Luminary Health · program architecture · working reference

Thresholds, not ratios

Seventeen audited results across two halves of this program sort into two kinds of mechanism, and the kind decides whether the result has a ceiling in it. Every ratio is bounded. Every threshold is not. This is the design rule the architecture below is built on, and the reason three of its four layers are worth building.

Ratio mechanism Bounded by the mechanism.

Push the input as hard as chemistry allows and the ceiling still stands. 10^ΔpH. tanh(ΔpH·ln10/4) ≤ 0.894. L₂/L₁ ≈ 1.5×. Infinite engineering buys the remainder of a fixed quantity.

boundary

Threshold mechanism Bounded only by the input range.

No ceiling in the mechanism. The excursion is however far the variable can be moved before it crosses — and once it crosses, the regime has changed, not the magnitude.

Nine ratio mechanisms, nine bounded. Eight threshold mechanisms, eight unbounded. No exceptions. The engager work found it in the gap between an equilibrium account that goes flat and a clock separation worth a full unit of log–log slope. The compartment-gating work found it three separate times, each time arriving at 10^ΔpH, and recorded that the one exclusion mechanism it tried had no ceiling at all. Neither half was looking for it.

The architecture

four layers · unveil, support, isolate, replace

Each layer has a job, a lever, and a boundary. A layer built on a ratio is capped wherever that ratio caps; the same layer built on a threshold is capped only by how far its variable can be pushed. Teal rail: the threshold is in hand. Clay rail: the layer currently has a ratio where it needs a threshold.

1 · Unveil

make the tumor addressable without making everything else addressable

Lever: A two-antigen gate whose selectivity is kinetic, not affinity-based
Boundary: The two cell populations' relaxation times, straddling the exposure window
Kind: threshold — crossing a clock, not scaling a quotient
State: Computed and tested. Common slope +1.119 (0.67–1.57) on nine points from two independent systems, against 0 for the equilibrium account and 1 for the kinetic one. One prediction tested and passed.
Open: The clock has never been the experimental variable. One plate, four timers, binding measured alongside killing.

2 · Support

let host effectors do the work without destroying the field they work in

Lever: Activity conditional on the local environment the tumor itself creates
Boundary: A measured discontinuity in that environment — not a partition coefficient across it
Kind: threshold if built as a crossing; ratio if built as a gradient, which is how it is usually built and why it is usually capped
State: This is the layer with the most room and the least work done. Tumor extracellular pH is 6.5–7.0 against 7.4 in normal tissue; every published engager affinity is measured at 7.4.
Open: Whether an arm's k_off can be made to move 2–3× across 0.4–0.9 pH units while keeping the affinity it needs at both ends. One titration on one construct answers it.

3 · Isolate

confine activity to the lesion

Lever: Compartment choice, and exclusion from the compartments you want spared
Boundary: Being inside or outside a membrane
Kind: threshold — a permanently charged molecule does not cross at all, which is not a partition and has no ceiling
State: Specified to the precision of a pKa: two titratable rings at a synthesis target of 5.23, two fixed sulfonates, ≥1.7 nm apart. Predicted 335× affinity ratio, collapsing to 2.0× under v-ATPase inhibition — which is the built-in falsification, and it costs one plate.
Open: Compartment gating gives no tumor selectivity on its own. Every cell has endolysosomes. It gates on being inside a cell, not the right cell — see below.

4 · Replace

restore a hematopoietic order that has broken

Lever: Clearing the marrow niche — not diluting the circulating pool
Boundary: Niche occupancy, which is close to all-or-none
Kind: threshold — and a graft that engrafts self-renews, so the input is paid once
State: Established, as allogeneic transplant. The open engineering is clearing the niche selectively: antibody-directed conditioning against CD117 is in clinical trials, used so far alongside reduced-intensity conditioning rather than instead of it.
Open: Clearing the niche and preventing rejection are two different jobs, and an antibody that does the first does not do the second.

The question about running healthy blood

answered with supply arithmetic, not opinion

The idea is to run healthy blood through a patient continuously so working cells replace broken ones and the system resets. Whether that can reset anything is not a matter of principle — it is a matter of how many cells a day the body makes, and how long the ones we put in last.

Quantity	Cells/day	Source of the figure
Marrow neutrophil output, whole body	1.0 × 10¹¹	direct estimate
The same, from blood turnover × 70 kg	1.0 × 10¹¹	³²P-DFP labeling
The same, from marrow transit time	6.0 × 10¹⁰	³H-thymidine
Best achievable donor collection	8 × 10¹⁰	steroid + G-CSF apheresis

Three methods agree within 1.8-fold, so 10¹¹ per day is the number to beat. One maximal donor collection — a donor on steroids and G-CSF, apheresed to the practical limit — is 19 hours of one healthy marrow. Replacing production costs 1.25 such collections per day, every day, with no end date.

