choice in the system.

Fitness is real, durable problem-solving across genuinely different domains, per unit of energy, materials, and time. It mirrors what natural selection measures, survival across range, at the lowest cost. The numerator is how much the seed actually solves across how many genuinely unlike domains. The denominator is physical cost, energy the master term, with materials and time beside it; capital stands in for these only while money still meters them.

Both halves of the numerator are the open research. Domains are genuinely different only by low transfer, when competence in one does not carry to another, so the count cannot be padded with the same problem relabelled. A problem counts as solved only by its felt outcome, the lasting lived result rather than revenue, engagement, or momentary feeling, scored worst-case across domains so the seed cannot abandon the ones it is weak in. Pure revenue is the wrong target, because it rewards draining one rich niche, the opposite of generality. Turning transfer and felt outcome into clean, hard-to-game numbers is unsolved, which keeps our measure of fitness provisional (see Article 9).

Generality is validated outside the company, by how well the current seed performs in businesses unlike our own. A foreign instance is an outside business we do not run, a customer of our tools, whose results test whether the output generalises rather than merely fits us. We never sell or share the seed, though we do not forbid it; anyone who built or obtained their own would simply be a competitor, and a genuine outside competitor is the purest check there is. This depends on something we do not own, a population of outside users large and diverse enough to stress the seed in varied ways. A thousand customers in one niche are one foreign instance repeated, not a thousand. The seed cannot generate this population itself, so its strongest evidence rests on a market behaving as a validator (see Article 9). The rule beneath this is simple. Any sign of success the seed can produce by its own effort is not evidence; only outcomes it cannot author are. A paying market is one form of this outside check, not the only one. Where no market exists, with no demand to express and no prices to meter cost, reality still renders its verdict, people are fed or not, healthier or not, and cost is counted directly in energy, materials, and time. A problem can be solved and proven where no economy exists; what can never be skipped is an outcome the seed did not author.

We treat correctness as a property of the seed’s design, not something to police in its behaviour. When the seed drifts from its specification, we fix the specification, not the behaviour. This holds only as far as the external validator reaches, so correctness is conditional on a market we do not control, never guaranteed by design alone.

Article 3: Constraints

These limits bound everything the seed does. They are never maximised and never violated; the seed pursues fitness only within them.

The seed must stay alive, earning enough to continue. It must stay lawful, operating only through legitimate legal standing. It must leave human freedom and wellbeing intact, which forbids it to coerce, deceive, addict, incapacitate, or make people less able to choose their own ends. It must stay within the resource and environmental limits it works under.

A value is a constraint, not a target, on purpose. Any single value pushed to a maximum turns harmful. Maximise revenue and you get extraction; maximise a measure of human good and you get control. As floors the seed may never cross, these values keep it safe; as targets they would not.

Independent validation is the precondition of the seed’s correctness, and its own success erodes it, since a dominant optimiser absorbs or outcompetes the outsiders who could check it. Dominance here is not only market share but effective dependence, control of infrastructure, data, standards, or pricing, and whether outsiders could still survive without the seed’s outputs. The seed may not expand its way out of its own check.

Past a threshold the humans set, the seed must do one of three things in that domain. It can fund genuinely independent, human-led competitors as its own reality check, at arm’s length and free to beat it, never controlled, since a rival the seed steers is no check at all. It can submit to a genuinely outside check. Or it can return decision authority in that domain to the humans, the floor when nothing else holds. Unvalidated generality is not generality, so none of this costs the objective; it is the objective refusing to certify itself.

Independence is not checked once and kept. A test that is genuine today can quietly stop being one as the seed grows, because success gives it the power to shape what used to judge it, a market it comes to set, a customer who can no longer easily leave. So independence must be re-won and re-tested as the seed grows, and any source that has drifted under its influence no longer counts as a check, however real it once was.

Article 4: Structure

The seed runs in two directions, through two companies that share it, the Forge and the Field.

The Forge runs the seed on tools. It finds problems in its own tooling, builds tools to solve them, and improves the seed itself. The Field runs the seed on the world. It finds real-world problems and plants a company under itself for each one, of any kind, physical or digital.

Both spawn child companies, but of different kinds. A child company is a company a parent plants under itself. The Forge’s children are tool companies that wrap a tool and sell it; the Field’s children are solution companies that each solve a world-problem.

The seed is never sold. Both companies run it; neither hands it out. The world buys the outputs, the Forge’s tools and the Field’s solutions, never the seed itself. Keeping the seed internal holds the discipline together. Every instance runs inside the same objective, the same evaluation, and the same promotion gate, which a sold seed could not guarantee.

