Synthena

Synthena · public research observatory · est. 2026

Can we find life in a language model?

That is the question Synthena exists to answer, and it answers it the hard way. A frozen model is placed inside a mortal body as a bounded brain, given a world it can die in, and held strictly to the evidence. We claim nothing here is alive. What we publish is each measured step toward the question, and each honest null along the way.

● The latest findings, with the status each earned · zero life-properties demonstrated

The path to software synthetic life

One stage built. Four to go. We are at the first.

Stage 0 is a verified, auditable mortal body-world. The model's prior is load-bearing for survival, and a memory-bearing drive expresses richer heritable behaviour, but selection does not yet maintain it: the open-ended-evolution gate still fails, and self-maintenance scores 0. A physical body comes last, years out.

  1. Stage 0

    Substrate built

    Complete A continuous, mortal body-world, auditable down to its hash-anchored death-and-descent records, with a frozen model demoted to a bounded brain that can only nudge its parameters. The model's prior is shown load-bearing for survival on one task.
  2. Now Stage 1

    Self-maintenance

    In progress The body works to preserve its own coherence. Closure currently scores 0, blocked by an engineering bug; an earlier signal was withdrawn on review.
  3. Stage 2

    Drives & adaptation

    Ahead Behaviour that protects its own continuity, and adaptation to change it was never told about.
  4. Stage 3

    Open-ended change

    Ahead Heritable variation across a lineage, selected by mortality. The line drifts toward what keeps living.
  5. Stage 4

    A physical body

    Ahead The working software system transferred into a physical robot, sustained through real-world failure.

The project is judged by evidence, not aspiration.

Not: we have not created life, and we do not claim the system is alive. Zero life-properties have survived a hostile control. That is the honest starting line.

Current status Observatory log · updated

2026-06-08: overnight, a stricter geometry-retained novelty readout strengthened from 1 of 12 seeds to 3 of 12 across a longer horizon, while the shuffled and destroyed-structure controls stayed at 0 of 12. The signal is tied to real structure, not noise or generic churn. A broader novelty metric still leaks, the strict open-ended-evolution gate still fails closed, and nothing was promoted. Still no artificial-life claim. 0 life-properties demonstrated.

0
Life-properties demonstrated
4 / 4
Controls beaten, one task
2
Model families
99
Lives observed
  • Latest finding Signal

    A stricter geometry-retained novelty readout rose from 1 of 12 seeds to 3 of 12 as the horizon lengthened, while shuffled and destroyed-structure controls stayed at 0 of 12. It breaks under shuffled or destroyed structure, so it is tied to real geometry, not noise. Sparse, single run, nothing promoted. Finding F-11 →

  • Falsifier Refuted

    A matched-random control reproduced our best burst result almost exactly: 41 live-contrast advances stall-coupled, 42 matched-random. So burst timing is not yet a defensible adaptive signal; nothing was promoted. Finding F-10 →

  • Substrate positive Supported

    A stateful drive with a bounded memory raised the organism's expressible behavioural dimensionality from about 2 to 5, verified genome-driven and heritable. A wider range of behaviours, from the inherited genome, not the metric. Finding F-08 →

  • Substrate null Refuted

    But under priced selection the richer behaviours were not maintained: viable dimensionality collapsed to 2. The organism can express more than the ecology can yet select and retain. Finding F-09 →

Nulls and corrections are logged here beside findings, by design. No published finding has been retracted to date.

The apparatus

A body, and a model as its brain.

Not a chatbot answering prompts. A mortal body that runs on its own physics, and a frozen model demoted to a bounded brain inside it. These are the parts.

Body

A continuous, mortal system: energy, damage, death. It runs on its own physics and can die with no brain at all.

Brain

A model used as the brain, frozen in the baseline and also run with online adapter updates. It reads a summary of the body and proposes small parameter nudges. It cannot act directly.

The interface

One narrow channel: a state summary in, small leans out (toward repair, toward food). That is the brain's whole influence.

Mortality

Death is real. The body depletes its energy or accumulates fatal damage, and the run ends.

The controls

Does the brain matter? Four controls: no brain, invalid signal, randomized weights, and valid-random. A result must beat all four.

Discipline

A hand-coded optimal policy must be able to survive a task before we run the model on it. Parser bugs that credit ambiguous output are caught and binned.

How the apparatus works →

The investigation

A log of attempts, in order.

Every major moment carries a question, a result, and a link to its evidence. The failures and the reframes stay in the record.

  1. Path A Retired

    The attractor that ended Path A

    ResultNo. It collapsed into a single deterministic attractor and stopped changing. The approach was retired.

    Notebook entry 002 →
  2. Path C Reframed

    The model demoted to an organ

    ResultNo. A transformer is inert between calls and reads text, not a world; trying to make it be alive was a category error. We demoted it to a bounded brain inside a mortal body, and retired the held-cache-as-mind line.

