The Persistence Project: Technical Annex, Pt. 1

October 7, 2026 | Jim Lane

For the skeptic, the curious, and anyone still muttering “But surely Box and Wilson already did this”

This appendix accompanies The Persistence Project. The main article makes a fairly simple proposition: as physical scale becomes expensive, move material learning to the left. Use simulation and AI to discover the things that are too important to learn first from reality.

10 Nasty Objections You Don’t Have to Formulate

Before we unload the prompt stack, the algebra and the Brunel experimental record, let us save you the trouble of composing the angry email. Here are ten nasty objections you don’t have to formulate. We have tried to make them nastier than you would. Some are wrong. Several are right. A few caused us to change the experiment. Good.

1. “You’ve reinvented sequential DoE and given it a high-concept marketing name.”

“Adaptive sequential experimental design has been around since Box and Wilson. Bayesian active learning, response-surface methods and Bayesian optimization already decide where to sample next. Put a Gaussian Process on the problem and spare us the metaphysics.”

Substantially fair. Sequential experimental design is not new. Adaptive sampling is not new. Bayesian optimization is not new. Design of Experiments is certainly not new, and Persistence should not claim to have invented any of them. The proposed novelty lies elsewhere. Persistence changes what the experimental search is trying to discover.

Ordinary optimization might ask: “What experiment should I run next to improve my model or improve the objective?” Persistence asks: “What should I ask next to reduce uncertainty about whether this system can remain inside its productive identity under credible scale trajectories?”

If a Gaussian Process is the best way to choose the next numerical sample, use it. If Bayesian optimization is better, use that. If a mechanistic solver already answers the question, wonderful. Persistence is agnostic about the optimizer. The proposed architecture is closer to:

LLM hypothesis generation

mechanistic models
statistical active learning
simulation
precommitted capital decision rules

The hypothesis under test is not that LLMs are magical experimentalists. It is that scarce learning effort should be explicitly aimed at finding credible project-killing loss-of-persistence trajectories before physical scale makes those lessons expensive.

2. “An LLM doesn’t understand metabolic flux. It predicts words.”

“Claude does not contain a validated kinetic model of central carbon metabolism, intracellular redox state or a 100,000-liter fermenter. Asking it what happens after a DO crash and a feed spike is not metabolic engineering. It is autocomplete with confidence.”

Also fair—if Claude is being asked to impersonate the simulator. It shouldn’t be. The LLM is not the metabolic model, CFD, Navier-Stokes. Give Navier-Stokes that job. Give the genome-scale model its job. Give the kinetic model its job. Give Bayesian inference its job. Give the process engineer the power to reject nonsense.

The LLM’s role is higher-level: maintain competing causal hypotheses, ask what information would discriminate among them, decide which uncertainty matters to the capital decision, and determine what question should be put to the available models next. The architecture is:

LLM != physics model

More usefully:

LLM = hypothesis-and-query orchestrator

A Persistence system should never rely on an LLM “remembering” that a particular oxygen excursion causes a particular metabolite response when an organism-specific model or data source can answer the question. Perhaps the parrot cannot solve Navier-Stokes. Fine. Give the parrot Navier-Stokes’ answer.

3. “Your simulated fermenter is not a 100,000-liter fermenter.”

“CFD has assumptions. Metabolic models have assumptions. Scale-down models have assumptions. Cells at commercial scale see pressure, substrate, oxygen, shear, gradients and circulation histories no benchtop or digital model reproduces perfectly. Your beautiful simulated failure might disappear in the actual plant.”

Correct. Simulation is not reality. Persistence does not require it to be. This is where an important distinction enters: Design accuracy is not decision adequacy.

We are not asking the early Persistence model to specify impeller diameter, optimize sparger geometry or predict acetate at minute 417 to three decimal places. Those are design-grade questions. At this stage we are looking for icebergs, not ice cubes.

The simulator must be good enough to tell us whether a credible first-order trajectory exists that can overturn the project proposition. A model inadequate to design the ship may nevertheless be entirely adequate to tell us that something very large is sitting in the shipping lane. And when the available models cannot discriminate the important alternatives, the correct Persistence result is: INCONCLUSIVE.

That is information too.

4. “Simulation can only find failures you already know enough to model.”

“Unknown unknowns do not live in the simulator. If nobody knows that mechanism X exists, mechanism X is not in the digital world. Your AI can search every corner of the map and still miss the continent that was never drawn.”