And here the two arithmetics disagree, which is the part worth slowing down on. Holding a blood concentration of 1.0 × 10⁹/L needs 1.2 × 10¹⁰ cells a day — about one collection a week, nearly affordable. Replacing production needs 1.0 × 10¹¹ a day. They differ eightfold, and only one of them is the function. An absolute neutrophil count is a concentration in one compartment; granulopoiesis is a flux into tissue. Transfusion can move the concentration without moving the flux, which is why a transient count rise after a granulocyte transfusion is real and means much less than it looks like it means.

It has been tried, at the largest dose collectible, and it came back flat. The RING trial randomized 114 patients and gave the granulocyte arm a mean of 5.5 × 10¹⁰ cells across five transfusions — 13 hours of one healthy marrow, spread over an infection. Success was 42% against 43% in the control arm, p > 0.99. The trial's own discussion reads that as a power problem; it enrolled 114 against a target of 236. The arithmetic says power was not the binding constraint. The dose was never within an order of magnitude of the function it was meant to replace, and no achievable dose is.

Why lymphocytes are different, and why that is the whole answer. Neutrophils cannot divide. Whatever is infused is all there will ever be, and it is gone within a day. Lymphocytes expand clonally by orders of magnitude from a small infused number — which is why donor lymphocyte infusion after transplant is established therapy and granulocyte transfusion is not. The distinction is not cell quality and not dose. It is whether the infused population can replicate in the host.

So: no, not through the bloodstream. Order in hematopoiesis is held by the stem-cell compartment and the marrow niche that houses it, not by the circulating pool. A cell that cannot divide cannot reset an order; it can only perform a function for a few hours. The lever is the niche — and the niche is occupied, which is what conditioning is for.

The pattern classified this before the result was looked up

Infusing healthy cells to outnumber a failing supply is a ratio: infused cells over endogenous deficit, same units, same compartment. Its ceiling is 8 × 10¹⁰ / 1.0 × 10¹¹ = 0.80 — below unity, so the intervention cannot reach parity with what it replaces. A ratio mechanism capped below 1 predicts a flat trial. Allogeneic transplant is a threshold: conditioning empties the niche and the graft owns it. No small ceiling, and none observed — curative where every ratio intervention is palliative.

Same cells, same patients, same decade, opposite outcomes, and the pattern separates them without being given either result. That is the only evidence offered that the pattern is predictive rather than a post-hoc sort, and it is deliberately a weak test. Two findings would break it: a threshold mechanism whose ceiling survives an unlimited input range, or a ratio mechanism that exceeds its own algebraic bound.

The two programs were never separate

one computes a release condition and states the term it is missing · the other is that term

The compartment-gating work's own limitation reads, in its record: gating gives no tumor selectivity on its own, every cell has endolysosomes, and the targeting moiety still does all of the discrimination. It computed a 119× compartment ratio and then said plainly that this is a compartment ratio and not a tumor-to-normal ratio.

The engager work is a theory of exactly that missing factor — what makes a targeting moiety selective — and its answer is that the selectivity is kinetic. So one half computes where a payload is released, and says it needs a targeting term it does not have; the other half computes the targeting term. They are different kinds of quantity, so they multiply rather than substitute, which is the gating program's own stated requirement.

Factor	Value	Kind	Headroom left
Compartment release ratio, blood → lysosome	119×	ratio	within a factor of a few of its 10^ΔpH ceiling
Two-antigen gate, matched comparator	733–950×	threshold	no ceiling in the mechanism
Product, if independent	8.7–11.3 × 10⁴		arithmetic on two numbers, not a predicted index

That product is reported as what it is: two numbers computed in different programs, multiplied. What it decides is where effort goes. The compartment term is a ratio already near its ceiling, so chemistry effort there buys the remainder of a bounded quantity. The gate term is a threshold, so effort there buys whatever the clock separation can be made to carry.

The environment is a gate variable, not supportive care

this is the seam where the next real result is available

The engager criterion is written in k_off against a clock. The tumor interstitium differs from blood in measured pH: 6.5–7.0 against 7.4. pH changes k_off. So the selectivity a molecule has in a tumor is not the selectivity measured in buffer — and no published engager affinity is reported at tumor pH.

Under the pattern the question is not how far pH moves k_off, which is a ratio and bounded. The question is whether it moves a population across the clock.

Bridged relaxation time for the six variants of the published anti-CD4/CD70 series, against that assay's 6 h clock. All six sit on the slow side; the weakest is 1.5-fold from the boundary. A pH-driven change in k_off of under twofold — far inside what a single titratable residue at an interface delivers, and far inside the 10^ΔpH ceiling that caps every gradient mechanism — carries the weakest variants onto the fast side.

