The Apollo guidance computer and state estimation in pharmaceutical R&D

The strongest objection: biology has no ephemeris

Here is the objection that should be given full weight before anything else is said. Apollo's navigation worked because two-body orbital mechanics, plus a short list of known perturbations, predicts spacecraft position to within metres over days. The dynamics model was nearly exact. That is what let Stanley Schmidt's filter do its work: propagate confidently between sightings, correct gently when a sighting arrived, and never be surprised by more than measurement noise.

Pharmaceutical research and development has no equivalent model. There is no equation that propagates a compound's efficacy forward through a phase II population the way a state vector propagates a spacecraft through cislunar space. Trial registries, adverse-event reports, preprints and retraction notices do not correct a known trajectory; in biology there often isn't one. Continuous intake without a trustworthy propagation step is not estimation. It is noise arriving faster.

If this objection stands unmodified, the whole analogy collapses. A translational lead who tries to run a continuous Kalman-style update on human physiology is not doing rigorous state estimation. They are chasing whatever paper posted last week, with no dynamics model to tell them how much to trust it against the last twenty. That is a real risk, and it is worth stating plainly before defending anything.

What Apollo actually solved

Look again at what the filter needed. Three ingredients: a prior, a dynamics model, and measurements arriving over time. Apollo had an unusually good dynamics model, which is why a handful of star sightings sufficed to correct a whole trajectory. But the filter's structure did not assume a good model. It assumed a model, of whatever quality, plus a covariance that says how much to trust it. When the model is excellent, the covariance shrinks and propagation dominates between measurements. When the model is poor, the covariance stays wide and measurement dominates instead.

This is the point the objection misses. A weak dynamics model is not a reason to abandon continuous intake. It is the reason continuous intake matters more. Numerical weather prediction has no ephemeris either — the atmosphere is chaotic on a ten-day horizon — and its answer is not to trust the model longer, but to reinitialise every six hours from fresh observation and run ensembles that carry the uncertainty forward explicitly. The weaker the propagation step, the shorter the leash you give it before the next correction.

Pharmaceutical R&D is closer to the atmosphere than to an orbit. A translational lead's belief about a target's validity should be held on a leash measured in weeks, not the eighteen months a programme typically runs before its next real review.

The translational lead's eighteen months

The characteristic failure in this domain has a specific shape. A translational lead builds a programme's early rationale on a published result — say, a knockout phenotype in a preprint, or a biomarker correlation from a small trial reported at a conference. The programme is resourced, a lead compound moves into preclinical work, milestones are set eighteen to twenty-four months out. In month three, the original result is quietly withdrawn: a retraction notice, a failed replication registered in a follow-up trial, an adverse-event signal that changes the risk calculus entirely. The programme does not stop. Nobody is watching that stream against that specific programme's founding assumption. The team continues integrating forward from a state that was fixed at the point the rationale was accepted, and no further measurement update touches it until the next scheduled review — by which point eighteen months of spend, headcount and opportunity cost have been integrated on a dead reckoning.

This is dead reckoning in the literal navigational sense. The initial estimate was excellent. The people were competent. The error did not come from a mistake at the start. It came from the absence of a correction step between the start and the point where the error was finally checked, by which time it had integrated twice: once into the belief, once into the resource commitment built on the belief.

Widening the prior, shortening the horizon

So the honest response to the ephemeris objection is not to claim biology behaves like celestial mechanics. It is to change what the filter is asked to do. Where the dynamics model is strong, propagate far and correct occasionally. Where it is weak, propagate briefly and correct constantly. In pharmaceutical terms: a programme's founding evidentiary claims — the specific trial results, the specific preprints, the specific biomarker studies it depends on — should carry an explicit, narrow trust window, re-checked against the live registry and retraction record on a cadence measured in weeks, not folded once into a due-diligence document and treated as settled.

This is not a demand for more data undifferentiated. It is a demand that every load-bearing claim in a programme's rationale be tagged: which trial, which arm, which preprint version, what the current status of that source is, and how much of the programme's plan depends on it remaining true. That is provenance in the literal sense the Apollo comparison uses it — not a citation for decoration, but a record of which instrument moved the estimate, so that when the instrument is later found faulty, everything downstream of it can be located and re-checked in hours rather than rediscovered by accident at the next portfolio review.

The other failure mode: confident error

There is a second objection that cuts the other way, and it is the more dangerous one for a field that already believes in continuous monitoring. Kalman filters do not always improve with more data. If the assumed process noise is set too small, the filter's covariance collapses: it becomes certain of a state it should still doubt, and it begins discounting new measurements that disagree with it, treating disconfirming evidence as noise rather than signal. This is covariance collapse, and it is a well-documented control failure, not a hypothetical one.

A translational lead can suffer exactly this. Continuous exposure to trial registries and preprints does not protect against overconfidence; it can manufacture it. A team that has watched a hypothesis survive four consecutive updates starts treating survival as proof, and the fifth signal — the one that actually matters — gets folded into "noise consistent with prior expectation" instead of triggering a re-estimate. More streams, wrongly weighted, produce confident error faster than a static plan would, because the confidence itself compounds.

Continuous monitoring just means you find out you were wrong on a shorter cycle. It doesn't mean you're less likely to convince yourself the update is noise.

That objection is correct and it is the honest failure mode of the whole architecture. The defence is not to claim it cannot happen. It is that the architecture makes it detectable in a way a frozen rationale cannot be. Practitioners who run filters for a living watch the innovation sequence — the gap between prediction and measurement — for statistical structure; if the residuals stop looking like noise and start looking like a pattern, the model is wrong, not the data. A translational lead's equivalent is a standing check on whether disconfirming reports are being logged and reasoned about explicitly, or silently reclassified as irrelevant. A programme with provenance on its founding claims can be asked, at any point, which sources still support it and which no longer do. A programme whose rationale was written once and never revisited cannot even ask itself the question.

What the Large Universe Model claims, narrowly

Put the two objections together and the claim that survives is smaller than the one that seemed to be on offer at the start. It is not that continuous intake replaces judgement, and it is not that pharmaceutical biology has an ephemeris waiting to be found. It is that on the axis of what a system is permitted to observe, three ingredients are all there are — a prior, a propagation model, and measurements arriving over time — and that a translational lead's founding rationale, held with no further correction for eighteen months, is dead reckoning by definition, whatever the quality of the original evidence.

The remedy is not a fourth kind of input. It is what Apollo's engineers had already worked out with a computer running on 2,048 words of erasable memory: keep the dynamics model honest about its own uncertainty, keep the trust window short where the model is weak, and never let an architecture decide, by omission, that a stream stops mattering after the day the plan was written. Registries, adverse-event feeds, preprints and retractions are the sextant and the ground network for a field that never had celestial mechanics to begin with. That is exactly why they need to stay switched on.