Eight steps, and the order is the point
Each step may only use what the step before it established. That is what stops a plausible sentence from being mistaken for an engineering result, and it is why a candidate can be rejected here even when it sounds like a good idea.
The real engines, on a fixed input
These results are computed on this page, now, by the same code the application runs. Only the input is fixed.
Live engine output · sample input
Scenario A: Plywood offcuts
Four clean 600 × 400 × 12 mm plywood offcuts from a cabinet job, and a workshop with basic cutting tools.
- Registered material
- Plywood (sheet wood) at 600 × 400 mm × 12 mm, 4 sheet(s)
- Workshop profile
- Sample workshop: basic wood tools
- 6 tools · 7 process steps
Candidates proposed from the catalogue
The model may only pick from this catalogue and set parameters inside declared bounds. It cannot invent geometry, tooling, labour times or consumables.
- 01
Photograph and describe
Upload up to four compressed photographs. The vision model describes only what it can see: category, condition, contamination and restrictions. Measurements are deliberately excluded from the prompt, because a photograph cannot establish one.
Client-side compression, server-side byte-level validation.
- 02
Confirm by hand
You correct and confirm the properties. Every attribute records whether it was AI inferred, user confirmed, measured, or unknown. Confirmation is blocked until critical fields are human verified.
An AI inference is never treated as a measurement.
- 03
Discover candidates
The model may only select from a curated catalogue of validated specifications, and may only set parameters inside their declared bounds. Anything that cannot be substantiated is rejected before it reaches the engines.
Unknown templates, out-of-range parameters and unsafe claims are all dropped.
- 04
Apply the gates
Six ordered checks: material family, sheet dimensions, thickness, condition and contamination, whether every panel fits, and whether the stock yields at least one unit. A failure removes the candidate from ranking.
Rejection reasons are always stated.
- 05
Cost it transparently
Eleven cost lines, each carrying its own assumption and its own derivation. Integer minor units throughout, with a single half-up rounding step. The platform fee is charged on the sale, so it is excluded from production cost.
Every figure is editable and shows its working.
- 06
Evidence the market
Prices come from a curated reference set, aggregated to a median with 25th and 75th percentiles. Each row is labelled as a transaction, quotation, listing, or estimate, and the set is scored for quality and freshness.
No evidence produces Insufficient Market Data, not a price.
- 07
Match buyers
Buyer output is a category of buyer with the matched factors shown, or a registered request whose owner consented to be shared. Quantity shortfalls against available capacity are flagged rather than smoothed over.
No company is named unless it registered itself.
- 08
Rank and explain
Hard gates run before any score. Viable candidates are scored on feasibility, economic value, material utilisation, buyer compatibility, evidence quality, and simplicity, with an explanation per dimension.
Unknown dimensions are excluded and reported, never scored as zero.
The five verdicts, and the five sort modes
Fifteen words do a lot of work in this product. These are the exact states a candidate can be in, and the exact ways you can order them.
Feasibility verdicts
- Process feasible
- Nests in the stock and every required tool is available. Still an opinion, not a certification.
- Geometrically feasible
- The parts fit the sheets, but no workshop profile was supplied, so the process could not be verified.
- Conditionally feasible
- Nests, but tooling is missing, the condition is poor, or defects were recorded but not located.
- Not feasible
- A hard constraint failed: wrong material family, a panel that cannot be cut, too thin, contaminated, or no unit fits.
- Not evaluated
- Required material properties are still unconfirmed, so nothing is decided either way.
Sort modes on the comparison screen
- Overall scoredimension value
- Economic potentialdimension value
- Material utilisationdimension value
- Production simplicitydimension value
- Buyer compatibilitydimension value
Sorting reuses the same normalised dimension values the ranking engine scored, so the order on screen can always be traced back to an explanation. Candidates that failed a gate stay at the bottom in every mode.
What this version does not do
Stated here rather than discovered later.
- Feasibility is a flat, rectangular, 2D model. No 3D, no structural analysis, no machine nesting file.
- Cutting yield assumes a guillotine layout and over-counts waste rather than under-counting it.
- Defect positions are not recorded as coordinates, so defects reduce expected yield instead of being excluded from the layout.
- Market evidence is a small curated sample set. Live marketplace data is not scraped.
- Buyer matching is category-level. Registered demand is the only thing treated as verified.
- Commerce is not implemented: no request-for-quotation, no quotation, no payment, no transaction.
- Outcome tracking and sustainability metrics are deferred. No carbon figure is claimed.
- The AI reads photographs. It cannot verify weight, grade, composition or certification, and it is instructed never to try.