What Factory Are You?
You run some piece of knowledge work over and over - turning out posts, models, decisions, enriched data. Once that line is stable enough to name its inputs, steps, outputs, and defects, you can run it like a factory. The useful question is which factory: the physical industry your line most resembles has already solved your quality problem, and you can borrow its playbook where the constraints actually match.
Answer six questions and you get a candidate archetype and the first diagnostics to run - not a final verdict. The match is read off a coarse input / output / value-shape signature (what actually binds you: how cheaply you can test quality, how much waste is in your inputs, how costly a defect is), with one follow-up question where the signature is ambiguous.
Where does the value come from?
The strongest single signal. How is your payoff shaped - by upside breadth, downside control, or both in sequence?
What does the line actually emit?
Output kind co-varies with archetype; it should not be forced to converge with the others.
What is your raw material?
The front of the line. High-grade input needs no concentration; free low-grade exhaust forces an assay station.
How much of the input is waste vs. signal?
Signal-to-ore. Extreme variance - a tailings pile with rare seams - forces exhaustive screening plus a cheap pre-filter. Uniform feed needs none.
Is the output a one-off or a reusable recipe?
The mechanism that turns a line into a meta-factory. Distinguish a reusable PLATFORM from a reusable unit OUTPUT - a configurator can be reusable while each unit it emits is a one-off.
Is this high-volume / scaled, or low-volume / bespoke?
Volume and marginal cost are separate (a line can be high-volume yet have real per-unit model/API/review cost). This asks scale; if your per-unit cost is non-trivial, throughput/capacity stay first-order regardless of archetype.
The shape of your line
A neutral schematic for now - it resolves into your line’s curvature once all six are answered. The value-add smile is drawn as a hypothesis, not a law - it holds only when the commodity middle can be cheapened without hurting risk-adjusted quality (otherwise the middle is the moat). J just records the order of convex (+) and concave (−) stages; whether it is a meaningful invariant is an open question on the formalism page.
Classify your line (0 of 6 answered)
Known collisions in the current working set
The signature is deliberately coarse, so some archetypes share a six-coordinate projection. When your line lands on one of these, the classifier returns a candidate and the discriminating question rather than a confident label. This list is not exhaustive - on a small, internal sample there are likely unlisted collisions; send counterexamples.
Classify your own line with an LLM
The six questions above are a compression of a longer elicitation. Paste this structured elicitation prompt into a capable model and describe your line in plain language. It is a guided interview, not a deterministic classifier - different models (and borderline lines) can land differently, so it asks the model to report all matched archetypes and its uncertainty rather than forcing one answer.
You are an operations analyst. I will describe a knowledge-work production line.
This is a structured elicitation, not a deterministic classifier - report all matched
archetypes and your uncertainty, and ask for evidence rather than asking me to self-label.
PRECONDITION: first confirm this is a REPEATABLE, stable production line (definable inputs,
transforms, outputs, defects). If it is exploratory, non-repeatable, or environment-shaping
(novel research, negotiation, bespoke strategy, incident response), say the typology does not
apply and stop. NOTE: a repeatable line that produces bespoke one-off UNITS is still fine -
that is the aerospace / job-shop case; the stop condition is non-repeatability of the LINE,
not one-off outputs.
Elicit these six coordinates using behavioral questions (e.g. "does 10x more candidates
help, or does tightening each attempt help?" for curvature; "how many units do you produce
after the recipe is fixed?" for volume). Do NOT ask me to pick a curvature/archetype label.
1. Risk curvature: convex (breadth of attempts pays; rare fat-tailed hit), concave (downside/defect control pays; can be high-volume), or mixed (discovery feeds production).
2. Output kind: artifact | function/policy/weights | data-slate/graph | type/reusable-process.
3. Input grade: ultra-high (scarce expert) | mixed/self-generated | low-abundant (free exhaust).
4. Input variance (signal-to-ore): low (uniform) | mixed | extreme (mostly waste, rare seams).
5. Transferability: reusable-recipe/platform (amortizes) | one-off. (A reusable platform can still emit one-off units.)
6. Volume / scale: scaled-high (many units) vs bespoke-low. (Volume and marginal cost differ - high volume with real per-unit model/API/review cost still binds on capacity.)
Selection rules produce a CANDIDATE, not a verdict. Gates exclude defining-coordinate violations; if none passes, ABSTAIN ("none of these as stated"). Several archetypes then need a second-stage discriminator the six questions do not capture - ask the relevant follow-up to separate the known collision (the output stays a candidate, not a verdict; sometimes more than one follow-up is needed):
- mixed curvature => candidate Vertically Integrated Process Chain. Confirm both halves are load-bearing (else classify the dominant stage). Discovery half is HTS-style only if cheap assayable breadth + fat tail (else a generic convex search); production half by volume+output.
- convex + (mixed/extreme) variance + reusable-recipe + function-ish => HTS.
- convex + low-abundant + extreme variance => Salvage (a low-value exhaust/waste stream with recoverable rare seams; internal exhaust - your own logs, tickets, failed generations - counts too, it just changes governance, not the archetype).