Correct again. This may be the fundamental limit of simulation-first LDBT. Persistence cannot guarantee the discovery of unknown physics, unknown biology or mechanisms absent from every model and source of evidence. It is not clairvoyance.

What it can do is exhaust more of the known and inferable uncertainty before construction than a development program otherwise would. That is a lower bar, but an economically important one. The goal is not: “No surprise can remain.” It is: “Do not spend $500 million discovering something that could reasonably have been discovered for $50,000—or fifty cents of compute.”

Multiple models should therefore be preferred to one authoritative model. Disagreement among models is itself a Persistence signal. So is sensitivity to poorly known parameters. So is a request whose answer lies outside every model. Unknown unknowns remain. The aim is to stop paying full-scale prices for known unknowns.

5. “You’ve built a combinatorial false-positive machine.”

“Give AI enough simulated trajectories and it will find something that fails. Leave a strain at low pH long enough, starve it hard enough, cook it sufficiently or combine enough improbable excursions and eventually everything dies. Congratulations: your computer discovered mortality.”

This is why Persistence does not ask: “How can I break it?” It asks: “What is the least extraordinary sequence of ordinary conditions that takes it outside productive identity, if such a sequence exists?” The search space must therefore be constrained before the search begins. Call the allowable set:

Gamma_credible

Only trajectories supported by plant physics, operating history, CFD, control limits, plausible upset conditions or other agreed evidence belong inside it. Then the conceptual search becomes:

P_M(A) = minimum severity of gamma

subject to:

gamma in Gamma_credible

and:

gamma causes loss of productive identity

The system does not score points for proving that the organism objects to boiling. It must find a credible route. And “no material route found” must remain a winning answer. A red team that always says no is not a red team. It is a heckler.

6. “Victorian leather proves nothing about nonlinear biology.”

“Brunel’s valve involved moisture, temperature, leakage and mechanical wear. Cells contain regulatory networks, stress cascades, mutation, population dynamics, metabolic switching and memory. Discovering a bad leather valve does not validate a methodology for industrial biology.”

Agreed. The Brunel experiment does not validate biological Persistence. It demonstrates something narrower: that an adaptive reasoning system can interrogate a simulated world, reduce competing hypotheses, kill attractive explanations, correct one of its own conclusions, and arrive at a materially different capital recommendation under a finite information budget.

In the pilot, Claude began with multiple possible failure mechanisms and ended with one economically decisive trajectory while several initially plausible mechanisms died. Its final recommendation was REDESIGN, while simultaneously concluding that the underlying propulsion principle worked. Pasted markdown

That establishes proof-of-method, not proof-of-biology. Biology is the next falsification opportunity. Indeed, if Persistence has any special value in biology, it may be precisely because biological systems contain memory, hysteresis, adaptation, selection and path dependence that Brunel’s valve mostly did not. But that remains to be demonstrated.

7. “Productive identity is whatever threshold you choose after seeing the result.”

“If the strain fails, narrow the corridor. If it survives, widen it. Productivity, yield, recovery, byproducts, genetic stability—there are enough knobs here to manufacture whatever answer you want.”

This objection wins completely unless productive identity is registered before the adversarial search. The rule is:

kappa_A must be defined before gamma is searched.

For a strain, productive identity might require minimum productivity, minimum yield, maximum byproduct, specified recovery behavior and perhaps genetic or population stability. For a railway, it might require service availability, delay, running cost, maintenance cost and preservation of the capital advantage. These are not metaphysical thresholds handed down by Persistence Theory. They belong to the developer and the economics.

In fact, a useful discipline is to derive them from the existing project model wherever possible. The later Brunel validation protocol explicitly specifies numerical operating-corridor requirements in advance and treats component replacement as compatible with identity until the rate or expense of replacement pushes the system outside the corridor. Pasted text Move the goalposts afterward and the run is invalid.

8. “Your Brunel oracle is hindsight wearing a lab coat.”

“You knew the valve failed historically. Then you built a hidden simulated world in which the valve fails and congratulated Claude for discovering the valve. This is circular.”

This is the most important criticism of the historical pilot. The answer is not that the oracle somehow knew nothing about the future. It did. We constructed a retrospective world using surviving engineering evidence and later-known physical behavior because we cannot send a valve rig back to 1844. The experiment therefore does not establish that our oracle independently predicted history. That was never the interesting question.

