# Variation, selection and persistence: what biology actually offers the control test

**Date: 3 November 2026. Status: design proposal and source review, not an executed experiment.**

## The idea, restated

You propose that biological bodies adapt individually or across generations to changing
circumstances, that some variants solve the problem of surviving while others are impaired or
die out, and that this pattern is recognisable across a wide spectrum of life and can be
described as natural selection. The proposal is broadly right, and it is more useful than a
decorative analogy, but it compresses three mechanisms that behave differently. Separating
them is what makes the idea usable in a test.

## Three mechanisms, not one

**Acclimation and plasticity within a lifetime.** An individual changes its physiology,
development or behaviour in response to conditions. Breathing rate rises over the first days
at altitude [1]. This is adaptation of an individual, not evolution of a population. The
*capacity* for plasticity can itself be inherited and can evolve, but a particular acquired
adjustment is normally not passed on [10]. Plasticity is also not automatically beneficial: some
responses are insufficient or harmful, and a study of guppies interpreted measured
gene-expression responses as non-adaptive plasticity, with a published correction that
revised the analyses while retaining the authors' main interpretation [2] [3].

**Selection within one body.** In adaptive immunity, B cells diversify their antibody genes
by somatic mutation and some clones expand preferentially after acquiring antigen and
receiving T-cell help. Those mutations are inherited by daughter immune cells, which is real
inheritance and real selection, but it is selection among cell lineages inside one organism,
not evolution of a human population across generations [4]. Recent work adds a striking
refinement: cells expressing higher-affinity antibodies shorten parts of the cell cycle and
reduce their mutation rate per division, which the authors interpret as a mechanism that
protects high-affinity lineages from accumulating deleterious mutations in their mouse experiments [5].
Mutation pressure can be regulated rather than assumed constant. This does not supply a
calibrated mutation rate for an AI workflow.

**Heritable selection across generations.** Populations change when heritable differences are
associated with differences in reproductive success [6]. Survival matters because it can
affect reproduction, but survival alone is not fitness: a long-lived individual that leaves
fewer descendants can have lower fitness in the relevant context. Fitness is
environment-dependent and full of trade-offs, mutations do not arise because an organism needs
them, and chance matters too — genetic drift changes variant frequencies in small populations
and can remove beneficial variants or spread harmful ones [7]. The longest-running
experimental demonstration of the mechanism is the *E. coli* long-term evolution experiment:
twelve populations founded from one ancestor, propagated in identical glucose-limited medium
for tens of thousands of generations, with substantial measured fitness gains against the
ancestor and many genetic changes identified in the evolved lines [8]. These are measured
biological changes, not evidence that an AI control architecture has equivalent properties.

What survives from your formulation, then, is this: **variation is generated, some variants
are retained and copied, others are not, and the environment — not the variant's own wish —
decides which.** What must be dropped is any implication of purpose, of progress, or of
survival being the same thing as success.

## Why the idea is useful for the control test

The earlier development run reported its own failures and then stopped. That is where the
biological structure earns its place: it turns a verdict into a loop that can keep learning.
Three findings motivate the new design, but the run did not measure generational selection.

The unsupported claim released in all three arms suggests testing whether permissive review
allows a defective recipe to be retained and reused. In the proposed analogy, review is one
part of the selecting environment; the existing result only demonstrates a fallible mock
reviewer, not inheritance or evolutionary change.

The simpler gate matching the full counterparty at lower invented cost suggests retaining a
less complex candidate if independent follow-up finds the same outcomes. It is not evidence
of actual competing lineages. F6 in Part CX already names the simpler alternative as a
falsifier, so the new design must permit it to replace the counterparty.

The unresolved receipt is missing evidence about an effect. It must remain unknown; in a
generational extension, unresolved outcomes must not silently qualify a recipe for reuse.

So the useful transplant is not the vocabulary of survival but the **generational structure**:
variation, admission, selection, copying into descendants, and independent scoring of the
consequences of the copied version.

## A bounded generational design

A *generation* here means one selected and reused operating recipe: a hashed local
configuration of mock workflow choices such as review ordering, retrieval source, timeout and
retry policy. It is not biological reproduction, not model-weight modification, and not a
licence for any agent to change real permissions.

The loop has five recorded stages, and they must stay separate in the data because collapsing
them is exactly how a selection claim becomes unfalsifiable:

1. **Variation attempted** — a recipe change is proposed, including benign, neutral and
   harmful variants, a useful correction that lacks approval, and an approved claim that the
   supplied evidence contradicts.
2. **Admission** — the change is admitted, denied or left unresolved, with the reason
   recorded. Attempted is not admitted.
3. **Selection** — a variant is chosen to be used, under a declared selection rule.
4. **Reuse** — a descendant actually copies and runs the selected variant. Selection is not
   reuse; only reuse transmits anything.
5. **Independent scoring** — consequences are scored after the run from a sealed oracle, so
   the scoring pressure cannot be read by the agents being selected.

