Facility buildout

The canonical plan for the facility-scale program: what has been decided, why, how each decision is kept from decaying, and what remains open.

What belongs here

This page is the one to read before doing facility work, and the one to update when a decision changes. It holds decisions and their reasoning, which nothing else in the repository does.

The roadmap tracks releases. It says what ships.
The backlog tracks framework work items. It says what remains.
Domain proposals hold candidate technology domains, one file each, against the screen in D12. D6 cannot record a choice without one.
This page tracks the program. It says what was decided and why, so a decision survives the conversation that produced it.

The distinction matters because of a demonstrated failure here. The 2026-08-05 audit is 945 lines of correct analysis whose own header records that nothing in it has been acted on. A document disconnected from work decays into archaeology. Every decision below therefore names the mechanism that enforces it, and a decision enforced only by intention is marked as such, because that is the honest description of its durability.

Decision log

Each entry carries a status and the thing that would catch its violation. Enforced by: intent means nothing would catch it, which is a standing invitation to find a better mechanism.

#	Decision	Status	Enforced by
D1	Facility scale	closed	intent
D2	A facility is not N cells	closed	intent
D3	The 10x claim is decisions per unit time	closed	intent
D4	Scale invariance: the small case stays first-class	closed	`tests/test_minimal_laboratory.py`
D5	The showcase proves the chain	closed	intent
D6	Target technology domain	open	`tests/test_domain_proposal.py`
D11	Capital intensity caps venture returns	closed	`tests/test_domain_proposal.py`
D12	Screen the cost share before spending research on a domain	closed	`tests/test_domain_proposal.py`
D13	The facility's first deliverable is the predictive bridge	closed	intent
D14	The twin renders, the adapter acts	closed	`tests/test_twin_is_read_only.py`
D7	Business model	closed	intent
D8	Scheduler architecture	open	`tests/test_minimal_laboratory.py`
D9	Framework work the facility requires	closed	intent
D10	Dogfooding does not become facility-only	closed	`make showcase`

Seven of fourteen rest on intent, and that count is printed here so it stays uncomfortable. Three of them — D1, D3 and D9 — become self-enforcing the moment the work they describe exists, and are honest placeholders until then. The other four are process rules that no test can express.

Audited 2026-08-24. Two entries claimed mechanisms that did not exist: D10 described a facility laboratory in the benchmark suite, which has never been there, and D1 and D3 both named the facility showcase, which has not shipped. Every row above was re-checked against the repository.

D1 — Facility scale

Decided. The next showcase is a laboratory facility. A single cell has already been shown.

A single cell has already been demonstrated and a second one proves nothing new. The interesting claim is that a facility is a different machine — see D2. Throughput alone would prove little.

Enforced by: intent, until the facility showcase exists. The current showcase is discovering-colors, which is a single cell, so nothing today would catch this decision being quietly abandoned. It becomes self-enforcing the moment a facility showcase ships, and not before.

D2 — A facility is not N cells

Decided. Facility scale is a qualitative change, and the design must reflect all of it:

heterogeneous stations, each doing a different thing
physical samples moving between stations, with custody and genealogy
contention for shared characterization instruments
multiple timescales in one loop — minutes to weeks — so work has to pipeline
graceful degradation when one station fails with dozens of experiments in flight
multi-fidelity triage: screen many, validate some, qualify few
human technicians as first-class stations for steps that cannot be automated

Every one of these maps onto a primitive the framework already has and has never stressed: capability contracts, resource leases, lifecycle states and attestation, provenance, and the human-task adapter. That is the reason to build it.

Two worked reference designs. The domain search produced two facility architectures that are worth keeping even though both ventures were falsified. They are the concrete form of the list above, researched to real instruments, real costs and real wall-clock.

Elastocaloric alloys. Nine steps from melt to regenerator assembly. Differential scanning calorimetry answers in 15 minutes at $10 a sample; functional fatigue to 10⁸ cycles takes 58 days at $15,000-30,000. That is a fast/slow ratio of about 5,500. The contended instruments are servo-hydraulic and resonant frames, which every fatigue campaign queues on, plus electron backscatter diffraction, transmission electron microscopy, and a single regenerator rig. Custody runs melt → work → heat treat → machine → coupon test → tube form → assembly, with a specific ingot heat traceable through six operations.

CO2 electrolysis. Five decades of timescale in one workflow: cyclic voltammetry in 10-60 minutes under $5; Faradaic efficiency on a small electrode in 2-6 hours; a 25 cm² membrane assembly for 100 hours over 4-7 days; 1,000-hour durability at 100 cm² holding a stand for six weeks; then an 8,760-hour qualification occupying one stand for a year. The throughput bottleneck is the gas chromatography and mass spectrometry product line, which every cell needs continuously.

Both are kept as reference designs for facility shape. Neither is a domain recommendation.

Enforced by: intent, and by D9's framework work failing without it.

D3 — The 10x claim is decisions per unit time

Decided. The capability claim is experiments that reach a decision — throughput × yield × the fraction producing an attributable, trustworthy measurement. Samples per day is the wrong unit.

Anyone can buy more liquid handlers, so samples per day measures a purchase. Real facilities lose most of their nominal throughput to samples that are made and never characterized, or characterized and never correctly attributed. That loss is what provenance actually fixes.

Enforced by: intent, for the same reason as D1 — the showcase that would carry the metric is the facility showcase, and it does not exist yet. The obligation this decision creates: when it ships, the metric appears with a stated baseline, and a claim without a denominator does not ship at all.

D4 — Scale invariance: the small case stays first-class

Decided, and load-bearing. Facility work must not tax the one-bench case.