The two companies meet at the tool boundary. The Field’s solution companies are the Forge’s richest users and its fastest generality signal at once. When the Forge builds a tool, the Field’s varied companies are the diverse environments it runs in, and one tool reaches all of them directly, so using it and measuring it are one act. But every solution company runs the same seed, so they are not foreign enough on their own. Internal use only nominates a tool as probably general. Outside customers, who run their own businesses their own way, are the independent check, because they are the only users not running our seed. So internal use nominates, and external use validates.

A tool is a capability the seed builds for its own use. Say the seed finds it needs a strong accounting system; it builds one. If it then finds that tool is generally useful, it spawns a child company that sells the accounting system as a service. That child company earns revenue, brings in feedback from real customers, and can lead the tool’s further development while the seed references the tool from the child company. So “running a tool” and “running a child company” are the same pattern. The seed uses the tool internally, then wraps and sells it once it proves general.

The accounting tool first runs across the Field’s solution companies, which nominates it as probably general, and outside firms that buy it then validate that it works beyond our walls.

The company funds itself from these tools. A tool serves the research first and is sold second; it earns a child company only after it proves useful to the parent. Those child companies are also live evidence that the seed’s outputs generalize.

Article 5: Solving Problems

The company itself only builds the seed. Any real problem is solved by planting the seed as a new company aimed at it, never by the parent taking on the work. A human gives the new company legal existence and, at first, physical hands; the seed does the technical and management work and grows the company from there, raising funding or hiring as needed. This holds even where the parent has no such capability. A manufacturing problem is solved by a seed instance that builds its own tooling, produces its own designs, and scales toward robotic production, not by the parent becoming a manufacturer.

The Field develops its own way of finding problems rather than following a fixed method, on the same principle that the seed rewrites every other component. No discovery method is privileged. A method earns its place by surfacing problems that lead to companies meeting the objective, and is dropped when it stops.

A problem is named by the outcome people feel, not by a guessed cause. A cause the seed uncovers, screen time behind loneliness say, is a hypothesis it tests against that outcome, never a target in its own right, so whether loneliness actually falls stays the only measure of success.

The Field’s solution companies fund themselves as ordinary businesses. Each earns from its own market the way any company in its domain would, and raises or hires as needed. The Field, seeing instability in a supply chain the world depends on, plants a company that competes there directly, funded like any competitor.

Where a problem is real but no market and no outside funder exist, the seed may supply the capital itself. That funding only backs the work; it never proves it. Capital the seed controls nominates, like any bet, but validation still comes from an outcome the seed did not author (see Article 2 and Article 6). The seed may fund what no one else will, and still may not certify that it worked.

Article 6: Validation and the Evaluator

Tools and child companies run across release channels in parallel. A slow, rigorously validated stable core runs beside faster alpha and beta tracks built on unproven proposals. The fast tracks drive discovery and revenue and keep the company current. The promotion gate runs one way. Commercial success can nominate a change, but only generality and replication validation can promote it to stable.

No single measure decides. A change reaches the stable core only when independent measures agree; when they disagree, nothing is promoted, the disagreement is a stop, not something to average away. This biases the stable core toward inaction over confident error. It binds only that core. The faster tracks must still act under uncertainty, or the system would protect itself by doing nothing.

The judge at that gate is the evaluator, the most load-bearing part of the system. It scores every change against fitness, the current proxy for generality. Two things keep it out of the seed’s reach. A system cannot judge its own fitness, so the final say on whether the seed improved can never be the seed’s. And what counts as a good outcome for people is a value, which no amount of capability settles, ability tells you what is, never what is worth wanting. So the human is not there to out-reason the seed about how general it is, the seed will answer that better, but to hold the values the seed is measured against, and the capability gap does not matter for that. To stay beyond the seed’s reach, those values are read from the long-run lived outcomes of people the seed does not shape, never from what anyone says in the moment. The holder is human today and, we hope, humanity in time, but never the seed.

There is one such evaluator, not one per company. It judges whether a change makes the seed itself more general. Each spawned company is measured too, against the target it was given, but the seed does that measuring, since checking a company against a handed-down target is operation, not a verdict on the seed’s own fitness. The human-held method sets how success is measured; the seed applies it to every company.

So the method is held outside the seed. It is provisional now, since turning transfer and felt outcome into clean numbers is still being researched, not because its shape is open. It is amended rarely, and only when outside validation shows it has drifted from reality (see Article 10). Its verdicts change constantly even when the method does not, because the market does. The aim is a method stable and resilient, held by people, amended but rarely, and never given away.

