Luminary Health · orientation · revised as the edge moves

Two frontiers

The same thirteen open questions, plotted twice against the same cost axis. Once by how much of our framework stops being guesswork — that is a decision tool about our budget. Once by how far past the field's knowledge they sit — that is what the work is for. The two orderings are nearly inverted, and noticing that is the point of the page.

These were one chart until someone asked which frontier it was. An efficiency frontier is a Pareto edge of cost against value — it answers what to do this week. A science frontier is a property of the world — it answers what we are for. Collapsing them makes cheap verification look like discovery, and the cheap items are cheap precisely because nobody else is stuck on them.

The ground

settled, and not worth re-litigating

Immune engagement cures some cancers. Durable complete responses exist. The question was never whether the immune system can do it.
Potency is solved; selectivity is the binding constraint. We can kill any cell that carries an antigen. What stops solid-tumor engagers is that the antigen is also somewhere that matters.
The architectures work in principle — and none comes with a number that tells you what to build.
Antigen count per cell does not predict killing. 143-fold in count against 4.4-fold in potency, unordered.
Equilibrium affinity does not predict selectivity. Flat across every admissible parameter value where the measurement moves 98-fold.

The fence

dead, with the evidence — the half that saves time

Ruled out	By what
Affinity ratios as a design rule	structurally flat across the whole admissible range
Antigen count per cell as the variable	143-fold count, 4.4-fold potency, not ordered
The σ² avidity idealization	realized exponent 1.45–1.74, never the square
Gradient and partition mechanisms past 10^ΔpH	reached independently three times
Electrostatic gating past tanh(ΔpH·ln10/4)	0.894 ceiling
Compartment gating for tumor selectivity	gives none on its own — every cell has endolysosomes
Diluting a broken system with healthy cells	ceiling of 0.80, below unity; RING 42% vs 43%, p > 0.99
Relieving a metabolic gradient (IDO1 as the case)	ECHO-301, 706 patients, PFS 4.7 vs 4.9 months, p = 0.517
Reversing terminal exhaustion	epigenetically fixed; landscapes barely move under blockade
Avidity as the route to the epitope-height effect	bounded by L₂/L₁ ≈ 1.5× in required density
ALS via compartment mislocalization	the nucleus–cytosol gradient is zero, not small
Published object-area imaging as local-density evidence	resolves 120–140 nm against a 3.1 × 10⁻⁴ µm² footprint
Partial bridge-selectivity in a proofreading step	flat across a hundredfold — the bridged side is already saturated at 0.996

Four of those rows are ours, built and then torn down. Each survived until somebody opened the source and looked. The fence is made of our own confident prose, not of our arithmetic.

Frontier one: ours

cost against what it settles for us · answers what to do this week

Vertical position is how much of our framework stops being guesswork — our judgment, not a measurement, and the first thing to argue with. The dashed edge reaches its ceiling at one plate and goes flat: nothing more expensive settles more of what we have built.

Three points, and everything else is dominated. The plateau convention, k_off at tumor pH, and arm A — one figure and two plates. But read the defense narrowly: arm A is not valuable because it is cheap. It is the only thing that separates "we found a mechanism" from "we found a correlation." k_off and K_D are collinear at r = 0.985 and 0.997 on both ladders, so every lifetime claim we hold is currently indistinguishable from a plain affinity dependence. Until that breaks, nothing downstream is load-bearing.

Frontier two: the field's

cost against how far past what anyone knows · answers what we are for

Same thirteen questions, same cost axis, different vertical. The ladder runs: a convention (bookkeeping), unmeasured (obvious to ask, nobody did), unasked (the question itself is new), unexplained (observed and not accounted for), unseeable (no instrument reaches it). Also our judgment, and also the first thing to argue with.

Point 1 moves from near the top of one chart to the floor of the other. That is the whole argument. The plateau convention settles a great deal of what we have built and adds nothing at all to what the field knows — it is a reporting convention in one paper. Meanwhile the edge on this chart never goes flat. It keeps climbing to the right, because the things nobody has solved are expensive precisely for that reason.

