Skip to content
Meritline

verify

Verify Meritline

Reproducible evidence, not assertions. Every hash and policy below is generated at build time from the files that actually ship and the headers that are actually served. If this page and the product ever disagree, this page is the bug — tell us.

Pinned model and runtime

The embedding model (Xenova/all-MiniLM-L6-v2) and all four ONNX Runtime WASM variants are served from this origin and SHA-256-pinned at build time against @xenova/transformers 2.17.2. The scoring worker checks each file against these hashes before the ONNX session is created.

  • /models/Xenova/all-MiniLM-L6-v2/config.jsonsha256-7135149f7cffa1a573466c6e4d8423ed73b62fd2332c575bf738a0d033f70df7
  • /models/Xenova/all-MiniLM-L6-v2/tokenizer.jsonsha256-da0e79933b9ed51798a3ae27893d3c5fa4a201126cef75586296df9b4d2c62a0
  • /models/Xenova/all-MiniLM-L6-v2/tokenizer_config.jsonsha256-9261e7d79b44c8195c1cada2b453e55b00aeb81e907a6664974b4d7776172ab3
  • /models/Xenova/all-MiniLM-L6-v2/special_tokens_map.jsonsha256-b6d346be366a7d1d48332dbc9fdf3bf8960b5d879522b7799ddba59e76237ee3
  • /models/Xenova/all-MiniLM-L6-v2/onnx/model_quantized.onnxsha256-afdb6f1a0e45b715d0bb9b11772f032c399babd23bfc31fed1c170afc848bdb1
  • /ort/ort-wasm.wasmsha256-bbdcb6b3c7d294577d806077630460be4e13ca35a345acb5e81188e9649fa74a
  • /ort/ort-wasm-simd.wasmsha256-9bd07bababc65f53d061f457233eeae501be7ceb8a2adb9eef52d87fe776d865
  • /ort/ort-wasm-threaded.wasmsha256-2ee2f715093aaacc0344bfb1610313eb49a3474e8afde5e4635f30ddfc7615c8
  • /ort/ort-wasm-simd-threaded.wasmsha256-ac23f2f3cbd519a65a0796f7c79eb34ead4c1f6f31eb06e14ed8a9579d697ef6

Content Security Policy, as served

Read from _headers at build time — the exact policy the browser enforces:

  • default-src 'self'
  • script-src 'self' 'unsafe-inline' 'wasm-unsafe-eval' https://accounts.google.com/gsi/client
  • worker-src 'self' blob:
  • connect-src 'self' https://accounts.google.com/gsi/
  • style-src 'self' 'unsafe-inline' https://accounts.google.com/gsi/style
  • frame-src https://accounts.google.com/gsi/
  • img-src 'self' data:
  • font-src 'self'
  • object-src 'none'
  • base-uri 'none'
  • form-action 'none'
  • frame-ancestors 'none'

connect-src 'self' is the exfiltration boundary: no request can reach another origin, so nothing scored on this device can be sent off it.

What can be reproduced

  • Verified on download and on cache read. Every model and runtime byte is checked against the manifest whether it comes from the network or the local cache. A file that is not in the manifest is treated as a failure, not skipped.
  • No external origin is needed to score. With every external origin blocked at the network level, a full scoring run completes with byte-identical results: Spearman rho = 0.9286 on the 21-CV determinism fixture, unchanged. That is a reproducibility check, not an accuracy figure — the accuracy numbers are in the limitations below, and they are lower. A blocked origin cannot lie.
  • A tampered model is refused, not silently replaced. Change one byte of the model and its hash no longer matches; scoring stops with the reason stated:expected sha256-afdb6f1a…  got sha256-f1cc63b6…

Disclosed limitations

What this page does not claim:

  • Inline scripts are permitted. script-src allows 'unsafe-inline' for the static-export hydration bootstrap (a static site has no server to issue a nonce, and per-build hashes would break on every deploy). The exfiltration boundary, connect-src 'self', still holds: an injected inline script cannot send anything out. Tightening this to per-build script hashes is planned.
  • Single-threaded runtime. COEP is not set, so ONNX Runtime runs single-threaded (numThreads = 1) with no SharedArrayBuffer — measured at about 249 ms per CV on an idle machine.
  • Two integrity paths are unit-proven, not browser-tested end-to-end. Cache-poison eviction and the unknown-file rule are proven by unit test at the decision boundary; a full in-browser cache-poisoning test is on the backlog.
  • Accuracy is early, and the honest numbers are low. On a 252-CV corpus across 7 job descriptions: Recall@20 = 0.807 (95% CI 0.714–0.850), Spearman rho = 0.708 (0.650–0.747), Precision@5 = 0.257 (0.143–0.486). That rho fails the 0.90 floor we set ourselves, and we have not moved the floor to make it pass. The reference ranking is constructed by the corpus generator, not written by human annotators, so it measures agreement with a specification rather than with hiring judgement — the benchmark can be wrong in ways human annotation would catch. Kendall’s W is pending human annotation. Treat the ranking as decision support, not a verdict.
  • The no-third-party-request guarantee is machine-checked. connect-src 'self' is backed by an automated regression test that runs the full pipeline with every external origin blocked and fails if any request leaves this origin. Reproducing it needs a local build plus the session secret — documented in the repository README.

How the ranking works is in the methodology.