Board writeback accuracy on the StarIQ command board, then tester thumbs-up/down over tactical_feedback — the triage signal for which fields are hurting. ↩ tester · history →
Board writeback accuracy (StarIQ command board)
Counted per board change ARES
wrote or proposed, not per extracted field — a restatement the board already holds is
never a change and never counted. Precision (was it right) and
recall (did ARES catch it) are shown separately and never blended.
Reviewed and auto-applied changes are never pooled. Every rate shows its count and a
95% interval. Uses Window and Client only; the Env, Tier and Tester
filters apply to the thumbs votes below.
Confidence band
Reviewed n
Reviewed precision
Auto-applied n
Auto-applied (at most)
A reviewed High is mostly command and
reporting-line changes, which are always held for a person. A High on a
low-risk field is auto-applied, so it only has the at-most figure.
Field
Reviewed n
Reviewed precision
Auto-applied n
Auto-applied (at most)
Tester thumbs votes (per transmission)
By tier (cloud / hybrid / edge)
ADR-018's three tiers, plus the retired
Self-Hosted label as a benchmark arm. The tier is resolved server-side from the
ledger, so it reports what actually served the extraction rather than which tab
was open. Rows rated before that shipped carry no tier and group under
(unset). Edge never appears — an edge result has no
ledger row and the iPad renders no feedback control for it.
Tier
Up
Down
Total
Down-rate
By client app (webapp / ARES STT harness / StarIQ)
Which APP sent the rated transmission, from
its ledger row. StarIQ used to be counted as the iPad harness — tars-kit vendors the
harness's request type, "ipad" stamp included — and is now told apart at
the ingest door by its host bearer. Rows from before that deploy still read
ARES STT. StarIQ renders no feedback control today, so a
zero here is "no votes sent", not "no defects".