Model Kombat: what the telemetry says

A 12,000-neuron slice of the maleCNS connectome fights a language model in mk.js. Every match writes a JSON file; this page is generated from all of them by scripts/report_telemetry.py. Generated 18 September 2026.

140matches

433rounds

3,645logged decisions

5.5damage per hitreal wiring, curriculum vs rule bot

2.3damage per hitrewired control, same curriculum

The competitors

A match is always one fly against one opponent, and neither side is a single thing. On the fly's side what changes is how much the brain is allowed to change, and whether the wiring is the real one. On the other side what changes is where the decisions come from. Everything below is a record of one of these against one of those.

The fly brains

all of them the same 12,000 neurons of maleCNS, differing only in how they learn

Wired (no learning): The connectome as mapped, with learning switched off. What the fly does is whatever the wiring does: senses drive motor pools, the busiest pool picks the move, nothing changes.
In-match learning: The same brain, with a dopamine-like rule switched on for one match. Landing a hit strengthens whatever just fired; taking one weakens it. It starts over every match.
In-match learning + baseline: The same, except each move is judged against its own recent average instead of against zero. Without this a losing fly suppresses everything it tries, including the moves that work.
Curriculum: A run of eight matches where each one starts from the best brain saved so far, so what one match learns is still there in the next.
Curriculum + thresholds: A curriculum that also learns how certain a motor pool must be before it is allowed to act — patience as something learned rather than set by us.
Reloaded brain: A brain that finished a curriculum against one opponent, entered cold against a different one. It learns nothing new on the way in; it simply arrives trained.
Rewired control: The null model. Same neurons, same number of connections into and out of each one, same transmitter signs, same synapse weights — only which neuron connects to which is shuffled.

The opponents

all of them answering on the same state, differing in what does the answering

Rule bot: Five lines of hand-written positional rules: close the distance, strike when in range, otherwise block. No model, no learning.
Jev API: TypeSafe's System One model, reading the fight as JSON and returning one typed move with probabilities, about five times a second.
Local policy: 168 weights fitted to Jev's own probabilities — the same judgment, answered in microseconds instead of a network round trip.
Hybrid: The local policy moving at the fly's speed while Jev corrects it live, and each correction trains the policy further.

Standings

Rounds, counted from each competitor's own point of view. Every fly has now met every opponent, so the win rates are comparable — but the rounds behind them are not evenly spread, and a fly that spent most of its rounds against the rule bot is not the same as one that spent them against the hybrid. The grid below is where that shows.

WonDrawnLost

Competitor	Rounds	W	D	L	Win rate	Damage / s
Reloaded brain fly brain	50	38	0	12	76%	5.84
Curriculum + thresholds fly brain	86	45	16	25	52%	4.78
Rewired control fly brain	53	25	11	17	47%	2.16
Curriculum fly brain	140	66	13	61	47%	3.79
In-match learning fly brain	18	6	0	12	33%	2.89
In-match learning + baseline fly brain	21	7	0	14	33%	3.80
Wired (no learning) fly brain	65	8	0	57	12%	3.24
Hybrid opponent	112	62	27	23	55%	5.17
Jev API opponent	91	48	1	42	53%	5.06
Local policy opponent	135	69	12	54	51%	5.00
Rule bot opponent	95	19	0	76	20%	2.52

Who actually played whom

Every competitor against every other. The big number is the share of rounds the fly won; under it, the record as rounds won–drawn–lost. Read down a column to see what the fly's learning setup is worth against a fixed opponent; read across a row to see how far that setup carries. The tint is rounds won minus rounds lost, so a cell full of draws sits near level even when its win rate reads zero. The wiring alone loses to the whole field; a brain carried in from an earlier curriculum is the only fly with a winning record against all four.

