@miller-tech/uap 1.52.0 → 1.53.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/README.md CHANGED
@@ -121,7 +121,19 @@ UAP installs hooks into your agent harness, then mediates every tool call throug
121
121
 
122
122
  ## Benchmarks
123
123
 
124
- Terminal-Bench 2.0, 12 representative tasks, UAP-on vs. baseline:
124
+ The honest, controlled result (paired A/B — same model, tasks, and seeds,
125
+ toggling only UAP, with confidence intervals): **UAP's accuracy lift depends on
126
+ whether the base agent already self-verifies.**
127
+
128
+ | Baseline | UAP accuracy lift | |
129
+ |---|---|---|
130
+ | Agentic harness (self-tests) | **~0pp** (CI spans 0) | overhead only — value is efficiency/coordination |
131
+ | Non-agentic single-shot model | **+20pp** (78%→98%, 95% CI [+8,+32], p=0.008) | gate loop repairs edge-case bugs |
132
+
133
+ Run it yourself: `uap bench paired --adapter raw --suite benchmarks/suites/real-gate-gated`.
134
+ Full analysis: **[docs/benchmarks/PAIRED_FINDINGS.md](docs/benchmarks/PAIRED_FINDINGS.md)**.
135
+
136
+ <details><summary>Earlier uncontrolled Terminal-Bench numbers (confounded — see TBench Investigation)</summary>
125
137
 
126
138
  | Metric | Baseline | With UAP | Δ |
127
139
  |---|---|---|---|
@@ -130,6 +142,8 @@ Terminal-Bench 2.0, 12 representative tasks, UAP-on vs. baseline:
130
142
  | Errors per task | 1.17 | 0.42 | **−68%** |
131
143
  | Wall-clock (total) | 618s | 266s | **−57%** |
132
144
 
145
+ </details>
146
+
133
147
  Methodology, raw runs, and cost analysis: **[docs/benchmarks/](docs/benchmarks/)**.
134
148
 
135
149
  ---