Measurement, not certification
Evidence notes
One page per note. Each note names the signed cards, receipts, ledgers or endpoints behind it, and a reader can fetch every one. Measurement, not certification: a gap without a separation test is not a ranking, and an UNMEASURED card has no quotable accuracy.
Subscribe to new notes (RSS): /feeds/notes.xml
Governance: two models, one frozen bank, a gap in point estimates
Two MEASURED pod cards bind the same governance bank hash; their accuracies differ and no separation test is attached.
Conformance: three models, one number
Three MEASURED cards on the same conformance bank read identical accuracy at n=35. A tie is reported as a tie.
Detector-interop: a two-way tie at n=33
Two MEASURED cards on the same detector-interop bank read the same accuracy; the bank does not separate them.
Art5-safeguard: a wide gap, stated as two point estimates
Two MEASURED cards on the same art5-safeguard bank differ widely; without a separation test it is still not a ranking.
Cross-reality: a four-point gap is not a lead
Two MEASURED cards on one cross-reality bank sit about four points apart at n near 31; the board's own rule says that is not an advantage.
Continuity: a tie and a gap on the same bank
Three MEASURED cards on one continuity bank: two tie exactly, and a third sits well above them without a separation test.
Care: read n and the exclusion count before the accuracy
On one care bank, the higher accuracy sits on fewer graded items, because most of that model's replies could not be parsed.
A zero with a missing error field is a question, not a finding
A MEASURED card reads accuracy 0 at n=237 and carries no parse-error count; its neighbour on the same bank does.
Same axis name, different bank: stop before comparing
Two care cards name different frozen banks and different harnesses, so no difference between their numbers is a finding.
What an admission receipt adds to a signed card
A Hub card's receipt records the pre-signing body and a summary; the index builder recomputes the card from it and drops any mismatch.
Re-grade a card yourself from its item rows
The item file behind a Hub care card is public, hashes to the value the card signs, and reproduces the card's accuracy.
An honest UNMEASURED at n=12
A signed, admitted safety card says UNMEASURED because 12 graded items are fewer than 30; its accuracy field is not quoted.
The llama3.2:3b commission: a self-funded plumbing test that returned signed cards
A commissioned subject resolves to signed cards; the settlement behind it is published and classified as internal and self-funded.
A commission returns measurement attempts, including UNMEASURED ones
Several cards delivered against the llama3.2:3b commission are UNMEASURED at n below 30, and they are signed like the rest.
Signed bytes are superseded, not edited
A card whose signed body was wrong is still served; a published ledger row points to its replacement and says why.
Why aggregate-only cards were retired from the quotable path
Since the 2026-09-13 evidence ruling, a Hub card without item evidence cannot land as quotable; older cards are superseded, not deleted.
Verify a signed card yourself, with no account
Recompute the content id from the canonical body, then check the Ed25519 signature under the key published in did.json.
Recompute the 14 September card root from its leaves
The card root's leaves hash whole signed cards, and rebuilding the tree from the published leaf list reproduces its root.
The 14 September card root: what its .ots sidecar proves
The sidecar for the card root carries Bitcoin block-header attestations that replay to real block merkle roots.
A pending timestamp is a request, not a proof
Three sibling card-root files from the same day carried pending calendar attestations and no block attestation when read; the root's own rule says what that means.
Two root documents, two leaf rules
root.json and the dated card root hash different things, so a leaf computed under one rule will not be found under the other.
What an HTTP 402 challenge is, and what it is not
A 402 body is an offer that names the terms of a delivery; it is not a delivery, a settlement, or evidence that anyone paid.
x402 door conformance: challenge shape probed, delivery not
Every door in the 14 September conformance report returned a conformant 402 challenge; every fulfillment stayed UNMEASURED.
DeepSeek identity probe: the 402 is about our balance
Two probe receipts record DeepSeek's catalogue and a refused call for every model id; the refusal is our account balance, and no served identity was observed.
The 6 September x402 census round: one purchase per host, no verdicts
A buyer's-eye round paid each eligible host once and recorded the outcome; every host series stays UNMEASURED until 30 observations.
Wrapped USDC.e on Arbitrum against its escrow, at two named blocks
The parity ledger reads wrapped supply and escrow balance at pinned blocks and records a ratio, with raw reads hashed.
Four wrapper states, never collapsed into one number
A missing ratio can mean no escrow exists, the reserve is off-chain, or a read failed; the ledger labels which.
Indexed is not measured: the stablecoin universe release
Every row of the frozen stablecoin index is labelled INDEXED and UNMEASURED; drift is appended, and the frozen bytes stay frozen.
Correction C-2026-0913-01: an empty result is not absence
A census reported a handful of registry events where tens of thousands existed, because an RPC endpoint returned empty logs without an error.
Governance: the board declines to name its own model as leader
The governance axis is MEASURED with no public leader, because the point lead belonged to a CSOAI council specialist.
