Guide · Layer 5
Written for: seed / Series A scientist-founder
The first four layers built and pressure-tested an engine: Layer 1 fixed the shape, Layer 2 put physical numbers on the flows, Layer 3 collapsed them into one figure — the levelized cost, ~$800/t for green ammonia — and Layer 4 found the two or three drivers it rides on and stress-tested how robust it is. The result is a number you trust and understand. This last layer is about getting it out of your head and your spreadsheet and into someone else’s — an investor, a partner, a diligence team — without losing its meaning on the way out, and without giving away the trade secret that produced it.
Every prior layer could fail only one way: a wrong number — a unit slip, a broken balance, a mis-annualized capex. The communication layer can fail with a number that is perfectly correct. The freedom to choose what to show — which metric, which scenario, which basis, gross or net — is the freedom to mislead without writing a single false digit. That makes the discipline here different from every layer before it. The arithmetic is done; the number is as right as it’s going to get. The skill now is honest compression under two pressures that pull against each other: you have to cut enough to fit a slide and to protect the IP, yet keep enough that the figure still means something and still survives a hard question. Cut the wrong things and you either mislead your audience or hand them a number they can’t defend — both of which end the conversation you were trying to start.
That tension defines the whole layer. A TEA travels as two different artifacts for two different moments: a deck that persuades a non-specialist to look closer, and a shareable model that defends the result when their technical team actually looks. Neither is the real working model — that one holds your IP and never leaves the building. The job is to derive honest, audience-fit versions of it that say exactly as much as they should.
This layer runs five moves — compress to an honest headline → match the artifact to the audience → make the shared version trade-secret-safe → make it defensible → dodge the credibility traps a true number can still spring. Each routes into the concept pages for the how; the reads below are about what to show, to whom, and what to never cut.
The whole analysis has to collapse to one figure a reader remembers and acts on: the headline number. Usually that’s the levelized cost, but it might be a net cost after credits, a breakeven price, or a margin — which figure leads is a presentation choice, not a fixed output. Pick it, then attach the three things that make it interpretable: its band (the ±accuracy class), its basis (the system boundary, capacity factor, and annualization it was computed on), and its drivers (the few inputs that move it).
🧭 Coach’s Read. The deliverable here is a sentence, not a number — ”~$800/t NH₃ (±~30%), gross of credits, power-to-ammonia boundary, CF ~0.90, driven by the power price” carries the figure, its band, its basis, and its top driver in one line, and a bare “$800/t” carries none of them. The band and basis are not detail you’re allowed to drop for brevity — they are the number; without them “X beats Y” is unfounded, and the basis is the part most often left off. The sharpest trap in this whole layer lives right here: the freedom to choose a headline is the freedom to mislead with a true number — picking the net-of-credit figure, the best-case capacity factor, or the one comparator you beat yields a defensible-looking number that misrepresents the central case without a single false digit. The discipline that defeats it is to show gross and net side by side, always. A clean-hydrogen credit can rival your entire gross cost, so reporting only the netted ~$260/t hides that the process makes ammonia at ~$800/t and policy closes the rest — and policy is jurisdiction-specific, time-limited, and conditional. Anchor on the gross figure — that’s what the process costs — and present the net as the clearly-labeled upside it is (what policy can close), never as a quiet substitution for it. A cheap test for whether your headline actually communicates: try explaining it to a sharp non-specialist in nothing but boxes and arrows. If you can’t make them nod without opening the spreadsheet, the compression isn’t finished — and the spots where you reach for jargon are exactly where a real investor will lose the thread.
A TEA does not travel as one thing. The pitch-deck version is the most compressed form it takes — a headline, a benchmark, the drivers, a one-line basis, on a slide — read fast by a non-specialist deciding whether you’re worth a closer look. The shareable model is the opposite artifact: a workbook a technical diligence team opens, interrogates, and runs their own sensitivities on. The deck persuades; the model defends. Build the right one for the moment.