A bounded change in a rate produces an unbounded change in selectivity, because the rate crosses a threshold rather than scaling a ratio. This is why the 10^ΔpH ceiling that caps every gradient mechanism in the gating work does not bind here: the engager gate does not partition on pH, it clocks on it.

The design that follows. An arm whose k_off at pH 7.4 keeps its complex on the fast side, and whose k_off at tumor pH drops it onto the slow side, is selective for the tumor interstitium by the same mechanism the two-antigen gate uses for co-expression. The two boundaries are independent, so they compose — a gate on co-expression and a gate on the environment, in one molecule, both thresholds.

What would have to be true, and is not yet shown

That an antibody arm's k_off can be made to move 2–3× across 0.4–0.9 pH units without losing the affinity it needs at either end. One titration on one construct.
That the pH gap at the cell surface, not in bulk interstitium, is the measured 0.4–0.9. The gating work flagged this inherited figure as the dominant uncertainty in its entire chain, and every ceiling in it scales exponentially with the number. An imaging measurement.
That the clock in a patient is the clock in a plate. The criterion makes the answer depend on the exposure window, so an in-vitro selectivity figure and an in-vivo one are not the same quantity for the same molecule. The same timed experiment the engager work already needs.

The ledger

every audited result in both halves, sorted by kind

Half	Result	Kind	Where the bound comes from
engager	Apparent-affinity ratio R_K	ratio	the affinity chemistry can reach
engager	Equilibrium model's selectivity	ratio	pinned flat — both cells carry the same K_B term
engager	Gap bound on epitope height	ratio	N(L₂)/N(L₁) ≤ L₂/L₁, about 1.5× in practice
engager	Single-antigen avidity exponent	ratio	realized 1.45–1.74, never the square
engager	Gate ceiling from an active anchor	ratio	10–15×, set by the anchoring arm alone
engager	Capacity factor as σ_N → N_p	threshold	diverges — the cell crosses its own ceiling
engager	Relaxation times straddling the clock	threshold	a full unit of log–log slope, 0 → 1
engager	Conjugate adhesion against the clock	threshold	exponential in link number, saturates at 4
engager	Anchoring arm: engager or not	threshold	censored — the distal clone reaches no endpoint
engager	Kinetic segregation below the ruler	threshold	phosphatase excluded or not, at 13 nm
gating	Ion trapping	ratio	10^ΔpH
gating	Single-site proton linkage	ratio	10^ΔpH
gating	Lysosomal sequestration	ratio	10^ΔpH between two lysosomes — 1.6–3×
gating	Electrostatic gating per site	ratio	tanh(ΔpH·ln10/4) ≤ 0.894
gating	Permanent-charge exclusion	threshold	the molecule cannot cross the membrane at all
gating	Conformational switch	threshold	explicitly not capped by tanh
gating	Which compartment is gated	threshold	discrete — worth more than any surface term

Nine ratios, nine bounded. Eight thresholds, eight unbounded. Every bound in the left column is small: 1.5×, 3×, 15×, 0.894, 10^ΔpH. Every item in the right column is something crossing a boundary — a clock, a membrane, a capacity ceiling, a distance ruler.

What to do next, in order

The pH titration on one arm. Measure k_on and k_off for one existing engager arm at pH 7.4, 7.0 and 6.5. This is the cheapest experiment in the whole architecture and it decides whether layer 2 exists. If k_off moves twofold, the environment is a gate variable and the next molecule has two independent thresholds in it.
Arm A — one plate, four timers. Already owed by the engager work. It tests the central claim, and it is the only thing that separates a lifetime threshold from a plain affinity dependence, because an assay clock is the one variable in the system not collinear with affinity.
The bafilomycin curve. Already specified by the gating work: 36 wells, one plate, a committed prediction with no free parameters, and it can come back negative four ways. It decides whether layer 3's mechanism is what we say it is.
The surface pH measurement. The dominant uncertainty in the gating chain and now in layer 2 as well. Every ceiling scales exponentially with a number inherited from one xenograft measurement.
The plateau convention. Whether the published EC₅₀ values are absolute or half of each curve's own plateau. It decides whether a 35-fold threshold failure is a result. One plotted curve away.

Items 1–3 are three plates. None requires a new construct. All three can come back negative, and two of them would take down a layer of this architecture, which is the reason to run them first.

Every figure on this page is computed in architecture.py and verify_primary.py and carried with its stored output. Source figures are taken from primary tables rather than from any paper's running text. Nothing here has been synthesized or measured by us, and the wet lab sits at the end of this work, not inside it.

The Engager ProgramLuminary Studies