vaxenfly
A fruit fly

Gates

5 gates, 18 arms. Every green arm has a red twin.

Proven red

Break what a gate guards, in memory, and watch it fail.

A matrix with one row per gate: a white cell for the green arm that passes, followed by dark outlined cells for the red arms, each named by the flag that breaks the thing the gate guards.

The five gates

What each guards, what it measured, and its arms.

  1. Sparsity

    gates/sparsity.mjs

    The Kenyon cell rate holds because of the measured counts: CV 0.1110 real, 0.2177 flattened, 0.3577 fixed, bound 0.15.

    • gates/sparsity.mjsmust pass
    • gates/sparsity.mjs --break=fixedmust fail
    • gates/sparsity.mjs --break=flat-aplmust fail

  2. Projection

    gates/projection.mjs

    One byte in, a little out: near beats far in 93.2% of pairs; sha256 scores -0.0847.

    • gates/projection.mjsmust pass
    • gates/projection.mjs --breakmust fail

  3. No learning

    gates/no-learning.mjs

    A small, bounded separation gain, median 12.8%, and a structure excess of 0.3584 shuffles cannot fake.

    • gates/no-learning.mjsmust pass
    • gates/no-learning.mjs --break=flat-controlmust fail
    • gates/no-learning.mjs --break=flat-basemust fail
    • gates/no-learning.mjs --break=structured-basemust fail
    • gates/no-learning.mjs --break=xcalmust fail

  4. Budget

    gates/budget.mjs

    1,210 of 22,586 raw rows dropped, 3.75% of synapses, by rules that name themselves.

    • gates/budget.mjsmust pass
    • gates/budget.mjs --break=unnamed-dropmust fail
    • gates/budget.mjs --break=extra-filtermust fail

  5. Documents

    gates/readme.mjs

    Every figure in the README, SKILL.md and this site re-derives from the build.

    • gates/readme.mjsmust pass
    • gates/readme.mjs --break=readmemust fail
    • gates/readme.mjs --break=skillmust fail
    • gates/readme.mjs --break=readme-factmust fail
    • gates/readme.mjs --break=readme-marginmust fail

How much room the thresholds have

10 judged quantities ranked by slack; inside the 0.02 warn band is shouted.

slackquantitymeasuredbound
0.0214(c1) p90 relative d' gain0.1786<=0.2000
0.0316(b3) p10 near-beats-far rate0.9316>=0.9000
0.0390(a1) CV of APL loop, worst seed0.1110<=0.1500
0.0916(c2) structure excess, ceiling0.3584<=0.4500
0.1084(c2) structure excess, floor0.3584>=0.2500
0.1997(b1) p10 near-pair median r0.9997>=0.8000
0.3032(c2) power-check excess, floor0.5532>=0.2500
0.4642(a3) x tighter than flat, worst seed1.8642>=1.4000
0.4791(b2) p10 near-far separation0.8791>=0.4000
1.5320(a2) x tighter than fixed, worst seed3.0320>=1.5000

A second connectome

On hemibrain:v1.2.1 the (c1) bound does not transfer; the structure statistic does, within 0.014.

Two dark cards name two connectomes with their cell and edge counts; below them two scales place each connectome as a marker, one inside a shared band for the statistic that transfers, one on either side of a dashed bound for the one that does not.
male-cns:v1.0hemibrain:v1.2.1
Kenyon cells4,0641,927
MBONs9771
KC to MBON edges61,21030,496
(c1) shuffle control, p900.17860.2929
(c1) flat control, p900.24290.3447
gap the bound must fit in0.06430.0518
(c2) structure excess0.35840.3440