⚑ Before anything else — we want your criticism
If you work in aerodynamics, polymer science, adhesion mechanics, or any adjacent field and you read these proposals and think "that won't work, and here's why" — please tell us. We mean it without reservation. Every substantive critique we receive gets fed directly back into the next revision. There is no version of this where pointing out a flaw is unwelcome.
Write to getdragons@dragonworx.bio. Harsh is fine. Specific is better. The fastest way to make this research real is to find what's wrong with it now, on paper, where fixing it costs nothing but a rewrite.
There's a fair worry hanging over any technical document produced with AI assistance in 2026: that it reads well, cites confidently, and falls apart the moment a real expert looks closely. Polished prose with no spine. We take that worry seriously — seriously enough that we'd rather show you exactly how these proposals were stress-tested than ask you to take the polish on faith.
So here is the honest account of how the second revisions came to exist, what the review process caught, and — just as important — what it cannot catch, which is precisely why your eyes on these documents matter more than ours.
The exercise: become the reviewer who wants to reject you
A research proposal lives or dies in the hands of a reviewer whose job is to find the reason it shouldn't proceed. So before sending these out, we ran a deliberate exercise: take each proposal and read it as the most skeptical qualified reviewer in its field would. Not a friendly reader. The one looking for the fatal flaw.
For the DragonSuit, that meant reading as a wind tunnel aerodynamicist — someone who has spent decades watching promising designs fail to survive contact with a real test section. For the GripSuit, it meant reading as a polymer and materials scientist who knows from experience exactly how nano-pillar arrays and bio-mimetic coatings disappoint in the lab. Two completely different skeptics, each with their own catalogue of "I've seen this go wrong before."
The goal was not to defend the proposals. It was to write down, in the reviewer's own voice, the objections most likely to sink them — and then either answer each one or admit it couldn't be answered yet.
What follows are the actual questions that exercise produced — eight for the DragonSuit, nine for the GripSuit. Each one is now printed inside the proposal itself, in a section we titled "Anticipated Technical Questions and Risk Register," written in the reviewer's voice, with our response beneath it. We would rather raise the hard questions ourselves than have them raised for the first time in a rejection.
A note on how this was actually done
You asked to peek under the hood, so here's the candid version. This review wasn't done by a single human expert, and it wasn't done by an AI working alone. It was done by a person directing an AI reasoning process — and the interesting part is what that process is unusually good at, and what it plainly isn't.
// the one real advantage: breadth held in a single pass
A career expert earns their depth by going narrow. The aerodynamicist spends thirty years on flows; the polymer chemist spends thirty years on networks and cure kinetics. That depth is irreplaceable — and it also means the two of them work in different buildings, read different journals, and almost never review the same document on the same afternoon.
A language model's strange gift is the opposite one. It holds the wind tunnel's Reynolds-number problem and the polymer lab's catechol-oxidation problem in the same working context, at the same time, over the same document. It can notice — in one reading — that an aerodynamicist will object to the blockage ratio and that a materials scientist will object to testing adhesion in shear when the field failure is peel. No single human carries both of those reflexes natively. Assembling a panel that does takes weeks of scheduling.
That's the honest value here: not depth that beats an expert, but breadth that surfaces the known objections of many fields in a single sitting — fast enough that self-critique becomes cheap, and cheap enough that you actually do it before you send the document out.
And now the part that matters more, because leaving it out would be exactly the kind of overclaim that earns the "AI slop" label in the first place:
What the process does well
- Surfaces the standard objections that already exist in the published literature of several fields at once.
- Adopts the skeptic's stance on demand — adversarial self-review without ego.
- Holds cross-disciplinary constraints together so a fix in one field doesn't quietly break another.
- Makes rigorous self-critique fast enough to actually happen before submission.
What it cannot do — and why you matter
- It lacks the tacit, hands-on knowledge that comes from running a real rig for twenty years.
- It can be confidently wrong, and won't always flag where.
- It doesn't reliably know which objection matters most in practice — only which ones exist.
- It can miss what isn't in its training. A specialist's instinct still catches things it won't.
The critical oversight in this work is human, and it stays human. The reasoning process proposes objections; a person decides which are real, which are answerable, and which require admitting a limit in writing. That last category is the honest one — the proposals now say, in several places, "this is not yet known, and the program is designed to find out" rather than papering over the gap. An AI that only ever reassures you is worse than useless on work like this. The point of the exercise was to make it argue against us.
