Luminary Health · computational program

Selectivity is a statement about time

A T-cell engager carries two arms. One holds a cell; the other decides whether it stays. This program asks a single question about that arrangement — how long does the bridge last — and finds that the answer, not binding strength, is what separates a tumor cell from a healthy one. Everything here is arithmetic on published rate constants. Nothing was measured by us.

11 October 2026 · avidity-gated CD70 × CD4 engager · manuscript v34, 35 tables, 26 references · 9 open items, 24 closed

the claim, in full

Two cell populations are separable if and only if they sit on opposite sides of the exposure window — and that puts selectivity on the clock, not on the affinity.

The rival account says selectivity comes from binding more tightly to the target. Written out, that account predicts a selectivity that does not move at all when you change the gating arm's affinity: a log–log slope of 0. Ours predicts a slope of 1. They are not variants of one idea — they disagree by a whole power, and that is the thing an experiment can settle.

Held precisely: the claim is that the two populations sit on opposite sides of the clock, not that the bridge is irreversible or that either population sits at an extreme. The dual-positive population is intermediate — partway to equilibrium when the plate is read — and the model says so in print.

New to this? Start from zero — four steps, then the two pieces you can play: the atlas of thirteen tissues and the synapse at three scales.

The idea in one picture

One time axis, from ten seconds to a month. Where each band sits is what the whole program turns on. The read is the moment the experiment is scored — the plate goes into the reader at six hours and whatever has happened by then is the result.

What "settling" means. The drug arrives, arms grab, arms let go, and after a while the cell reaches a steady balance between held and released. The time to reach that balance is the relaxation time, and it is what decides whether a timed experiment differs from one left to come to rest. A healthy cell gets there in under an hour. A target cell takes five to nineteen hours — so at six hours it is still in motion, and that is why a six-hour read can tell the two apart at all.

What "weakest variant" means. Six versions of the molecule were published, differing only in how tightly the gating arm binds — from 0.9 nM at the tightest to 70 nM at the weakest, a 138-fold range. The weaker that arm, the sooner the target population settles. The two weakest settle at 5.3 and 5.5 h, just inside the read, and those same two show the weakest selectivity in the original authors' own measurements. The picture and the data agree about which end is which.

Nothing in the numbers changed from the first version of this figure. What changed is which band carries the claim. The dashed band is real and is the molecule's headline fact, but a long-lived grip on individual molecules is not the same thing as a population that has not yet settled — and only the second is what a timed assay reads.

The pieces

Eleven documents, all open. The first is the argument; the second is the same argument you can watch happen; the last two are the honest ones — where the open questions actually sit, and what this program has had to withdraw.

the manuscript · v34 Selectivity Is a Statement About Time The full derivation, 35 tables, 12 figures, 26 references, and a limitations section that names every assumption we could not close. Start at §6 if you already know the field. simulation · three acts Observations at the Synapse From 150 µm of tissue down to 450 nm of membrane gap, drawn to scale and narrated. Nine gates, seventeen clickable players, and the glycocalyx as its own color. Start here if you are new. environment 00 · where it was derived The Assay The criterion was not derived in a tumour. It was derived in a well holding three kinds of lymphocyte, a drug and a clock — no vessels, no stroma, no polarity, no barrier. This page plays that well, and then lists the six things it does not contain, which is the argument for every other module. barrier map Why Here and Not There Six real engager trials mapped onto the nine gates, with a twelve-row matrix of what stopped each one. Every row now carries evidence. atlas · thirteen tissues Where the Drug Has to Get To Thirteen disease environments drawn with their own published numbers, each with a treatment span you can play. Of 37 targets surveyed across them, 3 are physically reachable by an antibody — and 58 of the numbers that would settle the rest do not exist, so they are printed as gaps instead. 3D · drag to turn it Inside the Contact A 500 nm window on the place where a T cell meets its target, rotatable, then stepped through the ten stages between touching and dying. Our model covers four of them. Being next to each other is necessary and about 5% sufficient. architecture Thresholds, Not Ratios The rule the whole program runs on: a ratio is bounded by the mechanism, a threshold only by the input range. It is why so many published selectivity numbers are unfalsifiable. the other half · manuscript draft A Compartment-Gated Construct A molecule built to switch on only inside the lysosome, with every part placed by a closed-form bound rather than by screening. It opens with its own errata: nine findings from an independent review, all correct, which superseded the numbers below them. transfer audit What Transfers Which published constants survive the move from a cuvette to a living interface — and which do not. This is where the two-dimensional kinetics correction was found to point the safe way for an antibody arm. orientation · two charts Two Frontiers The same thirteen open questions plotted twice: once by what they settle for us, once by how far past anyone's knowledge they sit. The orderings are nearly inverted, and exactly one question sits high on both. the board · live Program Standing Board Nine open items ranked by how much falls if the answer comes back badly, and twenty-four closed ones with their outcomes. The honest version of where this stands.

How to come in

You are curious, not trained

Begin with Act 0 of the simulation — a sheet of healthy gut next to a carcinoma, at a scale where a cell is the size of a house. Then "Meet the cast."

Open the simulation →

You work in this field

Section 6 is the argument and sections 6.1–6.3 are the three places a rival account fails structurally. Section 10 is where we say what we cannot support.

