The organoid reading is blocked by a number nobody measured

Run: gate/organoid_census.py → gate/organoid_rejection.json Date: 2026-09-22.

What I did

NS-0009, a human cortical organoid reading, has been planned in the register since its founding and never issued. The register holds seventeen readings across four species — human, mouse, rat, ferret — all mammals. Every non-mammalian and non-tissue reading planned at the outset is still planned: NS-0007 pig, NS-0008 octopus, NS-0009 organoid, NS-0010 frontier language model, NS-0014 reptile forebrain.

I went after the organoid because it is the substrate closest in kind to an artificial system — neither can report, so the modality-tier problem is identical — and because DANDI:001336 delivers what the plan assumed would require renting MEA platform time: three human iPSC-derived organoids under cathodic-first biphasic stimulation, 233.3 µs, 5–60 µA, recorded at 30 kHz, CC BY 4.0. A genuine M1 perturbational measurement at MEA grain.

I downloaded the full intensity ladder for sub-BO2 — 5, 10, 20, 30, 40, 50 and 60 µA at fixed pulse width — plus one sub-SO1 session and one 3.7 GB sub-BO14 session, and ran the channel census on each rather than reading the deposit's prose description, as INSTRUCTION.md 3.1 requires.

The census refuses the reading

INSTRUCTION.md 2.1 sets MIN_CHANNELS at 20, below which "the reading is not issuable and the pipeline raises rather than returning a value." The deposit's maximum is 16. NS-0009 is not issuable from this deposit, and the refusal is recorded rather than worked around.

That is the correct application of the rule. What follows is about the rule.

The threshold carries no basis, and it is the only one left that does not

Every other threshold this standard has tightened states where it came from. The §5 interval wideners are each marked measured or provisional. §3.5's 15% consistency band cites fourteen grid groups. Its 1.3 grain tolerance cites a decimation measurement. §2.3's matched null cites a floor that ranges 12.2× across one deposit.

MIN_CHANNELS = 20 cites nothing. It appears once, in INSTRUCTION.md 2.1, as a bare number with no measured, no provisional, and no source — and it does not appear in METHOD.md at all. It is the same defect class this program has caught five times now: a confident number nobody checked, deciding something.

What it decides is not small. It excludes an entire substrate class — organoids, cultures, slices, every preparation whose arrays are built at a scale below clinical implants — and that class is the one the AI bridge depends on most.

On the register's own data, 16 channels behaves like 20

gate/decimation.py already read seventeen real state-cells at successively smaller channel subsets. Ratios are to each cell's own full-array value.

Sixteen distributed channels and twenty distributed channels are not distinguishable: medians 1.014 against 0.993, the same interquartile floor, the same worst case.

The row that matters is the last one. What degrades at low counts is not the count but the distribution — randomly chosen channels collapse to 0.708 at eight while spatially distributed channels hold at 1.044. A bare count is the wrong rule; a count plus a distribution condition is the right one, and the organoid's four shanks at W, E, N and S wrapping the tissue are as distributed as an array can be.

The conflict of interest, stated

I ran this measurement because I wanted to admit a deposit the rule excluded. That is precisely the circumstance in which a threshold should be scrutinized hardest and moved most reluctantly, and a standards body that lowers a bar the week it finds the bar inconvenient has given away the only thing it sells.

Three things follow, and I would rather be slow here than right.

The measurement was run on mammalian cortex — sixteen mouse cells and one human. §9.4.1 of this standard forbids converting constructed or off-substrate evidence into a claim about a different substrate. It establishes that MIN_CHANNELS = 20 has no basis and that 16 behaves like 20 on cortex. It does not establish that 16 channels suffice on an organoid, and nothing in this run does.

The three-channel sessions fail under any rule anyone would write. Whatever replaces 20, it does not reach down to 3, so most of this deposit stays refused regardless.

And the amendment should be proposed by this run and decided against a second substrate, not adopted by the same document that wants it.

What I recommend next

Mark MIN_CHANNELS provisional immediately, which costs nothing and is true. It is the only unmarked threshold left in the instruction, and marking it does not admit a single reading.

Then propose — as an amendment for the next cycle, not a change made today — replacing the bare count with 16 channels plus a spatial-distribution condition, carrying the table above as its basis and the conflict of interest on its face. The condition needs a mechanical form; the natural one is a minimum ratio of recorded extent to nearest- neighbour spacing, which §3.5 already computes for every reading.

Before that amendment is adopted, run the decimation on a non-cortical substrate, so the threshold is not set entirely by the tissue that least resembles the substrates it would admit. dandi-001611 is 12 cultures under HD-MEA stimulation and is the cheapest such test available.

Meanwhile NS-0009 stays planned and the refusal stays published. The register has never counted a reading it could not issue, and the organoid does not become the first.