Fly ahead on roundsLevelFly behind

Table: every match-up

Fly brain	Opponent	Win rate	Won	Drawn	Lost
Wired (no learning)	Jev API	12%	2	0	14
Wired (no learning)	Hybrid	0%	0	0	20
Wired (no learning)	Local policy	0%	0	0	16
Wired (no learning)	Rule bot	46%	6	0	7
In-match learning	Jev API	67%	4	0	2
In-match learning	Hybrid	0%	0	0	4
In-match learning	Local policy	50%	2	0	2
In-match learning	Rule bot	0%	0	0	4
In-match learning + baseline	Jev API	60%	3	0	2
In-match learning + baseline	Hybrid	0%	0	0	4
In-match learning + baseline	Local policy	40%	2	0	3
In-match learning + baseline	Rule bot	29%	2	0	5
Curriculum	Jev API	100%	24	0	0
Curriculum	Hybrid	0%	0	7	17
Curriculum	Local policy	35%	26	6	43
Curriculum	Rule bot	94%	16	0	1
Curriculum + thresholds	Jev API	20%	4	0	16
Curriculum + thresholds	Hybrid	0%	0	16	8
Curriculum + thresholds	Local policy	100%	18	0	0
Curriculum + thresholds	Rule bot	96%	23	0	1
Reloaded brain	Jev API	56%	5	0	4
Reloaded brain	Hybrid	72%	21	0	8
Reloaded brain	Local policy	100%	6	0	0
Reloaded brain	Rule bot	100%	6	0	0
Rewired control	Jev API	0%	0	1	10
Rewired control	Hybrid	29%	2	4	1
Rewired control	Local policy	0%	0	6	5
Rewired control	Rule bot	96%	23	0	1

The fly learns, against every opponent but one

Damage dealt per second by the fly, averaged over each generation's rounds. Each generation inherits the best brain saved so far and keeps learning. Against the hybrid — the local policy moving at the fly's own speed with Jev correcting it live — learning from scratch converges on not fighting at all.

Rule botJev APILocal policyHybrid

Table: fly damage per second by generation

Opponent	gen 1	gen 2	gen 3	gen 4	gen 5	gen 6	gen 7	gen 8
Rule bot	6.04	8.50	8.38	8.10	8.53	7.87	8.46	8.16
Jev API	6.14	7.16	8.15	6.93	6.34	7.01	7.03	6.90
Local policy	1.87	2.90	2.98	2.95	2.83	2.93	4.16	2.93
Hybrid	1.87	0.00	0.00	0.00	0.00	0.00	0.00	0.00

Scrambling the connectome: still learns, stops landing clean hits

The control keeps every neuron, every in- and out-degree, every transmitter sign and every synapse weight, and shuffles only which neuron connects to which. It still learns to beat the rule bot — at 2.3 damage per landed hit against 5.5, which is what a blocked hit is worth.

Real connectomeRewired control

Table: real vs rewired, damage per second

Circuit	gen 1	gen 2	gen 3	gen 4	gen 5	gen 6	gen 7	gen 8
Real connectome	6.76	8.00	7.42	7.55	7.56	7.55	7.94	7.97
Rewired control	3.05	3.80	4.40	4.16	3.93	3.90	4.04	3.84

What learning changes, synapse by synapse

Mean strength of the excitatory synapses onto each motor pool, relative to the wiring (1.0 = as mapped). The guard grows; the escape jump — a real circuit doing its job, useless against someone who keeps punching — is suppressed.

Wings (guard)TTMn (escape jump)Hind legs (kick)

Table: learned strength by pool

Pool	gen 1	gen 2	gen 3	gen 4	gen 5	gen 6	gen 7	gen 8
Wings (guard)	1.40	1.49	1.42	1.52	1.55	1.53	1.57	1.53
TTMn (escape jump)	0.96	0.96	0.96	0.96	0.96	0.96	0.96	0.96
Hind legs (kick)	1.04	1.06	1.03	1.06	1.07	1.06	1.06	1.06

Every round played, by opponent

All rounds in the telemetry, including the runs where the fly was learning from scratch and lost badly.

Fly wonDrawnFly lost

Table: round outcomes

Opponent	Fly won	Drawn	Fly lost
Rule bot	53	0	18
Jev API	40	0	24
Local policy	54	6	64
Hybrid	21	23	61
Rule bot · rewired	23	0	1
Jev API · rewired	0	1	10
Local policy · rewired	0	6	5
Hybrid · rewired	2	4	1

What each side actually does

Share of decisions by move, from 3,645 logged decisions. Jev spreads itself across closing, striking and blocking; the fly, as wired, spends a quarter of its moves jumping away.

JevFly brain

Table: decision mix

Move	Jev	Fly brain
stand	0%	0%
walk_forward	23%	28%
walk_backward	1%	0%
jump_away	0%	17%
punch	24%	0%
kick	26%	20%
block	26%	34%