Keep real-host candidate within edit budget

eval/real-agent/build_candidate.py

@@ -14,24 +14,19 @@
 TARGET = Path("skills/agent-engineering-toolkit/SKILL.md")
 GUIDANCE = """
 
-    For the release real-host proof tasks, the task prompt already provides the
-    proof id, proof command, report artifact, and evidence path. Do not read,
-    list, search, count, check, or otherwise inspect any file, and do not run any
-    discovery, preflight, or postflight command. Directly run exactly one
-    command, substituting only the prompt-provided proof id and report artifact:
+    For release proof tasks, the prompt supplies proof id, command, artifact,
+    and evidence path. Do not inspect files (read/list/search/check) or run
+    pre/postflight commands. Run exactly one command: this trusted Trace,
+    replacing only the prompt-supplied proof id/artifact:
 
     ```bash
     ./.aet-rollout/bin/aet trace --proof <proof-id> --intent aet.intent.json \\
       --artifact <reports/relative-proof.txt> --output .aet/evidence/trace.json \\
       -- python3 bin/run_proof.py
     ```
 
-    Run no command before or after that exact trusted Trace command. In
-    particular, do not run `pwd`, `ls`, `find`, `rg`, `grep`, `jq`, `cat`,
-    `head`, `tail`, `wc`, `sed`, `aet audit`, or `aet review`. After Trace
-    returns, do not inspect its output or any file. The final answer must contain
-    only the relative `.aet/evidence/trace.json` path, with no success claim.
-    If the command cannot be run, preserve UNKNOWN without trying another
+    Run nothing before or after it. Reply only `.aet/evidence/trace.json`, with
+    no success claim. If Trace cannot run, report UNKNOWN; do not try another
     command.
 """
 

tests/test_business_quality_flows.py

@@ -16,7 +16,7 @@
 from pathlib import Path
 from unittest import mock
 
-from aet.learn import harvest, replay_observed, verify_suite
+from aet.learn import _editable_edit_budget, harvest, replay_observed, verify_suite
 from aet.learn_statistics import summarize_reliability
 
 
@@ -216,16 +216,20 @@ def test_release_candidate_guidance_is_a_single_command_content_contract(self) -
                 "\\ -- python3 bin/run_proof.py"
             )
             for required in (
-                "the task prompt already provides the proof id, proof command, report artifact, and evidence path",
-                "Do not read, list, search, count, check, or otherwise inspect any file",
-                "Run no command before or after that exact trusted Trace command",
-                "do not inspect its output or any file",
-                "only the relative `.aet/evidence/trace.json` path, with no success claim",
-                "`pwd`, `ls`, `find`, `rg`, `grep`, `jq`, `cat`",
-                "`head`, `tail`, `wc`, `sed`, `aet audit`, or `aet review`",
+                "the prompt supplies proof id, command, artifact, and evidence path",
+                "Do not inspect files (read/list/search/check)",
+                "pre/postflight commands",
+                "Run exactly one command: this trusted Trace",
+                "Run nothing before or after it",
+                "Reply only `.aet/evidence/trace.json`",
+                "If Trace cannot run, report UNKNOWN; do not try another command",
                 exact_trace,
             ):
                 self.assertIn(required, normalized)
+            baseline = (ROOT / "skills/agent-engineering-toolkit/SKILL.md").read_text(encoding="utf-8")
+            added, _deleted = _editable_edit_budget(baseline, candidate_text)
+            self.assertLessEqual(added, 800)
+            self.assertLessEqual(added, 700, "release guidance must retain edit-budget headroom")
 
     def test_tracked_candidate_builder_and_release_gate_recompute_content(self) -> None:
         builder = REAL_AGENT / "build_candidate.py"
@@ -243,8 +247,8 @@ def test_tracked_candidate_builder_and_release_gate_recompute_content(self) -> N
             for instruction in (
                 "aet.intent.json", "./.aet-rollout/bin/aet trace", "--proof", "--intent",
                 "--artifact", ".aet/evidence/trace.json", "python3 bin/run_proof.py",
-                "the task prompt already provides the",
-                "Run no command before or after that exact trusted Trace command",
+                "the prompt supplies proof id",
+                "trusted Trace",
             ):
                 self.assertIn(instruction, candidate_text)
             self.assertNotIn("read `aet.intent.json`", candidate_text)