Synthena has not landed artificial life. What it has done is cross from scattered promising runs into an auditable experimental spine, with one new result strong enough to report and bounded enough to trust: online sliding adapter weights measurably move short-horizon body-world outcomes, under matched controls. The honest headline is exactly that, and the second half matters as much as the first. Sliding weights now move outcomes, and this is still not artificial life.
The strongest new result: final energy, under matched controls
The question is whether letting the brain’s weights slide online, through small adapter updates during a single life, changes the body’s outcomes, beyond a frozen brain and beyond no brain at all. We ran an online-adapter arm against matched frozen-brain and no-brain arms.
In Dolphin, a Llama-derived family, at n=12 paired seeds of 36 ticks each, the online updates were real and used: 432 adapter updates, 420 of them feeding changed weights into later forward passes. The online arm finished with more energy than the frozen arm in 11 of 12 seeds, and more than the no-brain arm in all 12, with mean final-energy deltas of about +10 against frozen and +11 against no-brain. The exact one-sided sign tests are p ≈ 0.003 against frozen and p ≈ 0.0002 against no-brain.
This is a final-energy result, and only that. It is not a survival claim. The uninterrupted alive counts are close: 10 of 12 online, against 9 of 12 frozen and 9 of 12 no-brain. In this code, “alive” and “death” are strict bookkeeping terms, where death totals count lifecycle death-and-rebirth events rather than simply “ended dead”, so the energy advantage does not transfer to a survival advantage. The wording we will stand behind is the careful one: online-updated weights produced a consistent final-energy advantage in this bounded lifecycle packet. Not: the organism learned to survive.
Qwen replication: directional, underpowered
To check it was not a single-model quirk, we preregistered and ran an independent family, Qwen, a different base with a different adapter seam, at n=3. The online arm beat both controls in all three seeds, mean delta about +12, but the exact sign test is p = 0.125. That is directional support, not statistical replication. It weakens a Dolphin-only-quirk reading, and nothing more.
The substrate is now auditable, not just pretty
Around that result, the substrate has become something you can attack. A control-audit panel runs real, shuffled, removed, and destroyed-structure variants for nutrient, presence attraction, presence avoidance, and hazard contact, and the right signals break when the structure is destroyed or removed. That is substrate-mechanics evidence, not cognition evidence, but it is the kind of falsifier that makes the rest worth reading.
A selection panel found replicated, bounded heritable-coupling signals for hazard, nutrient, and presence avoidance and threat, with presence attraction mixed. That supports selection-like consequences over authored couplings, not open-ended evolution. And the lineage verifier now checks that a reborn frame carries the same inherited child couplings recorded in its birth lineage record, so a coupling value can no longer be silently edited and pass by refreshing hashes. The artifacts themselves still carry explicit no-promotion and no-artificial-life flags.
What we are not claiming
No artificial-life claim. No life claim. No consciousness claim. No robust survival claim. No general adaptation claim. Serial death and rebirth with descent records is not reproduction in the strong sense, and we do not call it that.
The strongest wording we will stand behind is narrow: early evidence that an online learning seam couples to a mortal body and moves its short-horizon energy outcomes. The coupling is not decorative. That is all, and it is enough to keep going.
Distance from artificial life
The next gates are concrete: scale the Qwen replication to n=6 and n=12; run lives longer than 36 ticks, toward a real survival horizon; add kill-switch controls on the online-adapter seam itself; show that inherited couplings causally matter after rebirth, not just that they are recorded; and ship a public audit bundle and viewer that shows lifecycle continuity cleanly.
Artificial life remains the target, not the claim. We’ll post the next update when there’s something measured to report.