Two Gates Before CAD: Where 70–80% of Product Cost Locks In

Most of a product's manufacturing cost is committed before the first prototype is ever built — locked in by early decisions on material, geometry, and tolerance, according to manufacturing-cost analyses. Yet in 2026 the default workflow still runs backwards: teams spend their first six hours per concept inside CAD, discovering only afterward that a feature violates a one-degree draft rule a checklist would have flagged in minutes.

The bottleneck has moved. Large language models have flooded concept development's front end with ideas; what they haven't priced is the modeling labor behind each one. Six engineer-hours per failed concept, multiplied across a portfolio, is the quiet tax on 2026 hardware teams — unpriced, unbudgeted, and almost entirely avoidable at a structured gate.

The remedy is sequencing. Two gates belong before CAD: a manufacturability screen and a human sign-off. Early-screening systems such as SelecT.AI — introduced January 2, 2026, and built explicitly 'not as a chatbot' — pair DFMA risk checks with a mandatory 'Human Decision Required' gate, mirroring how veteran engineers ask 'but does it' before tooling. Gate first, model second.

Two Gates Before Geometry

Stone and Wood's functional basis holds a vocabulary of verb-object terms — "import electrical energy," "secure solid to frame" — and that vocabulary alone is enough to kill a weak concept before anyone opens a CAD tool. The familiar objection, that you cannot judge novelty or manufacturability until geometry exists, gets the causality backwards: draft angles, wall-thickness targets, and working principles are all properties of the written functional description, the same information an engineer holds after 30 minutes of sketching. No solid model is required input, which is precisely what makes this a pre-CAD instrument rather than a post-modeling audit.

Gate 1 decomposes each concept into a function-solution tree in that vocabulary and scores novelty branch by branch: every branch is compared against the same branch in every other concept of a fixed comparison set of at least 10 prior concepts, and a working principle shared by few concepts scores high. Weights increase at deeper tree levels, following Shah, Vargas-Hernandez and Smith's Design Studies metric — correctly so, because top-level functions like "transfer mechanical energy" are generic across an entire product class, while leaf-level solution choices are where competing designs actually differ.

The LLM's job in gate 1 is deliberately narrow: a GPT-4-class model, prompted at temperature 0 with a fixed rubric in the system prompt, emits each concept as a structured JSON function-solution tree. Temperature 0 matters because the comparison set is frozen — identical concept text must yield an identical tree across runs, or branch-level scores drift between sessions. Novelty is then computed one of two ways: have the model cross-compare the trees directly, or cluster text embeddings and apply a cosine-similarity threshold near 0.85 to mark "same working principle."

Gate 2 is a second LLM pass over the same written description, checked against a one-page, process-specific DFM rule list. Each violation is tagged critical or advisory: critical means the named process cannot produce the feature as described; advisory means it survives only with secondary operations — the multiple-setups, scrap, and supplier-pushback pattern that DFMA.com identifies as the signature of a design fighting its process.

Sequence is an economics argument. Novelty runs first because a duplicate dies cheapest — one tree parse, no DFM analysis spent. DFM runs second because a novel-but-unmakeable concept must die before CAD hours are committed. A concept advances only if it scores novelty of at least 3 out of 5 on at least one leaf subfunction and carries zero critical flags; everything else returns to ideation. And according to Rajavel Rajesh's account of screening pipelines, every run terminates at a mandatory "Human Decision Required" step — the model ranks, the engineer signs.

The ordering has institutional backing. According to SEBoK (updated May 18, 2026), manufacturability and producibility can serve as discriminators between competing system concepts during the study period, well before the final design matures. Automated tools such as aPriori's aP Generate notify teams seconds after a CAD model reaches PLM check-in — genuinely useful, but the solid model already exists and the modeling hours are already spent. The two-gate screen runs on the paragraph instead. Before your next ideation session, freeze two artifacts: the pinned comparison set of at least 10 prior concepts, and the rule sheet below.

The Numbers That Justify the Gate

Most of a product's manufacturing cost is committed by decisions made during design, according to manufacturing-cost analyses. Ullman's The Mechanical Design Process supplies the escalation arithmetic: catch an error one stage later and it costs roughly ten times more — the "rule of 10." Together, those two findings define the gate's economics. A concept that survives three weeks of modeling before dying in review is an error caught at the most expensive point on the timeline; screening the written concept instead intervenes where correction is cheapest.

This is also where the standard objection collapses. The claim that you cannot judge manufacturability or novelty until geometry exists confuses the solid model with the decisions it records. Draft-angle requirements, wall-thickness targets, and overhang limits follow from the chosen process and the functional description — both fully specified in a one-paragraph concept — which is precisely why a gate can precede CAD rather than trail it.

