diff --git a/README.md b/README.md
index fe1b76f..fb64828 100644
--- a/README.md
+++ b/README.md
@@ -46,8 +46,10 @@ Every finding carries one judgment (`evaluator/Evaluator.go`):
 - `trajectory/` — the `Run`/`Step` model, the JSON loader, and `Validate()`
   (structural invariants: ids, timestamps, per-type required fields).
 - `evaluator/` — the `Evaluator` interface and the evaluators:
-  - `tool_calls` — pairs tool calls with results (by tool name, in order) and
-    flags orphans (neutral) and tool errors (bad).
+  - `tool_calls` — pairs each tool result with its call by `toolCallId` when
+    present (authoritative, order-independent), falling back to tool name in FIFO
+    order only when a result carries no id; flags orphans (neutral) and tool
+    errors (bad).
   - `loops` — flags repetition (same tool plus identical input, or same node)
     at least `MaxRepeats` times; catches runaways the pairing check cannot see.
   - `cost_latency` — flags per-step and whole-run cost/latency budget breaches
diff --git a/cmd/trazo/main.go b/cmd/trazo/main.go
index 6f14639..7f7613c 100644
--- a/cmd/trazo/main.go
+++ b/cmd/trazo/main.go
@@ -8,6 +8,7 @@ import (
 	"os"
 	"os/signal"
 	"strings"
+	"syscall"
 	"time"
 
 	"github.com/Cro22/trazo/config"
@@ -141,7 +142,7 @@ func main() {
 
 	// Cancel in-flight evaluation on Ctrl+C (SIGINT) or SIGTERM so a long run,
 	// notably one using the network-bound LLM judge, stops promptly.
-	ctx, stop := signal.NotifyContext(context.Background(), os.Interrupt)
+	ctx, stop := signal.NotifyContext(context.Background(), os.Interrupt, syscall.SIGTERM)
 	defer stop()
 
 	start := time.Now()