A.
The wind tunnel reviewer's questions
DragonSuit · 8 objections
These are the questions an aerodynamicist raises first. Several of them reshaped the experimental design substantially; one of them required walking back a claim in the original draft.
Is this measured at flight Reynolds number, and if not, why should I believe it transfers?
The single most dangerous question. Two of the suit's tricks — the shark-skin riblets and the whale-fin tubercles — behave differently depending on the flow regime. A test run too slow would measure the wrong physics and quietly mislead everyone.
Now addressed: the proposal stops claiming transfer and instead commits to finding the riblet optimum empirically at the real test conditions, reporting tubercle data against measured boundary-layer state, and treating Reynolds-sensitivity itself as a deliverable.
A rigid mannequin's glide ratio is not a flight glide ratio.
A real wingsuit pilot constantly manages their body shape in the air. A stiff dummy in a tunnel can't, so its measured number isn't the same quantity as in-flight performance — and conflating the two is a classic overreach.
Now addressed: we explicitly separate the two. The program measures the repeatable polar of a fixed geometry and states plainly that this is not a claim about what a trained pilot achieves through active control.
Your wing deforms under load — how do you know its shape during the test?
The whole point of the suit is that it changes shape under air pressure. So a force reading alone can't tell you whether the wing reached the shape you designed or some unintended one. You'd be measuring an outcome without knowing its cause.
Now addressed: every force measurement is paired with simultaneous full-field shape measurement (stereo digital image correlation), so each data point becomes a measured shape-and-force pair rather than a guess.
Blockage will dominate a full-suit test in an academic tunnel.
The deployed wingspan is huge — close to two and a half meters. Put something that big in a normal tunnel and it chokes the airflow, corrupting the readings. This went completely unmentioned in the first draft.
Now addressed: the isolated wing panel becomes the primary quantitative test for exactly this reason, with the full-suit configuration gated on blockage staying within correctable limits or being descoped.
The shape-memory polymer won't be in its working state inside an empty mannequin.
The suit's rib structure stays rigid using body heat. A tunnel mannequin has no body in it and gets chilled by the airflow — so the ribs might go soft and the test would measure a suit that isn't behaving as designed.
Now addressed: the test article carries controlled heating with logged temperatures, making the rib's thermal state a controlled variable — and the cold-loss penalty becomes a measurable result in its own right.
You cannot add up lift-and-drag contributions and discount them 30%.
The original draft estimated each technology's benefit separately and summed them with an arbitrary haircut. No aerodynamicist accepts that — the effects interact in tangled, non-linear ways that simple addition can't capture.
Now corrected: the per-technology figures are reframed as isolated directional estimates only. The proposal now states the combined performance must be measured, not summed, because the mechanisms are coupled.
What about unsteady effects, stall hysteresis, and dynamic stall?
Real flight includes gusts and maneuvers. A set of steady-state measurements misses the messier, time-dependent behavior — and stall in particular behaves differently depending on whether you're approaching it or backing out of it.
Now addressed: stall is characterized in both sweep directions to capture hysteresis, and genuine dynamic-stall testing is explicitly named as out-of-scope Phase II work — so the boundary is stated rather than implied.
What is your uncertainty budget?
A serious result carries error bars. A proposal that reports clean numbers with no stated uncertainty signals that the author hasn't thought about measurement error — an instant credibility problem.
Now addressed: every reported coefficient will carry propagated uncertainty from the balance, dynamic pressure, model attitude, and blockage corrections, with repeat runs establishing repeatability. No claim without its error bar.
B.
The materials reviewer's questions
GripSuit · 9 objections
A polymer scientist's skepticism runs along entirely different lines — fabrication reality, statistics, chemistry that fights you. These reshaped all four research threads.
Your fatigue test loads in shear, but these adhesives fail in peel.
A climber's pad is pulled off at an edge — a peeling motion — not slid off uniformly. Testing only the sliding (shear) direction characterizes the wrong failure and would massively overstate how durable the pad is.
Now addressed: the durability test now cycles in both regimes, and the governing peel case sets the longevity claim. The failure-mode census distinguishes the two.
Single coupons tell you nothing — where is the statistics?
Adhesion is inherently scattered; one sample per condition is noise, not data. Without replicates and proper statistics, a fatigue curve is just a story with numbers attached.