The throughput case rests on controlled trials, not vendor benchmarks. According to Noy & Zhang's experiment published in Science, knowledge workers completed professional writing tasks faster and with higher quality under LLM assistance; according to Peng et al.'s controlled experiment, developers finished a standardized coding task 55.8 percent faster with an LLM pair-programmer. Gate evaluation blends those two task profiles — constrained extraction plus checklist verification — so pushing dozens of concepts through both stages in a single day is an empirically grounded expectation, not speculation.

Error rates are bounded, too. According to the Vectara Hallucination Leaderboard, leading models hallucinate on only a small fraction of sentences in summarization benchmarks. Populating a function tree from a short concept paragraph is bounded extraction against a fixed verb-object vocabulary, not open-ended reasoning, so per-concept extraction error sits near that floor — and unlike a silent modeling mistake, it is auditable: any engineer can diff the proposed tree against the source paragraph in minutes.

Gate one also inherits a calibrated yardstick. According to Shah and colleagues' Design Studies paper, their novelty metric correlates with expert judges' rankings of the same concept sets. An LLM-populated tree scored against the fixed comparison set therefore runs on an instrument with demonstrated agreement with human experts — a validated scoring scheme, not an ad-hoc vibe check.

The boundary condition is the honest caveat. According to Dell'Acqua et al.'s Harvard Business School working paper, a large randomized experiment with consultants found that GPT-4 users improved output quality substantially on tasks inside the model's capability frontier but were markedly less accurate when they trusted its answers beyond it. The gate lives entirely inside the frontier: novelty thresholds and severity flags are codifiable rules with binary outputs. Where your criteria resist codification — brand language, market timing — the gate correctly defers to human judgment.

Before trusting the gate forward, validate it backward: take the last ten concepts your team abandoned mid-modeling, run both gates on their original paragraphs, and log which stage would have killed each one. If the gates reconstruct your kill list, you hold calibration evidence — and the next CAD session opens only for concepts that earned it.

Gate Order Showdown

Content for Gate Order Showdown is being prepared.

What the Data Doesn't Tell You

The gate's weakest joint is its evidence base, and pretending otherwise is how teams get burned. Nearly everything published about pre-CAD screening rests on small case collections rather than controlled comparisons, and the outcome that matters — wasted first-pass modeling hours — tends to be self-reported by the very teams that adopted the screen. Concepts killed at the gate rarely leave an archive, so you observe the saves and never the false kills. Because the scorer is an LLM, calibration also drifts whenever the underlying model version changes: a 3/5 issued before a vendor upgrade and one issued after it are not guaranteed to mean the same thing.

The tempting response to thin evidence is to retreat behind geometry — wait for the solid model, then judge. Resist it. Draft requirements, wall-thickness targets, and working principles are properties of the written functional description, which is exactly why both stages stay decidable before CAD. The deeper measurement problem is corpus-relativity: Shah's novelty metric is defined against whatever prior concepts you retrieve, so the ten-concept floor stabilizes the score without validating it. Pugh's concept-selection method dodged this by scoring strictly relative to a datum; an absolute 3/5 cutoff imports a precision the construct doesn't have.

Variance across cases runs wider than any single pilot suggests. The checklist thresholds — draft of at least one degree, uniform-wall rules, overhangs at forty-five — carry injection-molding and extrusion-printing instincts; apply them unedited to sand castings, sheet metal, or five-axis machining and you manufacture spurious criticals or miss real ones. Domain matters on the novelty side too: a die-cast housing scored against a corpus of printed enclosures lives in the wrong neighborhood, and swapping two entries in a ten-slot set can flip a pass into a fail. The gate is conditional on its inputs, not a property of the concept alone.

So when does the rule break? Not at its core — geometry still waits — but at its edges. A cross-domain breakthrough can score below threshold purely because its nearest neighbors sit outside the corpus. A cost-down refresh of your own catalog is intentionally un-novel and fails a floor it was never meant to meet. A concept with zero critical flags but a dense cluster of moderate ones sails through and burns modeling hours anyway. Each is an input problem, not a refutation: widen the corpus, document the exemption, read flag density as a signal rather than a second veto.

Before your next screen, audit three inputs: corpus breadth, checklist-process match, and model-version pinning. Those audits — not additional pilot hours — are what make the gate's thresholds worth trusting.

What the Gate Cannot See

Doshi and Hauser's 2024 experiment in Science Advances is the uncomfortable mirror for LLM-judged novelty: generative-AI assistance raised individual output creativity while reducing collective diversity. A model scoring "how similar is this to prior art?" measures distance against its own training distribution, whose center of mass is the average of everything ever published. Genuinely strange leaf subfunctions have no near neighbors, so their measured distance collapses and the score comes back ordinary. The gate rarely rejects weird ideas outright — it quietly re-scores them as familiar, tilting the surviving portfolio toward incremental designs. The loop self-tightens, too: research on generative AI flags a contamination pathway in which "AI-generated content included in new data crawls from the Internet" propagates into future training corpora, so homogenized outputs become tomorrow's comparison set.

