ECHO-2 — closure, results and demonstration
ECHO-2 closed on 9 September 2026. The agent retains the ECHO-1 core and adds viability, pattern recognition, stream operation, consolidation, inheritance of one predisposition and regulation of two internal variables. CAPACITY-1 justified a 512 LIF + 128 Adaptive-LIF monitor architecture, and NEURAL-VIZ-1 made it observable in a native GUI.
Direct demonstration
This is not a promotional animation. The application runs Agent.turn() and
shows its WSP, CAM, T, PATTERN, Q, gate, homeostasis, death, respawn,
inheritance and occasional cortex calls. The drone represents the real pose of
the discrete Body3D body. VTK/OpenGL renders the view; aerodynamics, IMU,
motors and PID do not exist yet.
Direct comparison with the previous scale
CAPACITY-1 repeated the exam with the 256-LIF architecture used as the ECHO-1 size baseline and with the ECHO-2 extension. This comparison ran inside the same bench with reserved seeds and frozen memory; it is not a retrospective score for the ECHO-1 release.
Held-out perceptual signatures
The gain is +1,219 hits, from 40.48% to 100%. Permuting the signatures drops the same system to 142/2,048: the result depends on the representation, not merely on adding rows.
Temporal discrimination with 640 neurons in total
The total number of neurons is identical here. Temporal adaptation is the changed cause: disabling it returns 0/256. The bench selected an eight-tick memory and adaptation gain 4.
Exercised sequence scale
The third chart shows exercised scale, not accuracy on the same task: the
workloads differ. STREAM-1 processed 48 chunks, reached 4,512/4,512 known
predictions in the coherent arm and kept dynamic_alias=0.
What each phase added
| Phase | Closure evidence |
|---|---|
| VITA-1 / FOOD-1 | H falls, death ends a life, and food/poison are learned from consequences |
| SURV-1 | medians with retained memory: 28/40/40 turns; with resets: 16/16/16 |
| SHIFT-S | adaptation over frozen Q: +421/+416/+446 turns; negative transfer is also published |
| PATTERN-1 | 32/32 held-out variants versus 0/32 for exact matching; no object id or position |
| STREAM-1 | 4,608 frames, 48 chunks and 4,512/4,512 coherent predictions |
| SLEEP-2 | 8,208 rows become 144 rules; 720/720 in the exam versus T at 0/720, without rewriting CAM/T/Q |
| GEN-1f | inherited budget 8: 360 late errors versus 602 for naïve; 52 wins, 24 losses, 52 ties |
| HEAT-1b | energy + temperature: 20,786 turns versus 7,221 without temperature and 7,186 without Q; load/cool exam 12/12 |
| CAPACITY-1 | 512 LIF: 2,048/2,048; 512 LIF + 128 ALIF: 256/256 temporal |
| NEURAL-VIZ-1 | one Python GUI displays the architecture and every auditable component live |
What improves over ECHO-1
ECHO-1 closed memory, prediction, objects, discrete physical actions, temporal patterns and transfer between worlds. ECHO-2 retains them and adds a consequence across episodes: the body can die, respawn, regulate energy and temperature, retain experience and transmit only an exploration predisposition to a descendant whose memories start empty.
The neural extension is not accepted because of the number 640. It is
accepted because it improves two causal exams and loses when the responsible
feature is removed or permuted.
Integrity and data
- WSP remains 16 bytes and there is no second thought bus.
false_facts=0,destroyed=0, with the cortex disabled in the main benches.- The neural branch in the GUI is a perceptual monitor; Q remains the causal policy.
- Held-out exams are frozen during scoring.
- Earlier negative GEN-1 and HEAT-1 results remain published as red.
- This closure has no physical robot, camera, LiDAR, IMU, PX4 or AKD1500.
Download the ECHO-2 summary and SHA-256 fingerprints. That file also includes the MP4 hash.
— R.N.