🧭 Coach’s Read. Don’t confuse the two artifacts, and don’t let one stand in for the other. The deck’s job is orientation, not documentation — it carries the conclusion and the reason to believe it, and routes every detail to the model behind it. Match depth to the audience: a partner skimming fifty decks needs one honest slide; their engineer in diligence needs the sourced workbook. The danger sign is a deck number with nothing behind it — a slide figure with no defensible model and no provenance a follow-up question can reach collapses on the first diligence query, and you’ve spent your credibility instead of building it. And respect that the deck is where compression is most dangerous, not just most useful: it shows the least and is seen the most, so every omission — a dropped band, an unstated basis, a cherry-picked scenario — misleads more readers there than the same omission buried in a model ever would. Maximum lossiness meets maximum reach. The compression is in detail; it is never in honesty. Two things worth keeping front of mind for this move. First, the deck is the TEA’s argument, not a screenshot of it — never paste the spreadsheet onto a slide. Each slide should make exactly one assertion and earn it with a single chart (a “killer slide”): here is the cost wall we beat, and here is the one mechanism that lets us beat it. An investor believes a cost advantage when you can name which unit operation is cheaper and why — a mechanism, with margin for error — not when you show a smaller bar or a top-down TAM/SAM/SOM bubble; a bottom-up number they can follow beats a top-down one they can’t. Second, whoever owns the numbers has to be in the room. The deck opens the door, but the instant someone pushes on a figure, the person who built the model has to field it live — a number nobody present can defend rarely survives the first hard question, however right it is. That’s the other half of building the model yourself.
The working model contains your IP: exact recipes, the novel step’s parameters, raw vendor quotes. The shareable model is a separate artifact derived from it — not the working file with cells deleted — built from three moves: black-box the proprietary unit operations (keep the honest interface, hide the mechanism); band sensitive numbers (a range or order of magnitude instead of an exact value); and aggregate line items (roll proprietary sub-costs into one block). The art is doing all three while the mass and energy balances still close, the headline and its basis stay intact, and the drivers stay visible enough for the recipient to run their own sensitivity.
🧭 Coach’s Read. Black-boxing protects the secret only if the crossings stay honest — the box can hide how the stack reaches its efficiency, but it must still declare every real input and output, or the surrounding balance breaks with no visible cause. Then band the interface, because an exact one leaks the very secret the box exists to protect: publish ≈10 MWh/t and ≈0.18 t H₂/t, not 9.73 and 0.1804, or a competent reader backs out your stack efficiency from three significant figures. The hard line — the one that separates protection from fraud — is this: honest abstraction blurs detail; it never shades the answer. Banding to the favorable end of a range, or quietly swapping in the best-case capacity factor to land under $700/t, isn’t protecting IP — it’s a different, misleading model wearing the real one’s conclusion. And do not over-abstract: a model so protected that no driver is visible and the headline can’t be sensitized is also too opaque to be believed — a reader who can’t test it can’t trust it. You live on a frontier here: the most abstraction that still leaves a result the recipient can sensitize and defend. One operational warning — the moment a shared artifact exists, it can drift from the working model; a correction you make internally but don’t propagate silently breaks the shared figure’s provenance even though every number still looks sourced.
A model travels to people who can’t re-derive it, so they test it the only way they can: by asking where each load-bearing number comes from. Provenance is the property that every figure traces to its origin — a citation, a quote, a comparable process, or an explicitly stated assumption — and defensibility is provenance made external, the ability to answer that question under scrutiny. Not all sources are equal: a vendor quote, a peer-reviewed value, and a press-release figure sit at different tiers of the source hierarchy.
🧭 Coach’s Read. Defensibility is set by the weakest load-bearing number, not the average — one unsourced driver undermines the result the way a single weak link breaks a chain, and a hundred well-cited minor inputs don’t compensate. So spend your sourcing effort on the drivers, the same handful Layer 4 found, and don’t burn hours citing a non-driver the answer is insensitive to — its provenance, however thin, barely moves whether the conclusion holds. Two traps that look like provenance but aren’t. First, provenance is traceability, not correctness: a well-cited number can be wrong and a real quote for the wrong configuration is fully traceable and still wrong — and the basis has to travel with the value, because a correctly-cited cost on an old cost year, a different scale, or a different boundary is sourced but not transferable. Second, and most often fatal in diligence: source the condition, not just the value. A ~$540/t clean-hydrogen credit cited to the policy’s headline rate but not to its eligibility rules — the carbon-intensity tier, the hourly-matched clean power — is defended on the number and undefended on whether you actually qualify. It survives “what’s the rate?” and collapses on “do you meet the conditions to earn it?” — which is the question that gets asked.