The second fragility is arithmetic. Shah's tree-based novelty metric is only as good as its reference set, and below roughly 10 comparators the score loses resolution — too few leaves to separate "somewhat novel" from "genuinely novel." Delegating retrieval to an LLM cuts both ways: weak or wrong prior art inflates novelty, and no published benchmark quantifies how large that inflation runs. Freeze the comparison set before prompting — a fixed roster of at least 10 documented concepts — and log which retrievals fed each score.

Stage two has a coverage problem, not merely a knowledge problem. Published DFM guidelines encode perhaps the top few dozen failure modes per process and stop there — silent on tolerance stack-up interactions across mating parts, resin-specific warp from differential shrinkage, and assembly sequencing conflicts. Dell'Acqua's field experiment, in which AI-assisted professionals fell measurably below baseline on out-of-frontier tasks, implies the deeper defect is miscalibration: the model issues confident verdicts on physics no checklist encodes. Confidence, not silence, is what slips past review.

Scope claims deserve equal skepticism. Gate accuracy has effectively been demonstrated only for discrete-rule processes — molding, machining, FDM — where rules are codified and largely independent. For castings, forgings, and multi-material overmolding, where rules interact nonlinearly, there is essentially no validation data. Treat verdicts there as unproven hypotheses, not screen results.

The error costs run asymmetric, and skewed the wrong way. A missed DFM flag surfaces in CAD and costs days; a wrongly killed novel concept forfeits the project's entire upside — and no study yet reports that false-kill rate, so nobody knows how often it fires. The defense is bookkeeping: a killed-concept ledger holding the full prompt, scores, and comparison set behind every rejection, each entry on a scheduled revisit.

Staleness compounds quietly. Training cutoffs predate material launches, so a high-temperature resin introduced after the cutoff gets audited against rule sets that have never heard of it. Refresh the checklist manually on a fixed cadence, synchronized with model-version updates, or the gate degrades with every release.

Note what none of these repairs require: geometry. Every blind spot above is an input defect — thin comparison sets, uncoded physics, stale rules, unlogged kills — fully decidable from the written functional description. That is precisely why the gate belongs before CAD rather than after: the failures are epistemic, not geometric, and modeling first would bury them under solids.

Worked Case

Content for Worked Case is being prepared.

Five Rules for Trusting the Gate Without Getting

A novelty score is not a property of your concept — it is a property of your concept relative to whatever comparison set you fed the model. That single fact drives most gate failures in practice: teams adopt the two-stage screen, run it on defaults, and then defend scores nobody can reproduce. The five rules below are input controls and output controls on the gate itself, not prompts to whisper at the LLM. And none of them require geometry to exist first — draft-angle requirements, wall-thickness targets, and working principles are fully decidable from a written paragraph, which is exactly why the gate can sit upstream of CAD instead of downstream of it.

Ground before gating. Write the one-page DFM rule list for your target process before scoring anything: minimum draft, minimum wall, corner radii, overhang limits. The ANONMGUR DFM Guide opens its Step 01 with the right framing — identify what the component must do, what loads it carries, how it interfaces with neighboring parts, and which features are truly critical — because the rule list falls out of those answers. An injection-molding page looks nothing like an SLS page; if you cannot produce yours, do not run the gate. Stay CAD-first until the list exists, or you will be scoring concepts against imaginary constraints.

Freeze the comparison set. Assemble at least 10 comparator concepts — internal archive plus patent prior art — before the first score is generated. Shah-style novelty metrics are comparative by construction: each leaf subfunction is judged against its peers in the set, so a score produced against eight concepts is not a weaker version of the real number, it is a different number. Label every such score provisional, in writing, wherever it lands.

Demand a double pass. Advance a concept to CAD only on novelty of at least 3 out of 5 on at least one leaf subfunction AND zero critical-severity DFM flags. Single-pass survivors go back to ideation, never to the modeler — a concept that clears novelty but carries a critical flag has not earned first-pass modeling hours, and neither has one that is manufacturable but derivative.

Fence off uncoded physics. Route safety-critical parts, tolerance-critical features tighter than ±0.05 mm, and simulation-dominated questions directly to CAD plus FEA, regardless of gate score. A checklist encodes lookups; a stress margin requires analysis. No checklist and no LLM owns those judgments, and letting a clean gate score launder a fatigue-critical feature is the most expensive way to trust the tool.

Audit on a clock. Every quarter, sample 10 gated concepts and compare the LLM's flags against what actually failed in prototypes or mold trials. Refresh the rule list and the model version together — vendors ship silent updates, and a checklist tuned to one version's behavior decays against the next. If the gate's catch-rate drops below your measured CAD-rework baseline, retire it; the gate must continuously beat the status quo it replaced, not merely exist.

The concrete next move: this week, pick your highest-volume process and draft its one-page rule list while pulling ten archived concepts into a frozen comparison folder. If the list takes longer than the folder, you are not ready to gate — and discovering that costs an afternoon, not a modeling sprint.

What to do next