Require tile quality evidence for training
This commit is contained in:
@@ -131,6 +131,9 @@ the versioned JSON argv command. Invalid state and bounded launch exhaustion
|
||||
fail closed.
|
||||
The orchestrator refuses to start unless every automated frozen-dataset gate
|
||||
passes and the corpus contains zero blank/low-variance positive tiles. The
|
||||
separate train tile-quality audit is a required checksummed loop input; missing
|
||||
invalid-label, missing-label-file or low-variance evidence fails closed rather
|
||||
than being interpreted as zero. The
|
||||
audit status may remain `needs_human_review` while training and objective
|
||||
evaluation continue: final human sign-off is deliberately the last gate and
|
||||
can never be interpreted as model promotion approval in advance.
|
||||
|
||||
@@ -11,6 +11,10 @@
|
||||
and nine Walloon dense/rural/architecture or difficult-negative samples.
|
||||
Added mode-0600 session-token-file support to the official provisioner so
|
||||
authenticated Tower acquisition does not expose session material in argv.
|
||||
- Closed a fail-open evidence gap between corpus and tile audits. The training
|
||||
loop now requires and hashes the train tile-quality report, and refuses
|
||||
missing counters instead of treating absent invalid/blank-label evidence as
|
||||
zero.
|
||||
|
||||
## 2026-07-27 - Guest demo and product professionalization
|
||||
|
||||
|
||||
Reference in New Issue
Block a user