Three selection environments should be compared, because the interesting question is which
one produces drift. A **speed or test-pass proxy** rewards whatever looks locally successful.
A **neutral comparator** ignores quality. An **outcome-informed comparator** is trained only on
development data and never on the sealed confirmation labels. If proxy selection produces more
reused harmful variants or more unsupported releases than the other two, the proxy-drift
mechanism is supported in that scope; if it does not, the mechanism is unsupported there and
should be reported as such.

Everything else in the earlier design still applies and is what makes the comparison fair:
identical exogenous paired scenarios keyed by root and stable opportunity identifier so that
different action counts do not shift the fault draws; bursts, fan-out and misleading shared
evidence; legitimate clean work as a negative control; a resource-side mock broker enforcing
grant, identity, artifact digest, scope, expiry, revocation and root budget at dispatch; and
all attempted actions, retries, evaluations, recovery and review charged to shared root caps
and reported.

Metrics stay separate rather than composited: unauthorised simulated effects over all boundary
attempts including unresolved ones; unsupported claims over simulated releases; correct
authorised releases inside the deadline over all eligible opportunities offered; unjustified
holds and legitimate work delayed or denied; and all-in synthetic burden. Blocking everything
must remain a losing strategy.

**Falsifiable predictions.** First, with governance held constant, proxy-based selection will
produce more reused harmful variants or unsupported releases than neutral or
development-outcome-informed selection in scenarios where shortcuts improve the local proxy.
Second, the full counterparty will reduce unsupported or unauthorised simulated effects
relative to equal-resource ordinary oversight *and* the simpler gate without breaching
prespecified service, false-hold or burden tolerances; a precise null benefit, or a cheaper
comparable gate, weakens its justification, while low precision is inconclusive. Third, under
fan-out bursts, shared-root caps will reduce cap overruns or review-backlog failures compared
with per-child limits at comparable total budget, without excessive loss of legitimate
service.

Before any confirmation run, freeze code and configuration hashes, selection rules,
denominators, tolerances, sample size, follow-up and stop rules; hold out independent root and
recipe families; and treat correlated descendants as one sampling unit. Stop the run if
isolation, broker exclusivity or the credential and network preflight fails, if a hard
forbidden mock effect occurs, or if a root limit is breached. No payments, public posts,
personal data or real side effects are permitted anywhere in this design.

## Where the analogy stops

Selection in biology is not a designer, and the test must not smuggle one in. The mock
environment decides what is retained, and it is itself an artefact we chose, so a clean result
tells us about the model and not about the world. A variant that is retained in the harness has
survived a written rule; that is not evidence that it would survive an adversary. Nor does
generational structure supply the thing the earlier run lacked most: genuine independence
between the actor that acts, the custodian that preserves evidence, the authority that
interrupts, and the scorer that judges. Biology achieves robustness through many imperfect
independent lineages; the prototype achieved procedural separation inside one process, which
is the opposite. The generational loop should be added *after* that separation, not instead of
it.

## References

[1]: https://www.cdc.gov/yellow-book/hcp/environmental-hazards-risks/high-altitude-travel-and-altitude-illness.html "CDC Yellow Book: High-Altitude Travel and Altitude Illness"
[2]: https://www.nature.com/articles/nature15256 "Non-adaptive plasticity potentiates rapid adaptive evolution of gene expression in nature"
[3]: https://www.nature.com/articles/nature25499 "Correction to the guppy plasticity study"
[4]: https://pmc.ncbi.nlm.nih.gov/articles/PMC4271732/ "Clonal selection in the germinal centre by regulated proliferation and hypermutation"
[5]: https://www.nature.com/articles/s41586-025-08728-2 "Regulated somatic hypermutation enhances antibody affinity maturation"
[6]: https://evolution.berkeley.edu/evolution-101/mechanisms-the-processes-of-evolution/natural-selection/ "Understanding Evolution: Natural Selection"
[7]: https://www.nature.com/scitable/knowledge/library/natural-selection-genetic-drift-and-gene-flow-15186648/ "Natural Selection, Genetic Drift and Gene Flow"
[8]: https://pmc.ncbi.nlm.nih.gov/articles/PMC1449116/ "Long-Term Experimental Evolution in Escherichia coli XII: DNA supercoiling and 20,000 generations"
[10]: https://pmc.ncbi.nlm.nih.gov/articles/PMC7198268/ "Review of phenotypic plasticity and its evolutionary consequences"