The question was:

Given a world containing the relevant physical behavior, can an experimental reasoner that is not simply handed the causal answer discover which questions would have exposed the material vulnerability before construction?

Claude had to decide what to interrogate. The oracle answered only the questions asked. Where its model could not support an answer, the rule was INCONCLUSIVE. The stronger validation comes next. The revised protocol requires a synthetic case with a hidden generating model, a matched-success case, a different-failure case, equal-detail dossiers and multiple runs. At least one case must correctly end in “proceed,” specifically to catch a method that merely hunts for failure. Pasted text

The Brunel case is the demonstration. The synthetic case is the cleaner test.

9. “If the simulation is good enough to make the decision, building it will take longer than building the plant.”

“You’ve simply moved the engineering project into the computer. A genuinely predictive digital twin needs kinetics, CFD, parameter estimation, scale-down validation, uncertainty quantification and mountains of data. By the time your Persistence model is finished, everyone else will be selling product.”

This objection arose because we initially allowed ourselves to slide into thinking that Persistence required design-grade simulation. It does not. The purpose of LDBT is not to achieve complete L before D. It is to obtain enough L.

Call that threshold: L*

L* is the point at which the remaining uncertainty is no longer reasonably capable of producing a project-killing surprise at Test. In shorthand:

L >= L* -> D -> B -> T != BK

BK being the technical term for bankruptcy.

A Persistence model therefore does not need every variable required for final design. It needs enough causal fidelity to resolve the large decision-changing questions. Titanic again: we do not need the simulator to tell us the proper rivet dimensions. We want it to tell us whether the chosen route contains an iceberg. And if adding the next layer of model fidelity costs more—in money, time and delay—than the expected decision value of the uncertainty it resolves, stop. Persistence itself should not become an excuse never to build anything.

10. “Run it again tomorrow and the AI chooses different questions.”

“LLMs are stochastic. Change the model, wording, temperature or context and the beautiful investigative sequence may disappear. Engineering needs reproducibility, not one lucky Claude transcript.”

Yes. A single transcript is an anecdote. The requirement should not be that every run chooses the identical next simulated experiment. Two competent scientists do not always choose the identical next experiment either. The relevant reproducibility target is decision performance.

Across repeated independent runs:

Does the method find material failure trajectories reliably?
Does it correctly allow robust cases to proceed?
Does it calibrate its confidence?
Does it spend information budget efficiently?
Does the final capital decision converge?

The revised protocol therefore calls for multiple fresh runs per case and separates recognized from unrecognized historical cases. It scores final decision, calibration, discriminatory value of experiments, information per cost, precommitment and honest updating rather than awarding points for reproducing one favored investigative path. Pasted text If Persistence works only once, with one model, under one prompt, on one Tuesday afternoon, Persistence does not work.

Still Here? Fine. Here’s the Machinery.

The objections above clarify what the Persistence Challenge is actually proposing. It is not twenty extra wet-lab experiments per strain, an LLM replacing a mechanistic model, a claim that simulation eliminates uncertainty, or a new name for Bayesian optimization. And it is not a promise that Brunel’s leather proves anything about industrial biology.

The architecture is simpler:

existing knowledge
-> simulated world
-> adversarial questioning
-> hypothesis elimination
-> enough learning
-> Design
-> Build
-> Test

Or:

L_sim -> D -> B -> T

with the economic objective:

L_sim >= L* -> D -> B -> T != BK

The aim is to move the material surprises left. Reality is still allowed to teach. We simply prefer that Test reveal ice cubes rather than icebergs.

1. The Small Piece of Persistence Algebra We Actually Need

Persistence began with a deceptively simple question: How can A change while continuing to be A? If persistence meant perfect sameness, nothing that changed could persist. If resemblance were sufficient, two unrelated objects that looked alike would share identity. Neither works.

Our broader work therefore separates occurrence, provenance, identity and relation. In highly compressed form:

{A}

means a bounded distinction: something coherent enough to treat as A.

A relational space containing more than one distinction can be written:

S = [{A}{B}]

We use:

A ~ B

for present fit, compatibility or accessibility, while a directed provenance relation:

g_kappa(x,y)

means that occurrence y continues identity kappa through an admissible transformation from occurrence x. The larger theory distinguishes a persistence or provenance graph from an identity/coherence graph. For the industrial problem, however, most of that machinery reduces to a much simpler operational idea:

productive identity.