- concave + type-process => Pilot plant. concave + function + scaled => Foundry (candidate - confirm shared-consumer count: many => Foundry, single => single-tenant function deployment). concave + artifact + reusable + scaled => Mass-customization OR Float-glass (confirm per-unit variety: configured => Mass-cust, uniform AND uniform feed => Float-glass, uniform output but heterogeneous feed => neither, re-examine).
- concave + low-abundant + data-slate + scaled => Refinery (candidate - confirm flow mode: continuous => Refinery, batch => batch data job).
- bespoke-low + one-off + concave => Aerospace (candidate - confirm defect consequence: catastrophic => Aerospace doctrine, low => low-stakes bespoke).
Do not commit a first move until any needed follow-up is answered; if unanswered, say "insufficient information - answer [follow-up] first." For the resolved archetype, attach a falsifier and output: archetype, exemplar, what transfers AND what does not, the per-layer QC regime, and the highest-leverage first move tied to observed bottleneck evidence.The triple signature, the archetype zoo, the smile curve, and the make-vs-buy boundary - in plain language.
The structural functor, the category error named honestly, and the open obstruction-invariant program.
Changelog9 entries · building in publicshow ↓
- Legibility + concreteness pass on the six-question diagnostic (Gemini-graded). Larger option-button text, higher-contrast hint and example tiers, more line-height. Every choice now carries a concrete real-world "e.g." line (e.g. "cold outreach, A/B headline tests, VC dealflow" for breadth-of-attempts) so the abstract coordinates are easy to map onto your own line.
- Adversary round 9 (gpt-5.5). Fixed the precondition contradiction: the stop condition is non-repeatability of the LINE, not one-off outputs (a repeatable line producing bespoke one-off units is the aerospace / job-shop case, which is in scope). Softened the smile-diagram label ("minimize ownership if it does not hurt end-to-end quality", not "cheap to ~$0").
- Adversary round 8 (gpt-5.5). Before all six are answered the diagram is now a NEUTRAL unlabelled schematic - it no longer derives a default-filled classification. Salvage no longer hinges on third-party-vs-internal feed (internal exhaust - logs, tickets, failed generations - is salvage too; the distinction is governance, not archetype). Softened "the one decisive follow-up resolves it" to "separates this known collision; output stays a candidate."
- Adversary round 6 (gpt-5.5). Confirmation redirects now check the redirected archetype's own gate - a uniform-output line with heterogeneous feed abstains instead of being mis-labelled float-glass (output-uniformity is not feed-uniformity). Salvage gate now requires extreme variance; aerospace gate drops the input-grade prerequisite so defect-consequence can decide it. Aligned the copyable LLM prompt with the gated + confirm + abstain semantics.
- Adversary round 5 (gpt-5.5). Explicit unanswered state: no diagnosis until all six are answered (defaults no longer masquerade as your answers after one click; partial URLs no longer pose as complete). The smile curve in the diagram is labelled a hypothesis (holds only if cheapening the middle does not hurt risk-adjusted quality). Relabelled the volume coordinate so high-volume is not equated with near-zero marginal cost.
- Adversary round 4 (gpt-5.5). Progressive confirmation: when a candidate's identity hinges on a coordinate the six questions cannot capture (shared-consumer count, per-unit variety, flow mode, feed source, defect consequence), the classifier no longer commits a primary - it asks the ONE decisive follow-up and resolves on the answer (confirm the candidate, redirect to the paired archetype, or resolve to a non-zoo alternative like single-tenant function / batch data job). Chains require a stage-dominance answer before they are called chains.
- Adversary round 3 (gpt-5.5). Replaced additive-only scoring with required gates per archetype: a line that violates an archetype's defining coordinate is excluded, and if nothing is eligible the classifier ABSTAINS ("none of these") instead of forcing a fit. Added an ambiguous-tie state and a "what six questions cannot tell apart" disclosure (the known indistinguishability classes and the follow-up question for each). Canonical cases locked by fixtures (not a claim of full calibration).
- Adversary round 2 (gpt-5.5). Result no longer shown as a diagnosis before you answer (labelled example; URL not written until you change an answer). Added a low-confidence state and absolute heuristic scores. Added float-glass, a salvage input-stage in chains, and aerospace as a bespoke production half. Fixed the HTS QC claim and softened universals to candidates.
- First public draft. Six-question diagnostic over the (input, process, output) signature; classifier wired to the same source-of-truth used by the strategy doc and the prompt set.
Rosetta Stone
Four circles, four readings of the same object. Each role reads the artifact through its own lens.
A diagnosis of which operations playbook your capital should be buying. The archetype fixes the QC regime, the make-vs-buy boundary, and where margin accrues - so it fixes where the next dollar goes.
A placement test for your whole line, not one task. Name your input, process, and output signature and you inherit a mature physical industry's runbook instead of reinventing it. Find the sign-flip and you know where your hardest gate belongs.
The architecture decision before the architecture. Convex discovery stages want cheap-screen-then-confirm; concave production stages want strict gates and SPC. Build them as separate subsystems with a buffer at the joint.
A classifier archetype = f(σ) over a triple signature, falsifiable by a bounded-zoo compression claim: the next independent line should decompose into the same finite atom set, or the typology was overfit.