Textile DPP · schema 2.1 · working prototype
DECAVERA checks DPP completeness and performance first, then tests whether the evidence supports an attribution claim. Unsupported claims are not promoted as conclusions.
HVJ-001 · INCOMPLETE: two materials totalling 95%, so one binary section fails. Recycled polyester evidence is absent. Four stages measured.
Not guidance the system tries to follow, but properties the code enforces on every request.
HVJ-001 is our reference product. Tap any point to see the fields it carries and the readiness section it decides.
Fibre composition passes only between 99.5% and 100.5%. Five missing percent is the whole reason this jacket sits at 87.5% instead of 100.
Photometric performance falls with every wash. Use, repair and maintenance counts as complete only when all three fields carry a real instruction, not a placeholder.
Reuse, resale, disassembly and refurbishment need all four fields. Hardware that cannot come out without cutting the shell turns a full answer into a hollow one.
End-of-life needs both disposal and recycling text. Mono-material construction is the rare case where the honest answer is short, and we still record which stage established it.
Economic operators and compliance documents are separate sections. An empty list is not a low score. It is a failed section, and we say so rather than averaging it away.
A hi-vis jacket is not made of hi-vis. It is shell knit, retroreflective tape, laminate and rib stock: four mills, four evidence trails, scored separately before anything rolls up.
Rated 500 cd/lx/m² new, and it must stay above 100 after twenty-five washes. Pick a wash count. This is a durability claim that today lives in a PDF nobody reads.
Readiness is completeness, not compliance. Each section either has what it needs or it does not, and the score is simply how many of the eight passed.
A language model will happily turn a ranking into a cause. We will not let it. Each claim declares which of four grades it earned, and the payload caps the ceiling before the model ever sees the data.
The validator rejects any claim above the ceiling, any unknown evidence ID, and any intervention-supported claim without verified before-and-after evidence.
Stage interventions on the left, the metrics they touch in the middle, the outcomes you are judged on at the right. Nothing moves until you apply one.
Apply an intervention and watch the effect propagate. The second one is deliberately smaller than the first, because the model shows diminishing returns instead of adding two brochures together.
Backend is authoritative. Expected benefits return as PROJECTED_NOT_ACTUAL until a snapshot is captured.
Recycled content looks like it drives defects. Separate the mills and the effect changes sign, because the mills quoting cheapest on recycled orders run the oldest machines.
The mill was doing the damage. The recycled yarn was taking the blame.
We run that separation for every claim we make, and print the variables adjusted for, so an auditor can repeat the work.
Adjustment set usedMill identity · Machine age · Order volume · Season · Shade depth
A proposed measurement does not change the numbers. It sits as an immutable record with its own hash until a second person verifies it, and every event links to the one before.
Previous and proposed values stored as separate immutable JSON, with changed fields listed and a submission-integrity hash. Operational metrics stay untouched.
The pre-intervention assessment is snapshotted at submission, so the comparison is against what was true at that moment, not a reconstruction.
A different reviewer applies or rejects. Applying supersedes the earlier verified revision and appends a verification event. Rejection preserves the proposal.
The post-intervention snapshot is taken. Only now does a projected benefit become an actual one, and only now can a claim reach intervention-supported.
A single index hides the trade you actually have to make. Each dimension is computed separately and explained, and the weighing is left to you.
Working prototype today: FastAPI, SQLite, a deterministic engine and an evidence-bounded attribution layer. Bring one passport export and we will model one real decision with you.
One product's DPP in whatever shape it is currently in. Incomplete is fine. Incomplete is the point.
Fields to the eight sections, stages to measurements, and an honest list of what is missing before anything is modelled.
A change you are actually considering, returned as a graded claim with the evidence IDs behind it.
Thanks. We will be in touch within two working days.
Enter your credentials to access the Decavera platform.