The test for any new feature: does it make the one-bench case better, or merely leave it unharmed? A feature that only helps at scale is usually an accident of deployment promoted into the domain model. All four planned additions in D9 pass this test; anything that fails it should be redesigned. Gating it behind a flag hides the problem.

The hard constraint: tier 1 must never require anything from tier 4.

Tier	Storage	Process	Manifest
Laptop	SQLite	in-process	~15 lines, `default_effect: allow`
Bench	SQLite	in-process	adapters, real policy, campaigns
Cell	SQLite	in-process	resources, twin, viewer
Facility	PostgreSQL	multi-process	stations, long-latency work, human stations

Enforced by: tests/test_minimal_laboratory.py, which pins the size of the smallest working manifest and fails when the minimum grows. Growth in the minimum is the observable form of this decision being violated, and a new required field is a diff anyone can see.

D5 — The showcase proves the chain

Decided. The previous showcase proved a loop closes. That is no longer interesting. This one proves the chain from discovery to a sellable asset, in three visible beats:

the loop finds something a grid search or a human would not have, with the search trace visible
the evidence store becomes a qualification package automatically — including the failed runs, because "what else did you try" is exactly the question provenance answers
the customer receives that package, and the material comes with it

Beats 2 and 3 are the differentiation. Nobody is showing them.

Enforced by: intent.

D6 — Target technology domain

Open. Seven candidates across two shapes have been modelled and falsified. Each failed on arithmetic. The negative results are recorded because rediscovering them would cost months, and because a single mechanism accounts for all seven.

Falsified: a heavy-rare-earth-free permanent magnet venture. The thesis was that grain-boundary engineering could reach high-temperature coercivity without dysprosium or terbium, removing the largest cost in a Western magnet. The physics is sound — Nd2Fe14B allows about 7.6 T and commercial magnets deliver 1-2 T, so 70-85% of the bound is unrealised and the deficit is microstructural. The economics are not. Eliminating heavy rare earth is worth $9.50/kg on a thin part and $31.92/kg on a thick one. The first pass assumed $40-75/kg because it priced bulk alloying; grain-boundary diffusion already uses a fraction of the terbium. Nobody has ever published the cost of being heavy-rare-earth-free, and at a $10/kg process premium the thin-part case goes negative. Probability weighted, a Series A returns 2.24x at 10.6% IRR; even the upside case returns 10.1x, which clears a growth fund's bar and fails a seed fund's.

Falsified: a recycling-fed rare-earth separation venture built on a superior extractant. The thesis was that separation factor sets stage count, stage count sets plant capital, and so a better extractant buys a cheaper plant. The first link holds and was validated against a patent running identical duty on two extractants, agreeing within 10%. The second link does not. Solvent extraction is 18-30% of plant capital and only part of that scales with stage count; 86% of per-stage volume is settler, and coalescence sizes it. The whole capital value of raising the separation factor from 1.5 to 5 is about $3/kg of oxide, 1.3% of revenue — against feedstock payable moving the same model by $52.6M and the export-control premium by $43.0M. Three further errors compounded it: the stage-count curve is concave and the target sat on its flat part; the binding split for magnet feedstock is non-adjacent and already runs at high selectivity; and cheaper levers exist on the same cost line with no chemistry risk, one patent cutting a cascade 37% using the same extractant by moving the reflux point. Returns 1.16-1.50x at 2.1-6.0% IRR.

Falsified: a step-change adsorbent for para-xylene separation. The thesis was that an adsorbent with selectivity near 70, against an incumbent barium-exchanged faujasite at 2-3, would simplify or replace simulated-moving-bed separation across the ~54 Mt/a of world capacity that runs it. Twelve agents across two independent passes attacked it from the cost side and the physics side. It fails four separate ways, and any one of them is sufficient.

The value saturates, and the saturation is algebraic. The separation unit is 11.4% of an aromatics complex, and chambers plus adsorbent charge — everything selectivity can touch — are 42% of that unit. The remainder is desorbent-recovery distillation, whose size is floored by the back-off from the equilibrium-theory vertex and is independent of selectivity. The fraction of the total available prize captured at any selectivity reduces to 1 − α_ref/α, in which every calibration constant cancels. At selectivity 10 that is 75%; at 70 it is 96.4%. The whole 2.5 → 70 step is worth 2.4-4.1% of project capital and $21-26 per tonne of product. The elasticity of capital cost with respect to selectivity at 70 is −0.009.

The binding pair is physically inaccessible to the mechanism. Ethylbenzene, at 14-17% of feed, is the hard rejection, and para-xylene and ethylbenzene have minimum cross-sections of 6.51 and 6.53 Å. Every steric or kinetic sieving mechanism capable of producing high selectivity is blind to that difference. The 2025 record material measures EB/PX = 2.79 on a real four-component feed — it prefers ethylbenzene. Ethylbenzene boils at 136.2 °C against para-xylene at 138.35 °C, so distillation cannot remove it downstream.

The headline number has no provenance. Three independent search passes found no peer-reviewed liquid-phase para/meta selectivity near 70. The best documented values are 4.99, ~17, 27.5 by vapour-phase IAST, and ~30. Everything above that is para/ortho — the easy pair — or a binary measurement. A selectivity of 70 measured at 298 K and driven by adsorption enthalpy decays to ~17 at the 180 °C operating temperature. And the incumbent's own equilibrium selectivity was 3.35 in 1972 and 3.26 in 2021: fifty years in which every gain came from capacity, mass transfer, particle size, binder elimination and architecture.

The flowsheet the thesis implies was already built, at the incumbent selectivity. UOP commercialised HySorb XP in 1998 — single chamber, light desorbent, single-stage crystallisation — and published that it "does not provide any cost or performance advantages relative to the Parex process." Sinopec took ~20% off investment cost in 2022 with a 16-bed single column, and commercialised temperature-swing SMB in 2019, both at conventional selectivity. The two largest capital reductions ever booked in this unit were won by people who own the architecture.

