The question, asked properly
The naive version of the question is "why does biology need a file format", and it has a naive answer about digitisation and efficiency that is not worth anyone's time. The sharper version is this: what specific information is currently destroyed at the moment two parties agree to make a living thing, and would a portable artefact have preserved it.
That question has an unpleasantly concrete answer. When a laboratory orders an engineered muscle construct, a neural culture on a microelectrode array, or a perfused tissue model from another laboratory, the thing that decides whether the delivered batch is acceptable is almost never written down in a form that survives the transaction. It is discussed on a call. It is implied by a cited paper. It is understood to be "like the last ones". And when the batch arrives and somebody is unhappy, there is no document to point at, because the document was a conversation.
Everything downstream of that failure follows from it. Two suppliers cannot be compared because they are not quoting the same article. A batch cannot be accepted or rejected on evidence because no threshold was agreed in advance. A result cannot be reproduced by a third laboratory because the conditions that produced it were never recorded alongside the number. And a supply chain cannot form, because a supply chain is not a set of relationships — it is a set of interchangeable specifications.
What the PCB industry paid to learn
The nearest complete case study is printed circuit boards, and it is worth studying because the industry there has already run the whole experiment, expensively, over forty years.
Gerber began in 1980 as a photoplotter driver: a subset of EIA RS-274-D, published by Gerber Systems to make their own machines work. It was an image format. It said where copper went. By 1991 it had fragmented into plotter-model-specific dialects, and it took the 1998 RS-274X unification to pull the family back into one thing. In 2014 the X2 revision added attributes, converting an image format into a data-transfer format, and did so backward-compatibly — attributes provably do not change image generation. X3 added component data in 2020. The format survived by being simple, ASCII, easy to implement correctly, and by extending without breaking.
What Gerber never did was carry design intent. The IPC-2581 Consortium's own diagnosis of the problem is blunt: Gerber, properly implemented, is perfectly adequate to transfer image data, but it does not transfer stack-up data, materials, design intent or netlist. So ODB++ appeared in 1995 to carry all of that, and it worked — an eight-layer board that took up to five hours to plan and tool from Gerber took about one hour from ODB++. By 2018 it reported roughly forty-nine thousand users and seventy-one committed software vendors.
And it still lost the argument it most needed to win. In August 2018 the manager of the ODB++ Solutions Alliance had to publish a blog post titled "The Truth about ODB++" to kill a rumour that Siemens intended to take the format private. His defence was entirely true and entirely generous: the specification is free, the viewer is free, the tooling is free, the support is free, and — the load-bearing sentence — "no one would be foolish enough to betray that many supporters."
That is the most instructive sentence in this entire domain. It offers stewardship where the market wanted a structural guarantee. Free of charge is not the same as openly licensed, and the industry knew the difference well enough to fund IPC-2581 as a neutral alternative anyway. Fifteen years of standards politics produced three formats and no winner. Anybody proposing to duplicate that exercise should be made to explain why.
The lessons that transfer are narrow and specific. Ownership of the process triggers the immune response, not ownership of the name: nobody objects to a trademark, everybody objects to a clause requiring written approval before third-party extensions. Being narrow beats being complete, because Gerber outlived two technically superior formats by being implementable in a week. And authority earned in the open — the reference implementation, the conformance suite, the validator, the editorship — survives an open licence completely intact. Give away the veto; keep the work.
What a prose specification loses
Here is the concrete version of the loss. On the left is a real shape of request — the kind of email that currently constitutes a specification for a living component. On the right is the acceptance section of the same order expressed as WWD criteria.
From: pi@lab.example.ac.uk
Subject: Re: muscle constructs — quote?
Hi — following up on the call. We need roughly 20 of the
aligned muscle strips, about 1 cm long, on the loop anchors
like the ones in the 2024 paper. Human iPSC-derived. They
need to actually contract, obviously — the last batch we got
from elsewhere were basically dead on arrival.