Open the manuscript →

You want to find the weak point

The board, and the closed log in particular. Twenty-four entries, and the ones marked in red are ours.

Open the board →

What the program knows now that it did not

Results that stand, each one reached by building a calculation and then trying hard to break it.

result	what it is good for
Selectivity is linear in the gating arm's affinity — log–log slope +1.016, 95% interval +0.92 to +1.11, against the rival account's 0	The central result, and the one an experiment can settle. The two accounts disagree by a whole power rather than a margin.
The retention advantage belongs to the arm nobody has mutated. The gating arm's off-rate cancels out of the lifetime ratio, leaving 1/[k_off,A · (1/k_r,B + 1/k_r,A)]	Verified against the exact first-passage result to 0.3%. Both published series laddered the arm that cannot change this — so it is an untested, cheap, falsifiable prediction.
Threshold × assay time is the invariant, not threshold alone — 241 µm⁻²·h	Fixed in advance by the multi-hit arithmetic, then passed on somebody else's data. The program's first external prediction to survive contact.
The killing floor is capacity, not transport. Each bridge consumes one antigen, so bound engager can never exceed accessible antigen	No rate constants in it at all. Run backwards it puts an external ceiling on other people's thresholds — which is how item 02 could be tested without their rate constants.
There is no single antigen-density threshold. One construct needs ~10,000 sites; another kills a cell carrying 875	Four independent arrivals at the same conclusion. It is why the threshold has to be computed per construct, which is what this program does.
Accessibility beats expression. Apical restriction and lost polarity decide where an engager reaches, shown primarily for two unrelated antigens	Carried to primate tolerability on one of them, with an in-vivo image of the antigen present on both sides and uptake on only one.
The two-dimensional correction points the safe way for an antibody arm — the penalty falls on association, not dissociation	Puts all six variants on the trapped side, including the two that were borderline uncorrected. A threat to the criterion that turned out to be a support.

And what it learned by being wrong

Each row below is a claim this program made and then broke itself. They are kept in public because the results above are only worth as much as this list is complete — and because every one of them was replaced by something better than what it removed.

the claim we had made	what happened to it
Two independent methods agreed to within 0.5% — strong corroboration	KILLED — the agreement was manufactured by a round number chosen for roundness. The quantity behind it is still unmeasured.
"No antigen-density threshold has been published for engager killing"	RETRACTED — a reader found one. ~10,000 binding sites per cell, 110 cell lines, calibrated beads.
"No model in this literature contains both requirements"	KILLED — the idea is thirty years old and its author's own review states it plainly.
Weaken the effector arm to recover serial triggering	KILLED as a design rule — contradicted in sign by the only published series with assigned numbers.
An untried construct: a gated engager that clears the glycocalyx	LOST TO PRIOR ART — the fusion is already published. Searched the same day the idea was written down, under a rule adopted for exactly this.
An untried principle: selectivity from tissue architecture	LOST TO PRIOR ART — it is the stated basis of an approved antibody.
An upper density wall — very high antigen should be harder to kill	WRONG IN SIGN, WITHDRAWN — conservation gives a floor, not a ceiling. Caught by the arithmetic minutes after being written.
The lower membrane in the synapse figure, labeled "tumor cell"	WRONG, FIXED — there is no tumor cell anywhere in that dataset. Caught by a reader asking what kind of cells they were.
This page's own first figure, plotting attachment lifetimes against the clock	OVERSTATED, REDRAWN — it plotted complex survival where the criterion is on population relaxation, inflating the margin 2–51× and putting two variants on the wrong side of the clock. Caught by a reader saying the picture "may or may not be fully true." It was the same confusion the model's own output warns about in print.
The bridged complex's retention advantage comes from the gating arm	REPLACED — the gating arm's off-rate cancels. The advantage belongs to the arm nobody in the literature has mutated. Verified to 0.3%.

the pattern worth noticing The arithmetic was wrong in none of them. Every one was a claim about what the arithmetic meant, or about what nobody had done yet — never about the numbers themselves. That is this kind of work's characteristic failure mode, it is now a known one, and the audit that catches it runs on every claim before it leaves. The calculations have held from the start; what keeps improving is our account of what they show.

The rules this runs under

Nothing here was measured by us. It is arithmetic on other people's published rate constants. Every number traces to a source read for its numbers rather than its abstract.
No idea leaves this program as "untried" without a prior-art search attached to it in the same session. Two hypotheses were lost on the day this rule was written. Both were real ideas; both were someone else's already.
Ratios, not thresholds, wherever a choice exists. A ratio is bounded by the mechanism. A threshold is bounded only by whatever range of inputs you happened to try.
A claim we have inspected and a claim we have been told are recorded differently — even when they point the same way. A null result from a supplementary table we have not seen is the weakest thing to lean on.
Corrections live in the record, not in the paper. The manuscript does not relitigate its own drafting history; the board and the project docs carry all of it.

The question is not whether the drug binds the tumor cell more tightly. It is whether the tumor cell is still holding on when the clock runs out.

Luminary Health · a computational program on avidity-gated T-cell engagers. The construct is a 1+1+1 format: one anchoring arm, one gating arm, one effector arm. No molecule described here has been synthesized and no measurement reported here was made by us. Every artifact linked above is private until shared.