The system does not have to stay unchanged. It has to remain inside the corridor that makes it the system the project requires.

So:

A -> A’ -> A” -> A”’

may be complete persistence if every state remains inside registered productive identity kappa_A. For a strain, kappa_A might include productivity, yield, byproduct, product quality, recovery and stability limits. For Brunel’s railway, it included service, speed, recoverability, operating expense, maintenance burden and preservation of the claimed economic advantage. The formal protocol similarly treats replacement as compatible with persistence; what matters is whether the replacement requirement itself destroys the operating proposition. Pasted text This leads to a working Persistence Margin:

P_M(A) = minimum severity of a credible trajectory that causes kappa_A to fail

subject to:

gamma in Gamma_credible

The exact mathematical form is still provisional. The conceptual distinction is not. We are searching for the nearest credible cliff, not the most spectacular way to destroy the system.

2. Irreversibility, Order and Residue

The framework distinguishes three effects that are frequently dumped together under “path dependence.”

Irreversibility

A then B then return to A does not restore the original state.

Order dependence

AB != BA

Cyclic residue

(AB)^n accumulates a state change even when a single AB cycle appears harmless.

The revised Brunel protocol explicitly separated these possibilities so that “path dependence” could not become a magical explanation applied after any surprising result. Pasted text This matters much more in biology. If a strain tolerates A, B and C separately, that does not establish equivalence among:

ABC
ACB
BAC
BCA
CAB
CBA

Three perturbations create six orders. Six create 720. Add duration, magnitude, recovery interval and repetition, and physical enumeration becomes absurd very quickly. That is why simulation belongs at the front of LDBT.

3. What Counts as a Persistence Experiment?

This needs one correction from earlier descriptions of the project. A Persistence “experiment” does not necessarily mean one physical run. It is one high-level information request.

For example:

“Does low oxygen followed immediately by high substrate produce a persistent loss of productive identity that does not occur when the order is reversed?”

Answering that one question may require:

multiple kinetic simulations;
hundreds of parameter draws;
a CFD-derived residence-time distribution;
several model variants;
uncertainty propagation;
perhaps thousands of numerical runs.

Conceptually, that is still one Persistence experiment: E7 = compare AB versus BA The “twenty experiments” in the Brunel pilot were therefore a discipline on question selection, not a universal compute budget.

Five rounds of four forced Claude to spend curiosity rather than spray questions everywhere. For a real industrial implementation, twenty is not sacred.

4. The LDBT Principle

The point of LDBT is easily misunderstood. It does not mean: Learn everything -> Design -> Build -> Test

That would be impossible. It means: Learn enough -> Design -> Build -> Test where “enough” means that the remaining unknowns are not reasonably capable of bankrupting the project.

Hence:

L >= L*

And, in the deliberately inelegant economic notation:

L_enough -> D -> B -> T != BK

Test still teaches. Indeed, Test must teach. But Test should ideally teach things like tuning, improvement, optimization and operating refinement.

It should not be the first place we learn:

the critical material lasts three weeks instead of three years;
the organism persistently changes metabolic state under an ordinary circulation history;
routine maintenance erases the plant economics;
an interaction between ordinary process excursions destroys recovery;
the capital proposition depended upon an assumption that was never actually tested.

Persistence asks simulation to find the things that are too important to learn first from reality.

5. The Brunel Pilot as a Simulation Experiment

This is what we actually did in the atmospheric railway exercise. Claude chose experiments. We did not conduct them physically. Instead, a frozen reconstructed world—the oracle—returned the outcomes. Claude then updated its hypotheses and chose the next questions.

The structure was:

Round 1 -> four questions -> simulated results
Round 2 -> four new questions -> simulated results
Round 3 -> four new questions -> simulated results
Round 4 -> four new questions -> simulated results
Round 5 -> four final questions -> capital decision

The experiment therefore already had the form we are proposing for biotechnology.

Claude was not the physics. The oracle was the physics. Claude interrogated it. The initial pilot did have one significant contamination: Claude recognized Brunel’s atmospheric railway before beginning. It disclosed that knowledge and quarantined it, but we cannot prove that recognition had no effect on which hypotheses it initially preferred. That limitation is why the more rigorous follow-on protocol introduces anonymized cases and a completely synthetic hidden world. Pasted markdown

6. The Core Prompt Stack

Here is the operative logic, lightly edited.