Everything above can be correct and the presentation can still quietly destroy trust. The recurring failure of this layer is not a false number — it’s a true one shown in a way that claims more than the analysis supports. The false-precision concept catalogs the family; the move here is to actively avoid them on the way out the door.
🧭 Coach’s Read. Match the displayed resolution to the real resolution. A spreadsheet returns “$805.27/t” because arithmetic preserves digits, but every input is a round anchor — a ~$40/MWh market price, a ~$1.0bn order-of-magnitude capex — so the four trailing figures are noise dressed as signal; report ~$800/t ±~30% and you’ve stated exactly what the method knows. The bare point estimate is the same crime by omission — a figure with no band reads as a commitment, more confident than an explicit wide range, and sets expectations the ±30% analysis can’t honor. Watch the comparison traps too: benchmark like-for-like (a first-of-a-kind plant against a decades-optimized incumbent, or a cradle-to-gate cost against a gate-to-gate one, overstates your position on the basis, not the process), and never benchmark against your own input anchor — if the incumbent figure is the one you seeded the model from, the comparison is circular. The throughline of all of them: on a slide, four small “just simplifying” omissions — dropped band, unstated basis, hidden gross figure, unflagged credit-dependence — combine into a promise of competitiveness the analysis never made. The honest version is less impressive and infinitely more durable, because it’s still standing after the first hard question. One specific instinct to fight: the urge to delete the number that looks bad. A founder I worked with wanted to cut a figure that looked absurd on its face — a roughly 900% first-year ramp — certain it would scare investors off. Hiding it is usually the wrong instinct; the fix was to show it, with the one line of context that made it reasonable. An investor who later stumbles on a number you buried assumes the worst, because concealment reads as evasion — and they redo the arithmetic themselves anyway. Show the awkward number with its honest context; you almost never gain by giving a diligent reader the feeling they caught you hiding something.
🪜 Leveling Up — the full diligence-grade data room. The one-slide headline and the banded shareable model above are built to open a relationship and survive a first technical pass. The deeper artifact is a complete, sourced data room: every assumption documented to its tier of the source hierarchy, the full sensitivity and scenario set, often an NDA-gated un-black-boxed model, and sometimes third-party validation of the key figures. Climb to it when the relationship has earned the disclosure — a lead investor in confirmatory diligence, a strategic partner’s engineering team, a project-finance lender underwriting the asset — not before. The cost is real: weeks of assembly, a defended provenance for every load-bearing input rather than just the drivers, and exposure of detail you’ve deliberately been protecting, which is why it sits behind an NDA and a term sheet rather than in a deck. At the maturity anchor, an honest headline sentence plus a banded model that keeps its drivers visible is deliberately the right tool — it opens the door the data room later walks through.
The five moves, run once on green ammonia — every figure a round, illustrative anchor carried from the concept pages and the reference sheet, where the full build and its open validation items live:
The pattern, same as every layer: the magnitudes and mechanics stayed on the concept pages; this layer sequenced them — compress, fit the artifact, protect, defend, and avoid the traps — and turned a trustworthy number into one that travels and survives contact with an audience. A late-stage guide could route the same six pages into a full diligence data room instead; that swappability is the point.
That closes the five-layer arc. Layer 1 gave the process its shape, Layer 2 its physical flows, Layer 3 its cost, Layer 4 the drivers and the robustness — and Layer 5 sent the result out into the world honest and intact. The loop the analytical layer opened — from the cost model back to the lab bench — is the real endgame: a TEA built this way doesn’t just answer “can this win?” for an investor, it tells you which experiment to run next. That’s the whole reason to build one this early.