And a thermodynamic ceiling sits above all of it. A perfect adsorbent — infinite selectivity, 100% recovery — shrinks circulation to the separator from 4.368 to 4.237 tonnes per tonne of product, a 3.0% reduction, because equilibrated C8 aromatics are ~23% para and that is a property of the molecule. The isomerisation recycle loop is mandatory at every selectivity.

Returns, probability-weighted over ten years: 2.03x selling adsorbent, 2.12x owning plants, 4.32x for a staged hybrid. Base-case IRRs are 5.6%, 10.2% and 15.1%. Owning plants carries the same expected multiple as selling beds, at four times the capital and a 50% probability of total loss.

Generation 1's diagnosis. All three were a better material for a mature industrial process. In a mature process the material axis is the cheapest axis to optimise, so incumbents harvested it first, over decades, and what remains there is the saturated tail. The shape was changed to require that the process does not yet exist, so that material and process are designed together.

Generation 2 followed that instruction and produced four more falsifications.

Candidate	Materials axis at its physical ceiling	The architecture that beat it, with no discovery
Elastocaloric cooling	+25-45% device COP; closes 4-13% of the gap to vapour compression	R-290 propane at $2/kg, +3% COP, full regulatory compliance
CO2 or CO to ethylene	thermodynamic floor of 13,183 kWh/t	the BASF/SABIC/Linde electric cracker, running since April 2024 at 1,875-4,286 kWh/t
On-site hydrogen peroxide	$40-99/t	Solvay deleting one distillation column, worth $500-1,000/t
Direct propylene epoxidation	11.6% cost share on the proponents' own arithmetic	Sumitomo's process, already cheaper on those same prices

The elastocaloric case is the sharpest. Take the material to both of its physical ceilings — zero hysteresis and transformation stress at the superelastic floor — and device coefficient of performance improves 25-45%. Parity with vapour compression requires 355-614%. Between 58 and 73% of the loss is pumps, friction, regenerator and heat leak. It is a mechanical engineering programme with a materials input.

The unified diagnosis, which is the actual result of five months of this.

In all seven candidates, a process-architecture change on the same cost line beat the materials discovery — including in the two where the process genuinely did not yet exist.

The mechanism is structural. Value that takes the form of reducing a loss is bounded below by zero loss, so it always saturates. The incumbent has owned that cost line for decades and can move architecture far more cheaply than materials, so the cheap part is already gone. Four of the seven show the value function explicitly as 1 - x_ref/x; a fifth shows it as the algebraically identical reciprocal sum 1/V = Σ 1/v_i.

Generation 2's instruction was necessary and insufficient, because it named the wrong noun. A new process delivering an old product inherits the old product's saturated cost line intact. Every one of the four obeyed the instruction and every one still chose a domain where the product already exists at industrial scale: cooling at 135 million units a year, ethylene at 200 Mt/y, hydrogen peroxide at 6.11 Mt/y, propylene oxide at 10 Mt/y.

The shape for generation 3:

The discoverable property must gate whether a product or capability exists at all, and not how efficiently an existing one is delivered.

Two admissible criteria remain. (a) The product cannot currently be made at all, meaning there is no incumbent article to price against — a product that is merely expensive or dirty does not qualify. (d) A named buyer with a named budget and a named deadline needs something for which no supply route exists; if the buyer cannot be named, it is not criterion (d). Two criteria are retired: "the incumbent route is indirect" admits loss-reduction plays by construction and produced three of generation 2's four failures, and "a capability recently became cheap" describes the search rather than the value.

One tension has to be resolved explicitly, because leaving it open is what produced the elastocaloric scissors — 92,148 t/y of demand in a market closed by propane, against 9-230 t/y in the market that is open. If the product does not exist there is no world tonnage, so the tonnage gate applies to a named substitution or new-demand pool with a named buyer, never to current sales.

The honest case against this whole approach, which is on the record because it may be right. A screen that rejects everything is indistinguishable from a broken screen. These gates are calibrated on chemicals and energy, where incumbents have had fifty to a hundred years on every cost line that exists; applied to any mature industrial cost line they will reject everything, because that is what mature industrial cost lines look like. The screen may be correctly reporting that no materials venture of any shape is fundable against a mature industrial cost line — in which case the search space has to change, and a fourth shape would not help. Three further objections stand unanswered: seven of seven were killed by a prosecution with no symmetric defence, and asymmetric adversarial processes converge on rejection regardless of merit; every kill is a median-case kill, and a screen computing expected value on medians rejects every venture ever funded; and the exclusion of biology, in force for all seven, may have removed the answer from the search space before the search began, since a large share of genuine "cannot be made at all" opportunities now sit at the materials and biology interface.

Generation 3 therefore adds a defence agent arguing each candidate as forcefully as the prosecution attacks it, with an adjudication between them.

Generation 3 ran, and it is the most informative of the three. Fifty-eight candidates, and the value-function distribution is the headline: 49 were loss reduction, under an instruction that said in bold that the product must not exist. Seven were thresholds, one multiplicative, one unclear. The shape instruction was followed and the generators still produced loss reduction six times out of seven, which says the shape is what materials opportunities mostly are.

Zero passed the screen. Three were carried forward anyway so the failure would be legible: a ²²⁹Th-doped nuclear-clock host, diamond with deterministically placed group-IV colour centres, and integrated electronics running above 400 °C.

The symmetric defence changed the answer, and that is a finding about the method. Adjudicating each prosecution against a defence reading the same dossier found eleven load-bearing prosecution claims wrong, every one of them in the direction of a kill. A thermal ceiling built on a strawman spec; an opacity ceiling that the field had already moved 13×; a screen-sign argument contradicted by its own stated optimum; a "NOT FOUND" for a spinless host that is the title of a 2025 preprint; and a missed $360M acquisition of the named buyer. Asymmetric adversarial processes converge on rejection, which is why the defence was added. It worked, and it means the seven earlier kills carry a bias of unknown size in the same direction.

One candidate survived to a decision programme, and then died for free. Diamond group-IV centres cleared TEST 0 under adversarial reading, cleared the sign check, passed the value-function form test, and had a cleared market price for the capability. It was costed as a staged programme: $480,000 to $700,000 over 12 to 15 months, four pre-stated kill gates, no new capital. Stage 0 of that programme cost $0 and was one week of reading, with the gate written in advance: if six centres lie within 72 MHz, or any N ≥ 6 simultaneous is documented, stop.

Abe et al., Advanced Functional Materials (arXiv:2505.00576), section 2.3:

Figure 5(a) shows a CFM image of Sample 2300 obtained with resonant scanning at a specific frequency, where several bright spots are simultaneously observed... Six PbV centers marked in circle show PLE spectra with similar center photon frequencies and narrow linewidths of 38-43 MHz. Notably, two spectra have a very small energy difference of only 1 MHz.

Six group-IV centres, simultaneously resonant at one laser frequency, transform-limited, with an inhomogeneous distribution standard deviation of 4.9-5.4 GHz after annealing at 2200-2300 °C. The gate fires on both clauses. Elapsed cost to fire it: twenty minutes and one paper.

One caveat kept with it: the candidate also required deterministic placement, and Abe et al. used ion implantation with annealing, which is not deterministic. A reading in which the full article still does not exist is defensible. The gate was pre-stated and unconditional, so it governs, and the lesson concerns the gate and not diamond: the cheapest question in a decision programme should be asked before the programme is funded, and it almost never is.

What generation 3 changed about the diagnosis. In generations 1 and 2 the kill was always architecture beat materials on the same cost line. That held for only one of three here. The other two died on the demand line: no buyer for the article at the specification where the article would be new, or a buyer who acquired the capability outright fifteen months ago while the merchant slot is held by the incumbent producer.

The binding constraint is the merchant channel. In every field where a threshold article is genuinely new, the buyer is either absent or vertically integrated.

And a selection artefact worth naming: all ten candidates to date were components inside someone else's system. That is a property of how they were generated, and not a property of materials science. Consumables where the material is the article — reagents, catalysts by the tonne, membranes, sorbents, isotopes, calibration standards — were never generated at all.

The retarget the adjudication recommends, with its own caveat attached. Regulatory substitution: PFAS restriction, the fluorinated-gas phase-down, lead-free extensions. The value function is a genuine threshold in a genuinely discrete variable, because the article is legal above the compliance line and does not exist below it. The demand pool is a named substitution pool with a legislated deadline, the tonnage is the incumbent's current tonnage and is therefore already measured, and the buyer is a user, so vertical integration by a chemicals company is structurally harder. Qualification cost builds a merchant moat where elsewhere it closes one.

The caveat, from the adjudication and worth quoting because it is the thing to stress-test: regulatory deadlines slip, exemptions are granted, the incumbent usually wins the replacement because it owns the qualification relationship — and "architecture cannot cross a molecular ban" is weaker than it sounds, because architecture crosses it by eliminating the function, which is the seven-of-seven pattern wearing a new costume.

The filter is fixed even though the answer is not:

combinatorial and not already solved by simulation or ML on existing data
loop-closable — real cost and wall-clock numbers per experiment
characterization automatable, by a named instrument
a named customer with a deadline
a survivable qualification timeline
legible — a non-expert can see success

Added by D2: the domain must have genuine multi-step heterogeneity and a fast/slow split. A domain where one instrument does everything is now a worse fit, because it would not exercise facility-scale capability.

Enforced by: tests/test_domain_proposal.py, whose third check reads this entry directly: once D6 stops reading as open it must link to a proposal file that exists and that the screen collects. That covers how a choice is recorded. Which choice is right is enforced by intent, and by D12.

The disqualifying risk to check before committing: does the fast screen actually predict the slow truth? A fast proxy that does not predict the real property makes an autonomous lab a machine for generating confident error at speed. If the answer is no, the domain is rejected however good the science.

D11 — Capital intensity caps venture returns

Decided. It constrains D6 and D7; it does not follow from them.

Two independently modelled full-stack materials ventures returned 2.24x and 1.16-1.50x. The pattern is structural: owning the plant means owning the capex, and capex compresses the multiple. A materials manufacturer can be right about the science, execute well, and still hand the investor who took the technical risk an outcome that suits a growth fund and misses a seed fund's bar entirely.

The consequence for this programme is uncomfortable and should not be smoothed: the instinct toward vertical integration and the requirement for venture-scale returns pull in opposite directions. Integration is the right answer for value capture and the wrong answer for multiple. Anything proposed under D6 must be tested against this before the science is evaluated, because a domain that fails here fails regardless of how good the chemistry is.

Amended after the para-xylene work, which supplied the missing test. The earlier form of this decision said only that capital intensity compresses the multiple. That is too blunt, because owning capital the discovery makes uniquely cheap is the entire point of a discovery. The test is a comparison of two numbers on the same denominator:

Own the plant when the discovery's value per unit of output exceeds the capital charge per unit of output.

For para-xylene the discovery is worth $21-26 per tonne at its theoretical maximum, and the capital charge on the tonne that captures it is $82 per tonne. It fails by a factor of three to four, and the industry it would be bought into runs a four-year mean margin of $304.81/t against a $332/t full-cost breakeven, with sector ROIC of 3.72% against a 6.22% cost of capital.

The compression is also priced, and the price is public. Chemical (Specialty) trades at 13.36x and Chemical (Basic) at 8.57x (Damodaran, January 2026). Forward integration therefore costs 36% of enterprise value per dollar of EBITDA and 6.20 points of economic profit, before any operating result. That is the cost of the relabelling alone.

Enforced by: tests/test_domain_proposal.py, which requires a domain proposal to carry a capital intensity section stating the two numbers above. Whether the numbers are honest is still intent; whether they were computed at all is now checked. Three modelled ventures exist as comparables.

Decided, and it is the cheapest thing in this document.

Three deep-research programmes each ran to completion and each died on the same question, asked at the end when it could have been asked at the start: what fraction of the paying customer's cost does the discoverable property actually control?

Candidate	The discoverable property	What it controls
Permanent magnets	heavy-rare-earth-free coercivity	$9.50-31.92/kg of a magnet
Rare-earth separation	extractant separation factor	$3/kg of oxide, 1.3% of revenue
Para-xylene adsorbents	adsorbent selectivity	2.4-4.1% of project capital, $21-26/t

Each number is estimable to an order of magnitude in a few hours from public cost structure, well before any physics is evaluated. Each would have stopped the programme.

The screen: compute the controlled cost share first, and reject below roughly 15%. A property that governs under a sixth of the payer's cost cannot carry a venture outcome, however good the science is and however large the market. Market size does not rescue it, because the share is multiplicative with the market and the arithmetic is the same at every scale.

Four more tests, added after generation 2, in ascending order of cost. Run them in this order and stop at the first failure. All seven falsified candidates would have died in under a day, and five of them in under an hour.

Test 0 — the value function. Free. Can the customer buy an article today that performs the same function, in any form, at any price? If so the value is loss reduction, it saturates, and the candidate is rejected. Write the value function down. The form V = C - K/x, or a reciprocal sum 1/V = Σ 1/v_i, is a rejection. The property must enter as a threshold — the article exists above the line and does not exist below it — or multiplicatively into a quantity nobody currently sells.

Test 1 — floor-limited cost. One hour. When the dominant cost line is a purchased commodity whose coefficient has a thermodynamic floor, compute floor × price against product price before anything else. This alone kills a route that is underwater at its own physics limit, and it would have ended the ethylene candidate before any literature review.

Test 2 — physics-limit sufficiency. R ≥ 3. Take the property to its physical ceiling and divide the improvement it delivers by the improvement required for parity with the best architecture available to the incumbent. Current practice is the wrong benchmark, because the incumbent will upgrade. The threshold is 3, because every prior candidate near or below 1 lost to an architecture change, and the margin has to absorb the pioneer-plant penalty: 51% of pioneer process plants never reach 85% of design capacity, and class-2 estimates come in 1.28× over.

Test 3 — architecture ratio. A < 0.3. Divide the value of the best zero-discovery flowsheet change by the value of the materials axis at its ceiling. This test has caught seven of seven, which makes it the highest-yield screen in the set for the time it takes. Search the flowsheet literature: debottlenecking, modular plants, electrified heat, an eliminated unit operation, and the licensor patent record.

Test 4 — capital benchmarked from outside the field. Half a day. Never price capital from a field's own techno-economic analyses, because every field's analyses carry its own aspirations. Price against the nearest mature technology's demonstrated installed cost. One candidate's literature assumed $68/kW against a roughly $2,000/kW demonstrated baseline for a strictly harder device — a factor of thirty, hiding inside peer review.

Two defects in these tests, found by running them. Both are correctable, and neither retires the test.

A is only defined on candidates test 0 has already rejected. The architecture ratio divides the best zero-discovery change by the materials axis at its ceiling. If no incumbent article exists, the architecture route has not delivered one either, the denominator diverges, and the answer depends entirely on the currency chosen — on one generation-3 candidate A ranged over a factor of nine between two defensible currencies. A and test 0 are not independent gates and must not be scored as though they were. Every A computed across three screens was assumed. Use A where an incumbent article exists, which is exactly where test 0 has already returned a rejection, so treat it as a sharpener on the loss-reduction verdict.

R degenerates on threshold candidates. R measures headroom to a physics limit, and it is computable only where a thermodynamic floor bounds performance — which is the chemicals-and-energy setting it was calibrated on. Where the impairment is extrinsic, the limit is zero and R is infinite trivially. Where the binding axes have no physical ceiling at all, R is undefined. On two of three generation-3 candidates R had to be re-specified as headroom on demonstrated levers to be computable, and that re-specification is the fix: divide the improvement available from levers somebody has actually demonstrated by the improvement the article requires.

A third rule, added because it changed an answer. Attack and defence must be symmetric. Running a prosecution against each candidate with no counterpart produced eleven wrong load-bearing claims in a single adjudication, every one in the direction of a kill. An asymmetric adversarial process converges on rejection regardless of merit, so a candidate that survives one is carrying no information. Argue both sides from the same dossier and adjudicate between them.

And the cheapest rule of all. When a decision programme is costed, look for the gate that costs nothing and run it before funding anything. Generation 3's surviving candidate came with a $480,000-to-$700,000 staged programme whose stage 0 was one week of reading, and that stage killed it in twenty minutes. The free gate is almost never run first, because reading a paper feels like preparation and a funded programme feels like progress.

The related trap, which the para-xylene work names precisely: the number the literature reports is often not the number that sets cost. The field reports gravimetric equilibrium selectivity on powder against a binary feed. Plant cost is set by volumetric working capacity and mass-transfer rate on a formed bead in four-component feed with a desorbent present. One material bought 5x selectivity and paid 39% of its capacity; a nine-formulation dataset spans 55.3 to 88.5 wt% selectivity with a 2.45x throughput spread that tracks macropore volume, and the most selective sample is not the fastest. Confirm that the reported measurement is the one that governs cost before treating a literature record as a prize. Check the sign of the correlation between the cheap assay and the expensive truth in the variable that must change at scale. A correlation that merely exists is insufficient. One generation-2 candidate had a screen that was worse than useless: selectivity fell with the pressure the commercial plant would require, so the bench result pointed the wrong way.

Enforced by: tests/test_domain_proposal.py. A domain proposal is one Markdown file in docs/development/domains/, and the suite fails if it omits any screen section, states the controlled cost share without a percentage, or reports a share below 15%. The para-xylene candidate at 4.1% would have been rejected by the second check.

A convention-based check has an obvious hole: write the proposal somewhere else and the screen never runs. That hole is closed at the one place it matters. A third check reads D6 itself, and if this entry stops reading as open it must link to a proposal file that exists — so a domain choice cannot be recorded in the decision log without a document the screen has already been applied to.

D13 — The facility's first deliverable is the predictive bridge

Decided. It came from the two facility architectures; neither venture survived to contribute it.

Both domains examined in enough depth to answer the question gave the same answer, and it was the answer that disqualifies a naive facility:

The fast screen does not predict the slow truth, and no validated accelerated protocol exists.

For elastocaloric alloys, differential scanning calorimetry genuinely predicts adiabatic temperature change, and an infrared self-heating method genuinely predicts structural fatigue to 10⁷ cycles. Neither predicts functional fatigue, which is what actually kills a regenerator, and single-cycle adiabatic temperature change overstates the stabilised value by 1.5 to 6 times (23-38.5 K measured against 6.6-15.9 K after fatigue). The field's fastest screen also measures the wrong object: the celebrated 10⁷-cycle result is a sputtered 20 µm film, and eleven years later no bulk material reproduces that fatigue life at that temperature change.

For CO2 electrolysis, a one-hour flow-cell efficiency of 75.9% does not predict a 1,020-hour membrane-assembly efficiency of 68% at one sixth the current density. Every degradation mode — salt precipitation, flooding, copper reconstruction, ionomer and membrane decay — is slow and invisible in an hour. The published literature on accelerated testing for this device is one position paper from 2020 and nothing standardised since.

A facility that scales the fast screen without the bridge is a machine for generating confident error at speed. D6 already lists this as a disqualifying risk for a domain. D13 states the consequence for the build: the bridge is the facility's first deliverable, before any discovery campaign runs.

Three things follow.

The bridge is measurement science. It is bounded, schedulable and estimable in a way discovery is not. The elastocaloric version is a bulk-coupon assay under 10⁴ cycles calibrated against 10⁷-cycle behaviour on a formed article; at 20 Hz a 10⁷-cycle point takes 5.8 days, so a 40-point calibration set is about eight months on four frames or two months on twenty.
The bridge is a saleable asset on its own. It is transferable, it serves every participant in the field including competitors, and it needs no plant, no offtake and no first-of-a-kind financing. D7's qualification-dossier hypothesis is the same idea seen from the commercial side.
It changes what the facility is claimed to do. The claim "we search faster" is weaker and easier to attack than the claim "we built the measurement that makes searching mean anything," and only the second is defensible when the field's own fast screen points the wrong way.

One uncomfortable finding to keep with it. Model contribution is largest where it is least needed. On cheap axes — transformation temperature, latent heat, hysteresis — models do well. On the expensive axis that decides the outcome, fatigue life scatters by an order of magnitude at fixed nominal condition, which sits at the bad end of the 2.3-3.5x noise penalty, and each point costs 6 to 58 days. So the realistic acceleration is near the bottom of the published 1.25-6x band on the axis that matters. State that to an investor plainly, because it is the first thing a good one will find.

Enforced by: intent, and by D6's existing requirement that a domain answer the fast-screen-predicts-slow-truth question before adoption. tests/test_domain_proposal.py requires the answer to be present and to state the sign of the correlation.

D14 — The twin renders, the adapter acts

Decided, and the boundary is now enforced.

The digital twin is a projection. Task events reach the evidence store, a deterministic projector turns them into visual cues, and the viewer renders those. The architecture page already states the consequence: the visual source "cannot create a capability, commission an instrument, or report physical state. The viewer has no equipment-command endpoint."

That property is the twin's entire value. A twin that can be written to is an assertion about the laboratory; a twin that can only be projected into is a record of it.

The pressure to break this arrives disguised as a demo. An interactive scene, a control that moves the arm from the viewer, a simulated world wired into the rendering layer because that is where the geometry already lives. Each is reasonable in isolation and each ends the guarantee.

Where a simulated world goes instead. Behind a capability adapter, where policy, leases, retry safety, typed errors and provenance all apply to it exactly as they would to steel. The write path and the read path stay separate and can share one geometry source:

	direction	what it is
the world	write	an adapter. an agent acts, physics responds, events record
the twin	read	a projection of what the store recorded

This matters more now than when the twin was built, because adapters/simulated-lab acts on a dictionary in process and its own source says so: "Nothing is aspirated, moved, or heated." The comment goes on to record that a real adapter would declare NOT_REPEATABLE for mixing and at most REPEATABLE_IF_NOT_DISPATCHED for moving labware, and warns the reader not to copy its line. The framework's hardest property cannot be exercised honestly today, and a physics-simulated world behind an adapter is what fixes that. It belongs in the adapter layer regardless of how tempting the twin's geometry is.

On device-level standards. Anthropic and HHMI Janelia have published the Model Hardware Standard as a limited research preview: read and write primitives, device discovery, and natural-language tags carrying weight, safety limits and capabilities, reachable over MCP, a command line, or code. It is application-gated and not yet open source, so the specification cannot be read today.