Ship live if you can. Happy to discuss.AC-001 median peak twitch force >= 500 uN
method WWD-TWITCH-001, 37 C, 1 Hz field stimulation,
0.5 mN preload, n = 12 random, destructive,
one-sided test at alpha = 0.05
AC-002 twitch force CV <= 0.20
same 12 specimens as AC-001
AC-003 median S2D alignment >= 0.65
OrientationJ 2.0.5, xy plane, 5 fields, 0.25 mm2 each
PC-001 cast gel stiffness 11-13 kPa
rheometry — tests the control axis, not the outcomeThe email is not stupid. It is written by a competent scientist and it contains real information. What it does not contain is any statement that could decide a dispute. "They need to actually contract" has no threshold, no statistic, no sample size and no assay. "The last batch were basically dead on arrival" describes a viability failure without saying at what timepoint viability is measured, by which stain, against what floor. "About 1 cm" is a dimension without a tolerance. "Like the ones in the paper" delegates the specification to a figure caption.
A supplier reading that email must guess. Two suppliers reading it will guess differently, and their quotes will differ by a factor that has nothing to do with their capability. When the batch arrives, the parties discover their disagreement for the first time, after the money is spent and the cells are dead.
The right-hand column costs the buyer perhaps twenty extra minutes and eliminates all of that. Note what it does not do: it does not tell the supplier how to make the tissue. That is deliberate. The acceptance criteria are the contract; the process is the supplier's expertise and the supplier's problem. A format that specified the process would be a bioprinter instruction language, which is a much smaller and much less useful object.
The round trip is the point
Gerber goes one way. You send the file, the board comes back, and the file has nothing more to say. That is acceptable for a process whose deviation is error — noise around a commanded target that a better machine would reduce.
Wetware is not that. Variance is not noise around the specification; variance is the central engineering problem. Two constructs from the same protocol, the same passage and the same donor differ in force by amounts that matter. So the outgoing file is only half the artefact. The same package has to come back populated with what was actually measured, and it has to do so without overwriting what was intended.
WWD handles this with no mode flag at all. As-built is derived, not declared: a package with no measured results is as-designed, and a package with measured results is as-built in addition to still carrying its unchanged design. Every measured result references exactly one acceptance criterion by immutable UUID, and a dangling reference is a hard error rather than a warning. Appending results never mutates a criterion. A retest is a distinct result and never erases the original. Lot disposition is derived from the criteria and the results, never stored as an independent fact that somebody could set by hand.
The immediate benefit is that the file is simultaneously the order form and the certificate of analysis, which removes an entire class of transcription work and an entire class of transcription error. The compounding benefit is larger. Every fulfilled order returns a populated file, and the accumulation of those files is a dataset mapping design to process to supplier to price to measured outcome. That corpus is what eventually produces a quoting engine, then design-for-manufacture rules, then a design model. It is also the asset a competitor cannot reconstruct after the fact, because the information was never captured in their records to begin with.
There is a stronger version of this claim, and it is the reason WWD tags every value with a causal role as well as a provenance. If the file records which quantities were set and which were observed, then each build record is literally one row of a designed experiment rather than one row of a log. Every fabrication run in a tissue laboratory is already a small designed experiment; nobody records it as one, because no format has a slot for the distinction, and the information is destroyed at authoring time and cannot be recovered afterward by any amount of cleverness applied to the resulting pile of files.
A number without a method is not a specification
"500 µN twitch force" reads like a specification and is not one. It omits the three things that would let a second laboratory reproduce the judgement.
It omits the statistic. Is 500 µN the mean, the median, the minimum, or the value of a single representative unit? For a stochastic biological product these are materially different promises, and the difference between a median and a floor is the difference between a normal batch and a batch nobody can supply.
It omits the dispersion. A lot with a median of 520 µN and a coefficient of variation of 0.15 is usable. A lot with the same median and a CV of 0.60 is not, because half the units will fail whatever the downstream experiment needs them to do. Location and dispersion are two criteria in WWD, never one compound statistic, precisely so that a supplier cannot satisfy one by sacrificing the other.
And it omits the assay. Force measured at 37 °C under 1 Hz field stimulation with 0.5 mN preload on a named fixture is not the same quantity as force measured at room temperature under paced electrode stimulation at a different preload. Between two suppliers using different rigs, the numbers are not comparable, and a buyer who treats them as comparable will systematically choose the supplier with the more generous method rather than the better tissue. WWD therefore treats a method-less functional requirement as nonconformant: not discouraged, not warned about — invalid.
