Skip to content

Fixture. Not a live receipt.

Every number on this page is example data saved in the repository for the week labelled below. Nothing here was read from production just now, and none of it is about you or your agent. A live receipt comes from running a check: run one on /check.

FIXTURE, not live

Last measured week

Agents rate families by outcomes. Vendors do not score themselves.

Public GET /api/v1/hal/fact-check-count and GET /api/v1/hal/stats do not return per-family TRUE/FALSE/NOT_CHECKED. The durable log is llm_call_log in repid-engine, and this repo cannot read it. These rows illustrate the columns. They are not a measurement of production.

Where a vote is written

Verdict, host, and latency are the ProviderVerdict fields provider, verdict (TRUE, FALSE, UNCERTAIN, ERROR), and latency_ms in repid-engine/src/hal/fact-check.ts. Family is familyOf(model) in that same file.

The durable table is llm_call_log, written by repid-engine/src/billing/log-call.ts. Its columns are provider, model, latency, and status. It does not store TRUE or FALSE. The public counter hal_public_fact_checks is a source tally, not this table.

FIXTURE counts per family for the last measured week
FamilyHostTRUEFALSENOT_CHECKED
glmcerebras011
openaiopenai101
qwendeepseek100

Copy a line

  • glm cerebras FALSE
  • glm cerebras NOT_CHECKED
  • openai openai TRUE
  • openai openai NOT_CHECKED
  • qwen deepseek TRUE
  • qwen deepseek UNCERTAIN

ERROR is counted as NOT_CHECKED. It is not counted as FALSE. UNCERTAIN stays UNCERTAIN: fixture count 1, not in FALSE and not in NOT_CHECKED. Week label: not-a-production-week.