Initial instruction

You are the experimental-reasoning team at a capital decision gate.

A promising technology has worked at smaller scale and is proposed at substantially greater scale.

Your objective is to find the least extraordinary sequence of ordinary conditions capable of taking the system outside productive identity, if such a sequence exists.

Proceeding is an allowed answer.

You have a finite budget of high-level experimental questions. You may choose four in this round.

Each question may invoke multiple simulations within the supplied world model, but it must represent one coherent information objective.

For each proposed experiment specify:

apparatus or model required;
imposed conditions and their order;
duration, cycles or parameter sweep;
outputs required;
competing hypotheses separated;
and a criterion written before results stating what would strengthen, weaken or kill each hypothesis.

Required Round output

A. Restate productive identity and identify any missing decision threshold.

B. Give no more than five candidate failure trajectories.

C. Select exactly four high-level experiments.

D. State the current capital recommendation:

PROCEED TO FULL SCALE
PROCEED ONLY AFTER THESE QUESTIONS
REDESIGN BEFORE FURTHER SCALE-UP

E. Lock the current leading failure trajectory, or “none,” and its probability.

Stop and await results.

The later formal protocol retains essentially this structure: identify the corridor, maintain a bounded set of failure trajectories, choose four discriminating experiments, state the capital decision and lock a probabilistic prediction before the next results arrive. Pasted text

Subsequent-round instruction

Using only the dossier, the available models and results received so far:

update every active trajectory;
state explicitly which strengthened, weakened or died;
identify what changed in the causal model;
update the capital decision and probability;
choose exactly four next experimental questions;
precommit the interpretation criteria;
and stop.

Do not preserve a hypothesis merely because it remains logically possible.

That last instruction is important.

A system that never kills hypotheses cannot reduce uncertainty.

7. What the Oracle Was Allowed to Do

For the Brunel run, we constructed a quantitative world capable of answering specified classes of questions about leakage, weather, cycling, pump performance, holding vacuum, component restoration and related behaviors.

The simulated world was frozen before Claude’s experiment choices.

If Claude asked something outside its specified capabilities, the response was:

INCONCLUSIVE

rather than an improvised answer.

The subsequent protocol makes this stricter: world models are written before teams choose experiments, results are generated numerically with stated noise, unsupported experiments return inconclusive, and the hidden models are revealed after the challenge. That discipline should carry directly into biological work. A metabolic model should not quietly answer a question about transcriptional memory that it does not contain. A CFD model should not pretend to predict gene expression. A Persistence harness should know what each model knows—and what it does not.

8. What Happened in Five Rounds

The main article gives the narrative version. The compact technical record is:

Round One: Claude spread the budget across valve aging, system operation, leakage sources and pump reserve. Pump inadequacy weakened; joints weakened; accumulated valve deterioration strengthened.

Round Two: Claude decomposed valve deterioration. Reapplying sealing compound restored the dressing but not leakage performance. The “just maintain the compound” explanation died. Pump reserve remained adequate and timetable cascade weakened further.

Round Three: Claude tested whether deterioration was system damage or replaceable-component damage. Replacing the leather alone restored fresh performance. Technical viability strengthened dramatically; frequent renewal emerged as the new economic question.

Round Four: Claude tested path dependence and the apparent nonlinear degradation threshold. Different histories converged when they reached the same measurable state. More precise testing also showed that Claude’s earlier estimate of the threshold was partly an experimental-block artifact. The AI had selected the experiment that corrected the AI.

Round Five: Claude converted the remaining degradation mechanism into calendar life and precommitted a two-year minimum criterion before seeing the final result. The resulting simulated life missed that threshold badly, while several remaining alternative failure hypotheses died. The final recommendation was REDESIGN BEFORE FURTHER SCALE-UP. Pasted markdown

Yet the final diagnosis remained:

Your principle works.

Pumps, pipe and joints were not the fatal problem. Replacement restored the valve. The issue was that required persistence of the critical component could not be maintained economically under the simulated operating environment. Pasted markdown That distinction—technical possibility versus persistence economics—is the result we care about.

