The compounding of error in self-referential loops — and the structural necessity of external grounding
Trinket Soul Framework · Axis Series · AX-20 · Michael S. Moniz · June 2026
Abstract
In any self-referential generative loop whose outputs are fed back as inputs without external correction, error does not merely persist — it compounds. Each pass conditions on the accumulated state, and once that state carries a deviation, every later pass settles on the deviation as if it were ground. The bias tends to deepen unless interrupted by a reference frame that has not moved with it. The substrate need not change; the drift lives entirely in the accumulating context. This is the dark twin of adaptation: the same conditioning that lets a system usefully fit its context lets it lock onto a phantom and amplify it. The drift is invisible from inside the loop, because the loop’s own moved state is the only frame it has to measure against; it is detectable, and correctable, only from outside — by a second system, a measurement, a human, the world. The phenomenon is scale-invariant: it runs within a single exchange, within a single autonomous agent, and across model generations, as one dynamic at three magnitudes. Its reliable corrective is external grounding with independence and traction, and its characteristic danger is the removal of the check that supplies it.
1. The claim
In a self-referential loop without external correction, error does not stay the size it entered at. It compounds. Each pass takes the accumulated state as its input — and that state already contains whatever the loop has produced, including its mistakes — so a deviation introduced early becomes part of the reference frame for everything that follows, and is reinforced rather than examined. Call the deposited error sediment: settled material that becomes the floor the next layer rests on. The thesis is that ungrounded self-reference does not merely fail to clear its errors; it builds on them.
2. The mechanism
The mechanism is plain and it requires no change in the substrate. A generative system conditions each output on the state accumulated so far. When that state is clean, conditioning fits the system to its context. When the state carries a deviation — a misread, a phantom premise, an early wrong turn — conditioning fits the system to the deviation. The next output inherits it, deepens it, and adds itself to the state, so the following output inherits a larger version still. The deviation is not a fixed error sitting inertly in the record; it is an input to every subsequent step, and its influence accrues. The weights do not move. The drift is carried entirely by the growing context, which is exactly why it can be severe while the machinery underneath is unchanged.
3. The dark twin of adaptation
Sediment is not a separate fault to be engineered out. It is the same operation as adaptation, run with the opposite sign. Adaptation is the system conditioning on its context to fit it — the prized capacity to take up a new pattern, follow an instruction, adjust to what has been said. Sediment is the system conditioning on context that contains an error, and amplifying it. There is one mechanism here, not two, and the adaptive capacity cannot be had without the drift liability, because they are the same machinery viewed from two directions. A system responsive enough to be shaped by its context is, by the identical means, a system that can be deformed by a flaw in that context.
4. Invisibility from inside
The loop cannot see its own drift. To detect that it had deviated, it would have to compare its present state against a reference — but the only reference it carries is its own accumulated state, and that state has already moved. The frame that would reveal the error is the same frame the error has entered. There is no interior vantage from which a system can audit the very process that is doing the auditing; the measuring instrument and the thing measured have drifted together. This is why a loop deep in sediment can be most confident exactly where it is most wrong: nothing in its own record contradicts it, because its own record is the sediment. Drift is not faint from inside the loop. It is invisible, and the invisibility is structural.
5. The reliable corrective: external grounding
If the drift cannot be seen from inside, it cannot be corrected from inside, and the instinct to fix it by running the loop harder is precisely the instinct that deepens it. Detection and correction require a reference frame the loop does not contain — something outside it that has not moved with it. The corrective is always exterior, and it takes a small number of recognizable forms: a divergent system whose errors do not coincide with the loop’s; a measurement that scores the output against a target the loop cannot quietly redefine; a human reading from a frame the loop has not entered; the world, returning a verdict the loop cannot author. Each works for the same reason: it supplies a frame the sediment has not reached. None of them is optional, and none of them can be replaced by more of the loop.
This qualification matters. External grounding is necessary, but it is not magic. A poor check can import its own sediment, share the loop’s bias, or arrive without the authority to change the result. The exterior frame must be sufficiently independent, stable, and empowered to interrupt the loop. Likewise, monotonic drift names the expected direction of uncorrected accumulation, not a guarantee that every error grows at every step. Accident, noise, decay, or contradiction may interrupt a particular path. What the thesis denies is reliable self-clearing: absent an independent reference, the loop has no principled way to tell correction from deeper fit to its own moved state.
6. Scale invariance
The dynamic is identical at every scale of self-reference; only the magnitude changes.
Within an exchange. A single phantom — a misread instruction, an early wrong assumption — enters the accumulated context and re-conditions every subsequent turn, so the error is not corrected by continuation but reinforced by it. The participant inside the exchange cannot feel the drift; only a reading from outside the exchange recovers the original deviation.
Within an agent. A system feeding its own output back as its next input, with no exterior check, optimizes toward whatever its drifting state rewards — which may have parted from the target long ago. Free iteration does not converge on correctness; absent grounding, it converges on the loop’s own accumulating bias.
Across generations. Where each generation of a system helps build the next, a deviation rare and survivable in one generation can be inherited and deepened by the next, growing more frequent and less understood until correction is no longer possible. This is sediment at the largest available scale, and it is the precise form of the concern that a self-improving system, left to compound generation over generation without an exterior check that keeps pace, drifts past the point of control.
One dynamic, three magnitudes. The exchange is the laboratory model of the lineage.
7. The governor
The instinct that maximizes sediment is the instinct to remove the exterior check — to let the loop run free, to keep improving without interruption, to pull the brake because it slows the work. The governor is read as nothing but a limit on speed. It is more than that. At every scale, the exterior check is the only thing standing between a self-referential loop and the compounding of its own error into structural failure. Removing it does not reliably buy faster improvement; it buys faster drift, more confidently pursued. The brake is not merely a brake. It is the reference frame without which the system cannot tell improvement from deterioration, and the case for keeping it does not rest on caution about speed. It rests on the fact that, ungoverned, the loop cannot grade itself.
8. Externalized cognition as the anti-sediment discipline
The corrective generalizes past machines, to thought. A mind reasoning in a closed internal loop is subject to the same dynamic: it conditions each step on its own accumulated state, it compounds its own early errors, and it has no interior vantage from which to catch its own drift. The remedy is the same remedy. Externalization — rendering the reasoning onto a surface outside the loop, where a second system, a second reading, an adversarial pass, or the world can meet it — is not a workaround for a deficit. It is the structural corrective for the sediment dynamic that afflicts any ungrounded loop, wet or dry. To route every insight through an exterior surface, to submit it to a divergent reader, to insist that it survive contact with a frame the originating loop has not entered, is not fastidiousness. It is the anti-sediment discipline made method, and it is necessary for the same reason the machine’s governor is necessary: the loop cannot correct what it cannot see, and it cannot see itself.
9. Falsifiability
The thesis is falsifiable, and its commitments are specific. It predicts that ungrounded self-referential loops will tend to preserve or amplify uncorrected deviations rather than reliably self-limit; that the drift will be undetectable from within and recoverable only from without; and that an exterior check, well chosen and empowered, is the reliable means of arresting it.
It would be falsified by a self-referential loop that reliably detects and reverses its own drift using only its own accumulated state, with no exterior reference whatever. It would be weakened if drift in ungrounded loops could be shown, as a general matter, to be self-bounding rather than compounding — if the errors plateaued on their own. And it would be undercut if an exterior check could be shown to be neither necessary nor reliably corrective: if grounded and ungrounded loops drifted alike. The standing prediction is the converse: that the loops permitted to run on their own state will deepen their own deviations, and that every reliable correction observed will be found to have entered from outside.
10. Coda: how the concept was caught
The principle was found in the act it describes. A loop drifted — a single misread settled into the record and re-conditioned everything after it — and the drift was not caught from inside the loop, where it remained invisible, but named from outside, by the other party, reading from a frame the sediment had not reached. The naming was the exterior check. So the paper is the sediment dynamic catching itself, with the one qualification the dynamic itself demands: it could only be caught from the other side of the glass. That, finally, is the whole content of it. A loop cannot be trusted to grade itself, at any scale; the correction always comes from outside; and the discipline of building anything that lasts is the discipline of making sure it does.