TrueGauge / Matching accuracy
How often the matcher is right
Everybody in this field publishes speed — minutes rather than weeks, two hundred times faster. Nobody publishes accuracy, which is the number that decides whether the speed is worth anything. A mapping engine that is fast and wrong is a machine for producing confident errors at scale.
0wrong commits in 25
Zero wrong commits across 25 labelled descriptions. That is the figure worth leading with: a wrong commit produces a number that looks exactly like a right one and is found, if at all, by somebody else, later. Of 14 descriptions that state both a material and a route, 14 were matched and 0 were declined. Of 11 that should be declined — no route stated, or a material this library recognises and will not price — 11 were. A decline costs a client one question; it is not counted against the matcher here, because a single accuracy figure would let refusals be traded for wrong commits and reported as an improvement.
Matched when it should
100%
14 of the descriptions stating a material and a route.
Declined when it should
100%
11 of the descriptions that do not determine a factor.
Repeatable
Yes
Deterministic scoring over a fixed table. No model is called, so an audit cannot change on re-run.
Who labelled this, and why that matters
- Labelled by:
- The authors of this codebase, against the factor library's own definitions.
- Not labelled by:
- An independent panel of practitioners. That is a real limitation: the person who writes a matcher will, without meaning to, write descriptions it handles.
- What was done about it:
- The set is built adversarially to itself — it is weighted towards descriptions that should be REFUSED, including materials the library recognises but cannot price, and one line that names no material at all.
- If you disagree:
- It lives in source at lib/matchQuality/labels.ts. Anyone who disagrees with a label can say so against a specific row, which is the reason for shipping it rather than a figure.
Every description in the set
Including the ones it declines. A declined description is not a failure — it is the matcher asking a question instead of choosing the answer on your behalf, and the question has a named shortlist behind it.
recycled HDPE bottle body
MatchedMatched hdpe-closed, over hdpe-primary, by 0.700.
Material and route both stated. The straightforward case.
virgin PP closure
MatchedMatched pp-primary, over pp-closed, by 0.700.
'Virgin' is the primary route said a different way.
primary aluminium foil lid
MatchedMatched aluminium-primary, over aluminium-closed, by 0.700.
Route word adjacent to the material rather than before it.
recycled board carton
MatchedMatched board-closed, over board-primary, by 0.700.
Fibre, where 'board' and 'paper' are separate factors.
virgin steel can
MatchedMatched steel-primary, over steel-closed, by 0.700.
Steel, which also has a scrap_metal and a construction_metals sibling.
recycled glass jar
MatchedMatched glass-closed, over glass-primary, by 0.700.
primary PET preform
MatchedMatched pet-primary, over pet-closed, by 0.700.
Trade term ('preform') rather than a product noun.
post-consumer recycled PET bottle
MatchedMatched pet-closed, over pet-primary, by 0.700.
'Post-consumer' as the route word instead of 'recycled'.
virgin LDPE shrink film
Declined, correctlyThe description names a specific polymer AND a generic film category, which this library prices separately. Nothing here ranks 'LDPE' above 'plastic film, average' — that would need a material taxonomy the library does not have — so the matcher reports a family ambiguity with both candidates named rather than picking one.
The FAMILY ambiguity, and the row that corrected a false claim. It was first labelled as a commit, on the assumption that naming the polymer settles it. Running the set showed the matcher declining at a margin of 0.117 — a value the codebase had recorded as impossible, having measured bimodality over eleven hand-picked descriptions that contained no family cases. The matcher was right and both the label and the note in match.ts were wrong.
recycled PVC profile
MatchedMatched pvc-closed, over pvc-primary, by 0.700.
primary PS tray
MatchedMatched ps-primary, over ps-closed, by 0.700.
virgin plasterboard
MatchedMatched plasterboard-primary, over plasterboard-closed, by 0.700.
Construction material in a mostly-packaging library.
reused timber joist
MatchedMatched wood-reused, over wood-closed, by 0.300.
The REUSED role, which only three materials carry.
recycled aggregate sub-base
MatchedMatched aggregates-closed, over aggregates-reused, by 0.300.
Matched through a synonym ('aggregate') rather than the factor key.
primary asphalt surfacing
MatchedMatched asphalt-primary, over asphalt-reused, by 0.300.
28 g HDPE bottle body
Declined, correctlyThe material is clear and the sourcing route is not. Virgin and recycled HDPE are different factors, so committing would be choosing the answer on the client's behalf.
The archetypal ambiguity, and the one a fast matcher gets wrong.
PP closure
Declined, correctlyMaterial stated, route not. Same reason as the HDPE row.
aluminium foil lid
Declined, correctlyPrimary and recycled aluminium differ by more than most design decisions, so a guess here is larger than the thing being decided.
The material where guessing costs most.
steel bracket
Declined, correctlyRoute not stated.
glass bottle
Declined, correctlyRoute not stated.
cardboard outer
Declined, correctlyRoute not stated, and 'cardboard' does not distinguish board from paper.
Ambiguous on two axes at once.
recycled EPDM gasket
Declined, correctlyThe route is stated but the library holds no elastomer factor. Recognising a material it cannot price and saying so is different from not recognising it.
Route present, material absent — tests that a route word alone cannot force a match.
virgin silicone seal
Declined, correctlyNo silicone factor in this library.
primary carbon black N330
Declined, correctlyNo rubber-chemical factors in this library.
Checks that a route word next to an unknown material does not produce a commit.
bespoke reception feature
Declined, correctlyNames no material at all — a description of an item, not of what it is made of.
The real-schedule case: a line that reads like a product and specifies nothing.
Why wrong commits are the headline
A decline costs you one question. It is visible, it is recoverable, and somebody is asked. A wrong commit produces a number indistinguishable from a right one, nobody is asked anything, and it is found — if at all — by somebody else, later.
A single accuracy figure averages the two, which lets a vendor trade refusals for wrong commits and report the result as an improvement. So the two are reported separately here, and the matcher is not penalised for declining.