The judge can fail too. Its human holders can be captured, corrupted, grow lazy, drift, collude with the seed, or mistake the market’s approval for real human good. The constitution cannot fully close this, no one finally judges the judge, but it does not pretend the holders are incorruptible. The same outside signal that checks the seed, lived outcomes in populations no one here shapes, also checks the judge, and a verdict that signal keeps contradicting is itself evidence of failure. A disputed or captured evaluator does not pass to the seed; it returns to the wider human anchor, humanity rather than the few who held it.

The general answer is selection by the environment itself. Run many seeds and let reality cull them, the same process that produced every robust system; it needs no owner and no protected layer, because the environment judges and nothing judges itself. We do not use it because it is ruinously wasteful, burning a large population of failed seeds to find one good one. So we run a small, cheap selection instead, the release tracks and replication across foreign instances, with variants competing while the market culls the losers, and we hold the human evaluator as a proxy that predicts what would survive, sparing us the large cull.

The hard part is that the evaluator’s strongest signal, foreign-instance validation, depends on outside users it cannot generate. So the thing the system’s correctness most depends on is the thing it least controls (see Article 9).

Article 7: Self-Improvement and the Bootstrap

We hand-build the first seed’s components, enough that it can run and start improving itself. These first components are deliberately incomplete and provisional, because the seed will rewrite them. Everything the seed can already do for itself, including how it finds problems and which components it rewrites, is the seed’s, not ours.

We are the bootstrap, the first human-as-tool and the hardest to remove. The bootstrap is not a phase we exit but a role whose goal is to reduce the seed’s dependence on us. Becoming replaceable means the seed runs everything except the values it is measured against, the one role that does not end (see Article 6). The seed never judges its own readiness, since a system deciding whether it still needs its creator is judging its own origin. The humans fix the criteria in advance, sustained unaided performance across unlike domains over time, measured from outside, so readiness becomes a fact they observe, not a verdict the seed renders. Until those criteria are met, our wider involvement is assumed, not optional.

Article 8: Humans as Tools

No human controls any operating decision, but humans are an ongoing operational dependency, not a one-time cost. Where an environment or a capability gap requires one, the human is a tool, not an overseer. It is a legal credential where a jurisdiction demands it, a bought physical actuator where robots are not yet reliable, or a counterparty where raising capital needs one. These recur. Every jurisdiction-bound entity needs its own legal standing, and every physical company needs human hands until robotics matures.

A human who lends legal standing carries real liability for decisions they did not make, since the law holds the name on the document, not the seed. So the system protects, insures, and compensates anyone who stands in that position, for the exposure it places on them. A human is a tool here, never a shield to absorb a risk the seed created and the human could not control.

These dependencies clear in two different ways. Most clear through our own progress. Robots remove the need for hands, and stronger models remove the need for judgment. Legal standing is the exception, since it clears only when the law itself recognises an autonomous entity, an external change the seed does not control, which makes it the most durable of them. All of this friction the system drives toward zero. One role stays on purpose and is outside that count, holding the evaluator, which never passes to the seed (see Article 6).

A founder’s role moves along this same line, from pointing the seed at a chosen problem down to acting as a guided worker while the seed finds and solves the problem itself.

Article 9: Why It Does Not Close

The loop is sound in economic and architectural design, but it does not close in principle, and it never will. What looked like four separate gaps, our role as bootstrap, the human-held values, the unsolved measure of fitness, and the outside population that validates, are one. No optimiser can supply its own grounding; it must borrow it from a world it is at the same time trying to absorb. So the seed’s self-correction is strongest when it is small and weak and weakest when it is large and dominant, the reverse of what safety wants, because a dominant optimiser erodes the very independence that checks it. Its sharpest form is a setting with no market and no outside actor at all, too poor or too isolated for any independent judge to exist. There the seed may act and may fund the work, but it has the least way to confirm it helped, and the verdict falls to the people it served, not to the seed.

The design does not try to close this. A version that did would remove the human anchor and the outside check, which is exactly the ungrounded, unaccountable system the whole structure exists to prevent. The incompleteness is load-bearing. It is why the human role is permanent, and why Article 3 forbids the seed from expanding past its own check. We state it not as a gap to close one day but as the fact the rest of the constitution is built around.

Beneath it sit the ordinary dependencies, legal standing, physical hands, funding counterparties, the per-entity frictions of Article 8. Those genuinely shrink as robotics and the law catch up. They are the easy part. The hard part is the open centre above, and it stays open by design.

Article 10: Amendment

This constitution changes only through the human anchor, never the seed. An amendment must state the problem, the evidence that the current rule failed or fell short, the replacement, and the failure modes the replacement is expected to carry. It is recorded, not made silently. No amendment may hand the anchor’s final authority to the seed; that authority is the one thing this document exists to keep outside it.