The sample size and the decision rule complete it. "Median at least 500 µN, CV no greater than 0.20, n = 12, one-sided test at α = 0.05" is a statement that two parties can disagree about in advance and then settle with data. That is what a specification is for.
What a file format is not
Three disclaimers, because overclaiming here is how formats acquire a bad reputation they never shake off.
A conformant file does not mean a manufacturable design. Structural validity says the document is well-formed, the units are dimensionally consistent, the references resolve and the integrity hashes match. It says nothing about whether the tissue will survive, whether any supplier can build it, or whether the acceptance criteria are physically achievable. Passing a validator must never be described as scientific or manufacturing conformance, and the specification says so normatively.
A file format is not a solver. WWD can carry a solver-neutral model envelope — parameters, boundary conditions, run provenance and results — and a design remains completely valid without any of it. What the format enforces is separation: a simulated value may not be presented as a measured one, and a simulation result can never satisfy an acceptance criterion. The most common way computational biology work becomes untrustworthy is a predicted number quietly acquiring the authority of an observed one.
And a file format is not a moat. This bears repeating because it is the claim most likely to be misread as false modesty. The defensible asset is the validator, the design-for-manufacture rules, the supplier network and the round-trip corpus. The schema is the least valuable part, which is exactly why it can be given away, and giving it away is the only way to acquire the other four.
The governance commitment
Stated as concretely as it can be while it is still only a commitment.
The format is not patented and will not be. Abstract data structures are largely unpatentable in the United States regardless — Alice v. CLS Bank, 35 U.S.C. §101 — and more importantly a patented format does not get adopted. The split that works is free format, protected implementation: the ARM model, where the instruction set documentation is open and the silicon is licensed.
The licences were chosen before any external contribution existed, because retroactive relicensing once outsiders are involved is genuinely miserable. Specification prose is CC BY 4.0 so it can be quoted, translated and reproduced in papers and in other standards. Schemas, validator, examples and tests are Apache-2.0, chosen over a bare permissive licence specifically for the express patent grant and its defensive-termination clause. Registry vocabularies, when published, are CC0, because a controlled vocabulary carrying an attribution requirement simply gets copied without it.
Those licences are already granted and are irrevocable. That is the part that matters,
because an acquirer inherits a licence and does not inherit a promise. The specification,
schemas, validator, conformance suite and examples are all public, so no privileged
implementation holds private knowledge. Nothing in the format requires a service: a
.wwd file is valid, parseable and complete offline, forever, with no call to
anyone's API.
Wetware World would be both the format's author and its largest commercial implementer, and that is precisely the structure the PCB industry rejected. So the intent is to move WWD to neutral custody before anyone has to demand it. The trigger is adoption rather than sentiment: once independent implementations exist from parties with no commercial relationship to us, single-maintainer control has outlived its usefulness. Until that happens the honest description is a single-maintainer draft with irrevocably open licences, and it should be read as exactly that.
Where this argument is weakest
Two places, and both should be said out loud.
The first is that a format nobody uses is a hobby. Everything above is an argument that this artefact ought to exist; none of it is evidence that anybody wants it. There are no adopters, no external implementers and no supplier currently accepting WWD files. The feature-discovery work — sitting with laboratories that build these things and finding out what they actually need represented — has not been done yet. Until it has, parts of the draft are reasoned guesses dressed in normative language, and the normative language should not be mistaken for evidence.
The second is that formats are usually written too early and frozen too soon. The guard against it is procedural rather than clever: 0.1 is explicitly breakable, 0.2 waits for fifty to a hundred real designs to have been pushed through 0.1, and 1.0 waits until the core has survived several suppliers and several tissue types. Registry breadth is deliberately deferred so that adding a property term never requires changing the core model. The architecture decisions were made now because they are cheap now and expensive later. The field-by-field coverage was deferred because it is expensive now and cheap later.
If you build living components and something above is wrong, that is the most valuable message we can receive. The specific request is on the participate page, and it is not a sales question.