What moves between them

#	The question	Cost	Settles for us	Past what anyone knows
1	The plateau convention	a figure	high	a convention — nothing
3	k_off at tumor pH	one binding run	high	unmeasured
2	Arm A — the clock as a variable	one plate	highest	unasked
4	The Hill coefficient for killing	one plate	high	unmeasured
5	Which account flattens the count panel	an E:T series	medium	unasked
6	Why epitope height decides engager-or-not	a few plates	high	unexplained
7	Exhaustion clock against selectivity clock	a few plates	medium	unasked
8	Contact area in an engager synapse	imaging campaign	low	unmeasured
9	Tumor pH at the cell surface	imaging campaign	high	unmeasured
10	Local antigen density at 10–50 nm	instrument campaign	high	unasked
11	Can discrimination pass one proofreading step	a new construct	high	unexplained and unbuilt
12	Does in-vitro selectivity transfer in vivo	a new program	high	unseeable
13	What converts bound engager into a lytic hit	a new program	highest	unseeable

Teal rows sit on our efficiency frontier. Violet rows sit high on the science frontier. Only one row is both — arm A. That is the single question that is cheap, settles the most of what we hold, and is genuinely unasked by the field. If there is a right next move, it is that one, and not because it is inexpensive.

Where we actually stand

Asset	State
A selectivity criterion in rates and a clock	derived, sensitivity-swept, slope +1.119 (0.67–1.57), n = 9, two systems
Predictions made	three
Predictions tested	one — arm zero, and it returned the shape, not the level
Measurements of our own	zero
Claims killed in audit	about seventy, nearly all prose around correct arithmetic
Claims that passed audit and later failed	none
A molecule specified to a pKa	computed, self-consistent, unsynthesized
A rule sorting seventeen results with no exceptions	held — and the lack of exceptions is itself a warning

The ratio of theory to data here is dangerous. An elaborate framework, one tested prediction, nothing we measured ourselves. Every one of the three cheap experiments can come back negative and two would take down a layer — which is the argument for running them rather than extending the framework again.

The dark

the top-right of the second chart, written out

What a cell actually requires to kill another cell. Every model here computes a bound density and assumes a monotonic map to killing. No such map exists in the literature. We showed the inferred threshold is 201-fold above anything adhesion needs, so it is a signaling requirement wearing a binding label — and that is as far as arithmetic goes.

Where the synapse is. No model in this program carries a coordinate normal to the membrane, and that coordinate decides whether an anchoring arm is an engager at all. A complete account would place the synapse, not only the complex.

Whether engineered discrimination can pass one serial step. The TCR achieves α = 2.0 on 2.67 steps. Every engineered receptor measured — CARs, cytokine receptors, our own gate — sits at or below 1. Our arithmetic now says what rate a step would need and that no known process sits there. We have mapped a piece of this frontier without crossing it.

Whether a patient's exposure window resembles a plate's. Our own criterion makes selectivity a function of the clock, so an in-vitro figure and an in-vivo one are not the same quantity for the same molecule. We are the ones predicting that, and we cannot test it.

How to use two frontiers at once

Spend from the first, aim at the second. The efficiency frontier says what to do this week. The science frontier says whether the week mattered. If a run of cheap work stops moving us rightward on the second chart, the schedule has drifted into tidying.
Prefer the overlap when it exists. Right now exactly one question sits on both — arm A. Take it.
Do not let the cheap end set the agenda. Points 1, 8 and 9 are honest work and none of them is why we are here. They are instrument calibration, and calibration is worth doing precisely once.
Re-draw both charts after the three results. They will move every point, and the second chart is the one to check for movement.

Two standing habits, both earned by being wrong

Before any new calculation, look at one figure nobody has tabulated, with no model in hand, and write down what is visible. Arm zero came from that; so did the imaging-scale row on the fence.
Check every new claim against the fence before building on it. Four of those rows are ours, and each survived for weeks because the arithmetic was correct and the sentence around it was not.

Figures from verify_primary.py, architecture.py, defense_screen.py and serialgate.py, each carried with its stored output. Source numbers come from primary tables rather than any paper's running text. Both vertical axes are judgments and are the first things to argue with; the horizontal axis is shared and is the only one close to objective. Nothing here has been synthesized or measured by us; the wet lab sits at the end of this work, not inside it.

The Engager ProgramLuminary Studies