Model finds
The judge extracts line-cited evidence. It never assigns a score.
The Bench · Code measurement
“Is this good code?” is the wrong question.
A throwaway script and a payment boundary can solve the same problem and score nothing alike, because they were built for different things. Facet refracts a file into 14 measured dimensions and reads it against the profile it was written for.
one number infourteen faces outnever a single grade
Profile a fileOpen a sample report
Free while in beta. No card. The standard web profile does not retain your plaintext source. A quick security check needs no account at all.
Building with an AI coding agent? Join the API, CLI, and MCP before-and-after beta.
01
Unmeasured code is not free. Someone is paying for it right now.
Every codebase is optimising for something - speed, safety, throughput, changeability. If nobody chose the trade, the deadline chose it. The documented bill for leaving that unexamined:
$2.41Tpoor software quality cost, US alone, one year (CISQ 2022)
42%of the average developer week lost to maintenance and bad code (Stripe)
45%of AI coding tasks introduced a known vulnerability class (Veracode 2025)
The full case, with receipts - and what it means if you build fast with AI →
02
One number in. Fourteen faces out.
Every other tool hands you a single grade. That grade hides exactly the part you need: the tradeoffs. Facet puts a file on the bench and lets one beam of measurement refract into fourteen dimensions, each measured apart, each with its evidence cited to the line.
03The instrument, operable
Not a grade. A shape you can inspect.
Hover or focus a dimension to see its cited evidence in the sample artifact. The probe pane lights the exact lines the judge read. Poles never net: capability and violation stay separate, always.
authorise_transfer.pypython
written for Security boundaryreads closest to P9 Security-Critical
D1 Runtime performance
D2 Memory efficiency
D3 Readability
D4 Maintainability
D5 Robustness
D6 Security
D7 Portability
D8 Prototype economy
D9 Testability
D10 Observability
D11 Auditability
D12 Concurrency safety
D13 API ergonomics
D14 Resource cost
probeD6 Security · authorise_transfer.py · lines 6-9
basis the account lookup is parameterised; no string-built SQL reaches the driver
capability present, right harm present, left N/A = no surface
13 measured1 N/Ano axis wins them all
04Why trust the read
Most AI reviewers hand you a verdict and ask for faith. Facet is built the other way around.
The model never scores. It extracts line-cited evidence, and deterministic code sets every level. Same code, same profile: test-retest 0.94 on real code, frozen per content hash.
The weights are not our opinion. Where a dimension is weighted, the weights come from external standards: SonarQube rule severities, corroborated by NVD CVSS scores.
We attacked our own scoring. A 40-agent adversarial review tried to refute every security weight. It found real defects. Every one was fixed.
Security fails closed. One detected critical defect caps the security score no matter how much good practice surrounds it.
0.94
test-retest on real code · floor 0.90 enforced
ρ ≈ 0.69
two independent standards, one tier order - provisional
- fabricated rule anchorremoved
- full-tier contradictionre-derived
- XSS below the gatefloored
40-agent review · 0 defects survived
capped
one critical defect sets the ceiling, whatever surrounds it
05See it run
Watch one file refract on the bench.
A short walkthrough: paste a file, watch the beam split into fourteen measured dimensions, and read each one against the profile the code was written for.
The beta question
Does this read feel fair on your real code?
Disagreement is useful when it is precise. Facet asks whether a dimension reads high, low, or about right, so beta feedback becomes calibration data. Free while in beta.
Code decides
Deterministic code sets each level. Same file, same profile, same result.
Profile frames
The same code can be right or wrong depending on what it was built for.
Plaintext is not retained
Managed durable scans temporarily hold an encrypted selected slice, then delete it after completion or expiry.