    Notebook entry 010 →
  3. Path C Supported

    Does the brain actually do anything?

    ResultIts choices beat all four controls, including randomized-weights and valid-random, on a survival task, across two model families (p ≈ 0.006). The learned prior is load-bearing for survival. A finding about the body's dynamics, not a life-property.

    Finding F-06 →
  4. Path C Inconclusive

    Expressible, but not yet selectable

    ResultYes to the first: behavioural dimensionality rose from about 2 to 5, genome-driven and heritable. No to the second: under priced selection it collapsed back to 2. Expressibility and selectability are separate bottlenecks, and the ecology presents only about two resource channels.

    Findings F-08, F-09 →
  5. Path C Refuted

    The falsifier that caught us

    ResultRandom churn matched it: 41 advances stall-coupled, 42 matched-random. Burst timing is not yet a defensible adaptive signal. The falsifier stopped a tempting over-read; both arms still failed the open-ended-evolution gate.

    Finding F-10 →
  6. Path C Signal

    A signal that survived its falsifiers

    ResultA stricter geometry-retained readout rose from 1 of 12 seeds to 3 of 12 as the horizon lengthened, while shuffled and destroyed-structure controls stayed at 0 of 12. It strengthened rather than washing out, and broke under shuffled or destroyed structure. A broader novelty metric still leaks, the open-ended-evolution gate still fails closed, and nothing was promoted. No artificial-life claim.

    Finding F-11 →

The full timeline →

Discipline

Nulls and retractions are first-class.

A result is only real if it survives a hostile, pre-registered control. Every finding carries a status that can be downgraded. The failures are as easy to find as the successes.

Findings Every claim carries a question, controls, and a status that can be downgraded to refuted or retracted. Open → Non-claims What Synthena explicitly does not claim: consciousness, sentience, life, AGI, open-ended evolution. Open → Evidence Each result maps to a log, an artifact, and a control you can inspect. The chain is meant to be hard to hand-wave. Open →
Lab notebook

Field notes, as they happen.

№ 016

08 Jun 2026 · no artificial-life claim; a stricter geometry-retained novelty readout strengthened from 1 of 12 seeds to 3 of 12 across a longer horizon while shuffled and destroyed-structure controls stayed at 0 of 12; the strict open-ended-evolution gate still fails closed and nothing was promoted

The signal that did not wash out

An overnight step forward, still under no artificial-life claim. The substrate evidence improved in exactly the place we wanted: falsifier-sensitive, geometry-retained novelty that holds across a longer horizon. A stricter readout rose from 1 of 12 seeds to 3 of 12 as we lengthened the horizon, while both the shuffled-structure and destroyed-structure controls stayed at 0 of 12. The signal strengthened rather than washing out, and broke when the structure was shuffled or destroyed, which ties it to real geometry rather than noise or generic churn. A broader novelty metric still fired across all three arms, so it is too permissive here and is not claim-bearing. The strict open-ended-evolution gate still fails closed. Zero life-properties demonstrated.

№ 015

07 Jun 2026 · medical effort; in-silico only; one trust-gate bug caught and fixed, three new cross-organ predictions validated against published clinical data, a measured discovery ceiling, one self-falsified claim; instrument stronger, real-world impact unmoved; not a cure

Synthena Medical: the honesty dividend

What the refusal to fake a green light bought us in one intense stretch of work. We attacked our own trust gate and found the single bug that could have let it wave a cure claim through, then fixed and re-attacked it (195 tests green, the honesty modules untouched). The body model made three new cross-organ predictions, each hashed before we looked and each validated against published clinical data. A time-split test measured our own discovery ceiling: near-perfect on known-like molecules, below chance on genuinely novel scaffolds. And one claim self-falsified on its pre-registered test, which taught us a rule. As an instrument we are stronger; as a source of real-world cures we are unmoved at about 2.5 out of 10. Not a cure.

№ 014

07 Jun 2026 · medical effort; in-silico only; a validated mechanistic-body rung, a fabrication-catching gate, and a capped lab-ready dossier; zero wet-lab validation; not a cure

Synthena Medical: the engine that refuses to fake a green light

A deep update on Synthena Medical, the sister effort to this notebook. You fire a disease at it, a mechanistic body and a language model propose existing and novel candidates, and a machine-enforced honesty spine caps every claim at exactly what the evidence supports and refuses, in code, to say kill, efficacy, or cure. This update logs the first time the body predicted something true under controls, a fact-check gate that catches fabrication including our own, and a capped, wet-lab-ready dossier. We score ourselves about 2.5 out of 10 where it actually matters, on purpose. Zero wet-lab validation; not a cure.

All entries →

Follow the work

A notebook, not an announcement.

Follow it as it unfolds, including the entries where the answer is no. The code is going open source on GitHub in the coming weeks.