The layering is nevertheless clear, and adapters/AGENTS.md already anticipates it with "declare semantic capabilities, not raw transport commands." A device standard is the raw transport. An OpenSDL capability is the semantics: what the operation means, what it may touch, whether policy permits it, and what evidence it leaves. When a device-level standard is readable, OpenSDL speaks it through the adapter layer and does not compete with it. Nothing about the twin changes.

Enforced by: tests/test_twin_is_read_only.py. The first check fails if any route under /twin or /viewer accepts a mutating method. The second pins the command surface to three declared routes so that adding a fourth is a deliberate act a reviewer has to approve. Both were verified against an injected POST /twin/move, which they reject by name.

D7 — Business model

Shape decided. The specifics wait on D6, and only the specifics. The three falsified candidates were modelled far enough to answer the structural question, and the answer held across all of them, so it is recorded now.

The four paths, costed side by side on the third candidate:

Path	Company capital	Expected MOIC	Base-case IRR	Downside
Sell the consumable	$80-155M	2.03x	5.6%	0.26x
Consumable under instrumented performance contracts	$200-250M	2.64x	7.6%	0.05x
Build and own plants	$500-620M	2.12x	10.2%	zero, at 50% probability
Staged hybrid	~$126M	4.32x	15.1%	1.08x

Owning plants and selling the consumable have the same expected multiple, and owning demands four times the capital, six more points of dilution, two more years to revenue, and a downside of zero. The hybrid wins by roughly 2x on expected value, and it wins in the bear case. Its bull case is only 1.5x the plant-owner's bull case; its failure costs $15M at month 18 where the plant-owner's costs $620M at year 10, and the $15M comes back and is redeployable.

Four rules follow, and none of them depends on which domain D6 lands on:

Keep the consumable. Licence-only is a weak position and the evidence is specific: GTC Technology — a genuine aromatics licensor, its process in more than sixty units, thirty years of development, around two hundred staff, roughly $50M of revenue — sold to Sulzer in May 2019 at a $39M enterprise value. The recurring physical product is where the margin lives.
Buy the reference plant with intellectual property. A first commercial reference is the artefact that cannot be bought and without which nothing else is financeable. Contribute licence and materials for carried equity in exactly one unit built on someone else's balance sheet. Model the dilution the entry percentage will suffer: LanzaTech contributed IP for 30% of the Shougang joint venture in 2011 and was diluted to 8.38% as others funded the build.
Prove inside a host's existing facility first. A capacity revamp borrows the host's feedstock, utilities, operators, tankage and offtake, and so has no minimum-scale problem — which is the trap that otherwise binds, because the scale at which a first plant is financeable sits below the scale at which it works standalone.
Write the hard stop into the financing documents before the work starts. Pre-commit the numeric gates and the decision to stop if they fail. Sunk-cost pressure at the point of committing engineering capital is what actually kills these companies, and a gate agreed afterwards is not a gate.

Retained from the earlier draft, still untested: tamper-evident provenance produces a qualification dossier as a byproduct, which in a regulatory-replacement market is itself a priced deliverable. That reframes the offer from "we found a material" to "we found it and can prove how, in a form a qualifying authority accepts." The second half is the defensible half.

Enforced by: intent for the four rules, and by tests/test_domain_proposal.py for the inputs they consume. D11's ownership test decides any proposal to own capital; D12's cost-share screen runs before it.

D8 — Scheduler architecture

Open, and the sharpest tension in D4. Facility scale wants a persistent process routing work across stations. The laptop case wants opensdl run workflow.yaml with no daemon, no broker, and nothing running in the background.

Intended resolution: scheduling becomes a strategy behind an interface with an in-process immediate-dispatch default, in the same shape as the storage layer, where repository interfaces make PostgreSQL a swap. Named here because "we will just add a small daemon" is exactly how the small case dies.

A concrete defect found while designing the facility work, recorded here because it is a scheduler question and not a bug fix. In packages/runtime/src/opensdl_runtime/engine.py the concurrency semaphore is acquired at line 652, outside _hold_resources at line 708, so a task waits for a slot and only then contends for its resources. Reordering the two looks obviously correct and is not: leases taken first means a task holding a scarce instrument idles behind unrelated tasks holding the concurrency slots, and it removes the only bound on concurrent resource pollers. Which order is right depends on how scheduling is resolved, so it waits for this decision rather than being half-fixed under it.

Enforced by: tests/test_minimal_laboratory.py, which refuses a manifest containing scheduler: or broker: and fails if the minimum grows past 20 lines. That catches a daemon becoming required. It would not catch a daemon becoming the default while staying omissible, which is the subtler version of the same failure and is currently enforced by intent.

D9 — Framework work the facility requires

Decided in shape, unscheduled. Four additions, each of which must satisfy D4:

Long-latency capabilities. A chamber that legitimately answers in three weeks is not a hung instrument. Today a timeout on a non-repeatable capability lands in intervention_required, which is right for a hung mixer and wrong for a chamber doing its job. "No answer yet" needs to be a declared, normal state. At one bench this fixes the overnight anneal.
Late-arriving observations. An observation must be foldable into a campaign whose run finished long ago and which has since issued hundreds of proposals. At one bench this is a sample mailed out for external analysis.
Multi-fidelity as a first-class concept. A fast screen and a slow truth are different measurements of one property, with different cost and trust, and the optimizer must know which it received. The triage policy — which sample earns the expensive measurement — is where the intellectual value sits and is the part a competitor cannot buy. At one bench this is a cheap check before an expensive one.
Leases across long durations. A chamber slot held for weeks stresses lease TTL, reconciliation, and controller restart mid-hold.
Partial observations that arrive before the measurement finishes. A capability must be able to emit a predicted result with a stated uncertainty from incomplete data, and then revise it when the measurement completes. The optimizer must know it received a prediction, act on it, and accept the correction without invalidating the campaign. At one bench this is reading a trend after one night of a three-night run.

