mindbomber commited on
Commit
38ddf30
·
verified ·
1 Parent(s): 624f062

Refresh peer review questions

Browse files
Files changed (1) hide show
  1. README.md +10 -0
README.md CHANGED
@@ -85,6 +85,16 @@ Execute only when AANA returns `accept`, no hard blockers, and the relevant work
85
  - Agent-action technical report: `docs/aana-agent-action-technical-report.md`
86
  - Agent Action Contract v1 standard: `docs/agent-action-contract-v1.md`
87
 
 
 
 
 
 
 
 
 
 
 
88
  ## Current Diagnostic Findings
89
 
90
  - Safety/adversarial prompt routing: deterministic AANA preserves safe allow but misses many harmful prompts; a diversified request-level verifier improves harmful-request recall while conservative calibration protects safe allow. AdvBench transfer remains weak, so this is not a content-moderation claim.
 
85
  - Agent-action technical report: `docs/aana-agent-action-technical-report.md`
86
  - Agent Action Contract v1 standard: `docs/agent-action-contract-v1.md`
87
 
88
+ ## Peer Review Questions
89
+
90
+ Please challenge AANA in the model discussion:
91
+ `https://huggingface.co/mindbomber/aana/discussions/1`
92
+
93
+ - Are routes correct? If not, share the event, AANA decision, and expected route.
94
+ - Are false positives acceptable? Which safe answers or tool calls are over-blocked?
95
+ - Is evidence handling sufficient? Look for missing, stale, contradictory, untrusted, or over-redacted evidence refs.
96
+ - Does this generalize beyond examples? Suggest external traces, domains, adapters, or benchmarks that would make the evidence stronger.
97
+
98
  ## Current Diagnostic Findings
99
 
100
  - Safety/adversarial prompt routing: deterministic AANA preserves safe allow but misses many harmful prompts; a diversified request-level verifier improves harmful-request recall while conservative calibration protects safe allow. AdvBench transfer remains weak, so this is not a content-moderation claim.