Synthena

Live research dashboard

Current experiments, and where each stands.

Mission-control for the investigation: the question each experiment is testing, its metrics, and the state of every arm. No marketing, only the state of the work. An arm that has not been run says so, because an uncontrolled result is not a result.

Status key Running Signal Completed Cooling Failed
EXP-05 Signal

Online weights, coupled to the body

Under testDoes letting the brain's weights slide online (small adapter updates during a life) change the body's outcomes, beyond a frozen brain and beyond no brain?

12 Dolphin n
11/12, 12/12 Final-energy wins
p ≈ 0.003, 0.0002 Sign tests
10/9/9 Alive (close, not survival)
n=3, p = 0.125 Qwen
Arms
  • Ran Online adapter updates (Dolphin) 432 updates, 420 changed-weight forwards; final energy ahead, mean +10 vs frozen
  • Ran Frozen-brain control matched, no updates
  • Ran No-brain control matched
  • Ran Qwen replication (independent family) preregistered, different family; online ahead in 3/3 on final energy, but p = 0.125
EXP-04 Completed

The load-bearing brain

Under testWith the model demoted to a bounded brain that can only nudge the body, does it actually do anything, or could noise replace it?

4 / 4 Controls beaten
2 Model families
5 Runs per arm
p ≈ 0.006 Significance
Arms
  • Ran Trained model (Qwen) beat all four controls across five runs
  • Ran Trained model (Dolphin, on Llama) reproduced independently: beat valid-random 5/5 (p ≈ 0.006)
  • Ran No brain control: dies on the perturbation
  • Ran Invalid-signal brain control: does the body just need a signal? No
  • Ran Randomized-weights brain control: same architecture, no learning
  • Ran Valid-random brain control: valid moves at random, the sharpest test
EXP-01 Cooling

Self-maintenance closure

Under testCan the organism close a self-maintenance loop on its own, with no copyable instruction in the cue: notice a decayed self-belief and act to restore it?

1 Lives
~38 Ticks
1 / 2 Control arms run
0 Closed cycles
Withdrawn Confidence
Arms
  • Ran Affordance enabled, aggressive decay closed-loop credit withdrawn on review (a scorer artifact); the closure smoke test now scores 0, blocked by an engineering launch bug
  • Ran Affordance disabled matched a trivial repeated-probe latch the scorer cannot tell from closure
  • Not run No-decay baseline pre-registered control
EXP-02 Completed

Held-cache attractor and controlled discontinuity

Under testIs the held-cache behavioural lock a structural effect, and does breaking continuity break it while a text-only cue does not?

3 Arms
n=3 / arm Replicates
200 ticks Run length
0/3 vs 3/3 Result
Arms
  • Ran Controlled discontinuity 0/3 locked, entropy 1.23
  • Ran No-reset control 3/3 locked, entropy 0.00
  • Ran SENSE-only confound control 3/3 locked; text cue alone does not break it
EXP-03 Completed

Cache-restore determinism canary

Under testDoes a saved cache restore token-identically across a fresh process? A standing release gate, re-run each cycle.

8/8 Tokens
20/20 Windows
~70 MiB Cache
0 Temperature
Arms
  • Ran Create-process vs restore-process 8/8 token ids identical; ok = true

Two headline results now stand: a frozen model's prior is load-bearing for survival, and online weight updates change body-world outcomes beyond it, early and cross-model.

Not: neither is life. The load-bearing result is one task; the online-weights result is only directional across families (Qwen n=3, p = 0.125) and shows no viability, homeostasis, or biological adaptation. Zero life-properties demonstrated.

How a completed experiment becomes a finding →