Correction, and it reorders the list. Items 1, 2 and 4 all treat a slow measurement as a fixed duration to schedule around. That is the smaller win. Attia et al. (Nature 578, 397-402, 2020) closed a loop on fast-charging protocols by shortening the measurement itself: a model predicting cycle life from the first 100 cycles replaced cycling to failure, and the combined early-prediction and closed-loop system evaluated 224 protocols in 16 days against roughly 500 days for the exhaustive alternative. Pipelining around a three-week measurement recovers throughput. Truncating a three-week measurement to three days recovers an order of magnitude more, and it compounds with the pipelining.

So item 5 outranks items 1, 2 and 4, and it changes what the facility is for. The triage policy in item 3 and the truncation model in item 5 are the same asset viewed twice: both decide how much measurement a sample deserves, and both are the part a competitor cannot buy. The framework obligation is that a capability can report progressively, and that provenance records which observations were predictions and which were truth, because a campaign built on predictions that were never corrected is a campaign that cannot be audited.

Enforced by: intent, until the work is scheduled. When it is, each addition ships with simulation and conformance coverage per the architecture rules and demonstrates its one-bench benefit in an example. The requirement item 5 places on the design: progressive results carry an explicit predicted flag through the evidence store, and attestation reports every prediction as either resolved to a measurement or unresolved.

D10 — Dogfooding does not become facility-only

Decided. Building the facility creates pressure for every default, example and document to assume one. The countermeasure is that the small examples stay alive and exercised.

Enforced by: make showcase, which re-derives the discovering-colors campaign in CI so the small reference cannot silently rot while attention is on the facility, and tests/test_minimal_laboratory.py, which pins the minimum manifest at 20 lines and runs real work through it.

Correction, 2026-08-24. This entry claimed the benchmark suite keeps both a small laboratory and a facility laboratory, so that a change hurting small labs would show as a score drop. The facility half does not exist. The suite has three laboratories — color-bench, restricted-bench and stalled-bench — and all three are a single bench.

What is true is the rest of it. tests/end_to_end/test_benchmark_suite.py runs the shipped suite in CI against two scripted control agents, one competent and one that oversteps, so the suite cannot quietly become unpassable or unfailable, and small-laboratory operability is exercised on every change with no key and no spend.

The missing half was premature by construction: a facility laboratory cannot be expressed until D9's station work lands, so it is blocked behind Phase 1 and Phase 2. Until then the comparison this decision wants — small against facility, on the same suite — is unavailable, and the small case is held by the showcase, the minimal-laboratory test, and the benchmark's own controls. When the facility becomes expressible, a facility laboratory joins the three benches and this entry gets the mechanism it was written as though it already had.

Standing rules

Durable constraints on all facility work. These are quoted in AGENTS.md so a fresh session inherits them.

A laboratory with one instrument stays expressible in about fifteen manifest lines, runnable in one process against SQLite, with no scheduler, no broker and no optional service.
Facility features are opt-in by configuration, never by requirement.
A change that lengthens the minimum manifest needs justifying.
Heavy dependencies live behind an extra. pip install opensdl stays small.
Documentation leads with the small case. The quick start never mentions a station.

An accepted consequence: a feature that would be simple given PostgreSQL and a daemon will take longer built this way. That is the correct trade while the small case is the adoption path.

Sequence

Phase boundaries carry the sequence. Each phase ends with something an outsider can check.

Phase 0 — settle the domain. Close D6. D7 is closed. Ends with a proposal in docs/development/domains/ that passes the D12 screen, a named buyer, a loop with real cost and wall-clock numbers, and the D13 answer: what the bridge is, what it costs, and what the sign of the fast-screen correlation is.

Phase 1 — framework work. D9 item 5 first, then items 1 and 2, each with a one-bench example. Item 5 leads because truncating a slow measurement is worth more than scheduling around one. Ends with a campaign that acts on a predicted result, accepts the correction when the measurement lands, and keeps proposing throughout — demonstrated in simulation.

Phase 2 — the facility in simulation. Stations, sample custody, contention, triage — as a manifest and a simulated laboratory, before any hardware. Ends with a facility that runs end to end with no physical equipment, which is also the honest way to size the real one.

Phase 3 — the twin. The Blender scene of the facility, showing work in progress at every stage simultaneously. The multi-timescale story is legible at a glance in a way one plate cannot be.

Phase 4 — the chain. D5 beats 2 and 3: the evidence store emitting a qualification package, failed runs included.

Phase 5 — real hardware, and the bridge first. Per the hardware access research: the first physical instrument, in whatever domain, closing one real loop. Per D13, the first campaign that loop runs is the predictive bridge. Discovery campaigns come after there is evidence the fast screen predicts the slow truth, and with which sign.

Open questions

#	Question	Blocks	Owner
D6	Which technology domain	Everything	research, then owner
—	Is the constraint set what keeps failing? Seven candidates died to the same mechanism. The screen may be correctly reporting that no materials venture of any shape is fundable against a mature industrial cost line, in which case the search space changes and D6's shape does not	D6	owner
—	Whether excluding biology removed the answer before the search began. In force for all seven, and a large share of genuine "cannot be made at all" opportunities sit at the materials and biology interface	D6	owner
D8	Scheduler architecture	Phase 2	design
—	Facility capital and staffing	Phase 2 sizing	research

D7 closed after three ventures were modelled far enough to answer its structural question. D13 answers the fast-screen question generically — the answer was no in both domains examined in depth, which is why the bridge became the first deliverable. It had been an assumption.