Now addressed: a minimum of five independent samples per condition, fit with Weibull statistics to capture both strength and scatter, with confidence intervals on every curve and a power analysis to confirm the sample count.
Carbon-nanotube composites live or die on dispersion — how is it controlled?
Clumped nanotubes are the usual reason these composites underperform. If you don't characterize and control how well they're dispersed, your results scatter for reasons you can't explain.
Now addressed: dispersion is characterized for every batch and reported alongside each mechanical result — plus a documented respirable-fiber safety protocol, which any academic lab will require.
Catechol oxidizes — your mussel-inspired coating will be dead on arrival.
The chemistry that lets mussels stick underwater only works in its reduced form. Exposed to air, it oxidizes and loses its grip on the surface. This is the single hardest problem in the whole mussel-adhesion field, and the first draft waved at it.
Now addressed: synthesis under inert atmosphere and controlled pH, antioxidant screening, inert storage, and verification of oxidation state immediately before every adhesion test — so each result is tagged with its measured chemistry rather than an assumed one.
Can a sharp stiffness gradient actually be fabricated in one disc?
The pad backing needs to be stiff at the rim and soft in the center. Making that transition sharp in a single molded part is hard — unreacted material diffuses across the boundary and blurs it. Easy to draw, hard to build.
Now addressed: the achieved gradient is mapped directly by nanoindentation rather than assumed, with two fallback fabrication routes specified if diffusion blurs the boundary too much.
Does the harvested power actually close the electrostatic budget?
The suit proposes to power its electrostatic grip from the wearer's own movement. That only works if the energy harvested per step exceeds the energy each grip cycle consumes — a number the first draft never actually calculated.
Now addressed: the proposal states the order-of-magnitude math up front — and admits continuous operation likely won't close on harvesting alone, reframing the design around an intermittent top-up cycle with a small energy buffer. The measured margin, not an assumed success, is the deliverable.
A one-to-three kilovolt wearable raises safety questions.
High voltage on a garment worn against the body, possibly in the wet, is a real safety concern — breakdown, leakage, fault behavior. You can't hand-wave it for anything a human will wear.
Now addressed: breakdown margin and wet-condition leakage are characterized on the bench first, with a current-limiting fault architecture validated before any energized human-worn test — and any wear test under formal human-subjects review.
What is your adhesion measurement protocol?
Gecko-style adhesion depends heavily on how hard and how long you press before testing. Without a fixed, stated protocol, two labs measuring the "same" pad get different numbers and nobody can compare results.
Now addressed: a fixed protocol with defined preload, dwell, and detachment rate, referenced to a recognized lap-shear standard, with those variables held constant within each campaign and varied only in a dedicated sub-study.
Your test substrate is clean lab glass; the real world is filthy.
Adhesives that shine on pristine lab surfaces often collapse on the dusty, contaminated surfaces of actual buildings and rock. Testing only on clean substrate flatters the technology.
Now addressed: a dedicated contamination sub-study introduces realistic fouling and tests the self-cleaning recovery against it — with no claim that clean-surface durability transfers to field conditions without that data.
What changed, in one line
The DragonSuit proposal grew from 25 to 32 pages; the GripSuit from 25 to 28. Almost none of that growth is new claims. It's the opposite — it's the addition of constraints, caveats, uncertainty budgets, controlled variables, and several explicit admissions of what isn't yet known. A good revision of a research proposal usually gets more humble, not less, and that's what happened here.
That humility is the whole argument against the "AI slop" worry. Slop is confident and frictionless. These revisions are full of friction we put there on purpose — the friction a good reviewer creates. The reasoning process helped us find it fast and across two fields at once. But the judgment about which friction was real, and the decision to write down the limits plainly, stayed with us.
And it stays with you, too, if you'll have it. The exercise we ran is a strong first pass. It is not a substitute for someone who has actually built these things. If that's you, the contact link is below, and the offer at the top of this page is sincere: tell us what's still wrong, and we'll fix it in Revision 3.
Found a flaw? Good. Send it.
Every substantive critique is read, taken seriously, and fed back into the next revision. The proposals are openly published precisely so that they can be openly improved.
getdragons@dragonworx.bio →
About DragonWorx
DragonWorx Biomimetic Technologies applies materials science and biomimetics to wearable systems, aquatic platforms, and structural materials, based in the Dallas–Fort Worth metroplex. Both research proposals — and the engineering documentation behind them — are openly available at dragonworx.bio.