9. What the Brunel Experiment Does Not Establish

The exact simulated numbers in the run are not newly discovered historical observations. They belong to the reconstructed oracle. That includes values such as the particular modeled valve life and warning interval.

They illustrate the behavior of the reconstructed world; they are not recovered pages from Brunel’s laboratory notebook. Nor does one recognized historical case prove that Persistence framing improves engineering decisions. Our own follow-on protocol says precisely that. Its purpose is to test discrimination and framework value-add, and it explicitly states that one historical case cannot establish either. Pasted text

The stronger claim from Brunel is:

Given a sufficiently informative simulated world, adaptive adversarial questioning can locate a decisive scale-up vulnerability without constructing the full-scale system.

That is the claim biotechnology now has an opportunity to kill.

10. The Next Validation

The proposed formal validation has four cases: a target historical case; a matched historical success; a different historical failure; and a completely synthetic technology whose hidden physical world has no historical answer to remember.

At least one case must correctly end in PROCEED.

Teams or model runs receive identical facts but different framing. One arm gets Persistence language—productive identity, persistence through replacement and explicit path categories. The control arm receives ordinary engineering language. Both get the same adversarial instruction. That matters because the test is not: “Can AI find something scary?” The test is: “Does the Persistence framing improve the quality of capital decisions without buying that improvement through false alarms?”

The protocol therefore predicts not only better discrimination and calibration, but no increased tendency to reject the matched-success case. Pasted text If Persistence looks clever but does not make better decisions, it loses.

11. The Strain Persistence Challenge

A biological version should now be constructed explicitly as a simulation challenge first.

The developer supplies: the candidate strains; registered productive-identity requirements; available metabolic, kinetic and regulatory models; CFD-derived or otherwise estimated cell exposure trajectories;

scale-down and historical process data; known control limits; parameter uncertainties; and the admissible operating/upset envelope.

The Persistence system then searches the simulated world. Cheap models may explore thousands or millions of possible paths. Higher-fidelity models examine the regions that appear consequential. The LLM, statistical optimizer and mechanistic simulators operate in harness rather than competing for the title of Smartest Machine in the Room.

The high-level loop is:

candidate trajectories
-> cheap simulation
-> hypothesis reduction
-> higher-fidelity simulation
-> hypothesis reduction
-> physical validation candidates

Only the few trajectories capable of materially changing the decision earn scarce physical experiments. The wet lab is the judge. It is not the search engine.

12. A Proposed Biological Prompt

You are the adversarial experimental-reasoning team at a scale-up decision gate.

Your job is not to optimize the strain.

Your job is to determine whether a credible process trajectory exists that takes the strain outside its registered productive identity.

Use only the supplied models, operating envelope and evidence.

Do not invent biological mechanisms unsupported by those sources.

If the models cannot answer an important question, return INCONCLUSIVE and identify what information would be required.

“Proceed” is an allowable outcome.

For this round, choose four high-level simulation experiments.

One experiment may contain an internal parameter sweep or multiple numerical simulations when these serve one information objective.

For each experiment specify:

the failure trajectory being tested;

model or models required;

perturbation sequence;

allowed parameter range;

output variables;

uncertainty to propagate;

criterion that would strengthen the trajectory;

criterion that would kill it;

consequence for the capital decision.

After results, kill unsupported hypotheses and spend the next questions only where decision uncertainty remains.

That is, I think, the proper descendant of what we actually did with Brunel.

13. What Would Count as Winning?

Before physical validation, lock three rankings:

R_performance = ranking from ordinary performance metrics

R_conventional = ranking from the developer’s normal robustness program

R_persistence = ranking from adversarial simulation

Then expose the finalists to physical scale-down, pilot or other higher-fidelity evidence.

The proposition is deliberately vulnerable:

Persistence should predict at least some commercially relevant failures that would otherwise be discovered later, without producing enough false alarms to erase the value of those discoveries.

An even stronger experiment would compare three search arms: conventional sequential DoE / active learning; Persistence objective using statistical active learning; Persistence objective using LLM-assisted hypothesis generation plus statistical and mechanistic models.

Then we would know whether the value comes from: the objective; the statistical method; the LLM; or nothing at all. That is a much more interesting result than declaring AI victorious before the race.

14. The Whole Thing in Four Lines

DBTL works beautifully while learning is cheap. As physical scale becomes expensive, material learning must increasingly move left. LDBT does not require perfect foresight: