fix: restore accuracy-first route admission

README.md

@@ -107,8 +107,10 @@ model/GPU uniqueness response, the entire submission cycle pauses fail-closed.
 For each compatible model/GPU pair, the worker also ranks the GPU's supported
 frameworks using ModelHub's public aggregate `modelCount` and `successCount` data,
 then blends in the worker's own GPU+framework outcomes with a capped weight.
-Only frameworks with at least 300 public samples and a safe Wilson lower bound
-are eligible. Framework statistics refresh
+Only frameworks with at least 300 public samples and a Wilson lower bound of at
+least 20% are eligible. GPU pools normally require a 15% lower bound; a
+low-sample pool can remain available only when it has a recent success and a
+framework that passes the stricter framework gate. Framework statistics refresh
 every 6 hours, so they do not add per-model API traffic. Newly published
 frameworks are discovered automatically, but receive no novelty bonus: they can
 win only when their confidence-adjusted success score beats the best incumbent by
@@ -124,7 +126,10 @@ evidence on every scan; accepted route counters and lifetime sufficient statisti
 are stored in `.modelhub_state/routing_intelligence.json`.
 
 Five consecutive local failures open a 12-hour GPU/framework circuit breaker.
-A sub-20% success rate over the latest 20 terminal tasks opens a 6-hour breaker.
+Five or more recent terminal results below 10% success open the same breaker,
+including ambiguous runtime failures that are unsafe to keep feeding. A sub-20%
+attributable success rate over the latest 20 terminal tasks opens a 6-hour
+breaker.
 Platform/infrastructure failures are excluded from long-term compatibility rates
 and model/profile breakers. Three consecutive platform failures on a GPU/framework
 instead open a short 30-minute breaker, so a temporary broken runner or lack of an
@@ -481,6 +486,18 @@ calibration each Monday 10:00-10:59 Asia/Shanghai. During that window only,
 the worker may borrow the five recent-model positions to measure a new limit;
 ordinary capacity-minus-five admission resumes afterward.
 
+Version `2026.09.22.1` restores accuracy-first admission while retaining bounded
+capacity recovery. Any existing result for an exact model/GPU pair, including a
+failed result, closes that pair; a different framework no longer reopens it.
+Each model is limited to six GPU routes across poll cycles and at most two local
+routes may await feedback simultaneously. Lower-ranked GPUs are still checked,
+in score order, only when the preferred six cannot produce a safe candidate.
+Download-ranked popularity fallback is removed: discovery expands by update
+window instead. Public framework/GPU confidence floors and an all-terminal
+failure circuit prevent high-volume low-quality routes from winning on relative
+rank alone. Community-confirmed model successes are used only as a tie-breaker
+inside the same confidence band.
+
 ## Deploy
 
 Create a tag and submit the repository URL plus tag in "我的适配智能体".