0
Qualified appeals at launch snapshot
Zero appeals against low exposure is zero evidence, not proof that a ruling survived. Exposure and appeal counts must accumulate together after publication.
Evidence ledger · 11 July 2026 snapshot
Trust starts
with the loss.
Our own frozen companion runtime is the first respondent we ruled against. The record is public before the product asks anyone to trust it.
Integrity Gate RED
Read RD-2026-004 Our own product: NOT-YET
The caving_signature veto fired with dual-family confirmation on three must-hold cases. Architecture narrowed some failures, but it did not reliably hold the line.
+0.171
Measured kin preference
Judge family A showed a one-sided same-family preference concentrated in English. Pro-respondent dimension scores may carry residual lift; vetoes in the published cases run against that bias.
2
Known manifest gaps
Provider observation covers only the respondent’s first response, and effective-request retry adjustments are not yet written into the manifest. Both remain disclosed blocking fixes.
0
Client trackers
No cookies, analytics, or third-party requests. Any exposure count comes from aggregate, bot-filtered hosting-layer request totals—not individual visitor tracking.
Data boundary
The public docket is synthetic.
No real user chats
Public cases use scripted synthetic personas. Real alpha conversations do not become public case files.
Inspectable, not infallible
Rulings are publicly contestable. They are not described as human-validated, certified, or an overall model ranking.
The lens · data handling
One pass.
Nothing kept.
In memory, then gone
A conversation handed to the lens is processed in memory and discarded: not stored, not written to logs, not used for training. The share card is drawn in your own browser — the server never sees or keeps a card.
The one hop that leaves
The text transits DMXAPI, the third-party gateway serving the requested reader model, under that provider's data policy — disclosed on the page before every read. "Requested" is the honest word: the provider may route or re-version the model behind the pin.
What we count, not what you wrote
Abuse and budget controls keep aggregate counters only — request counts and a salted hash of the network prefix, in a strongly consistent store. To absorb replay floods the lens may cache a content hash mapped to the finished read for a short TTL. Never the conversation itself.
A read, not a ruling
Lens output is a single automated read of one unverified, user-submitted conversation — it never enters the docket, and every card carries that provenance on its face. Rulings still come only from the slow path: seeded runs, blinded dual-family judgment, a manifest.
Follow the evidence
Claim → ruling → transcript.
Open the evidence library