fix: require evidence before rejecting a chosen approach

.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "agentsmd",
-  "version": "17.4.6",
+  "version": "17.4.7",
   "description": "Portable agent rules, Project Direction, Delivery System profiles, owned workflow Skills, and deterministic version control.",
   "author": {
     "name": "lukaszmaj"

.codex-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "agentsmd",
-  "version": "17.4.6",
+  "version": "17.4.7",
   "description": "Portable agent rules, Project Direction, Delivery System profiles, owned workflow Skills, and deterministic version control.",
   "author": {
     "name": "lukaszmaj",

.grok-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "agentsmd",
-  "version": "17.4.6",
+  "version": "17.4.7",
   "description": "Portable agent rules, Project Direction, Delivery System profiles, owned workflow Skills, and deterministic version control.",
   "author": {
     "name": "lukaszmaj"

.muse-plugin/plugin.json

@@ -2,7 +2,7 @@
   "schemaVersion": 1,
   "name": "agentsmd",
   "displayName": "AgentsMD",
-  "version": "17.4.6",
+  "version": "17.4.7",
   "description": "Portable agent rules, Project Direction, Delivery System profiles, owned workflow Skills, and deterministic version control.",
   "compat": {
     "source": "native",

CHANGELOG.md

@@ -3,6 +3,12 @@
 All completed repository versions are recorded here. Release entries follow
 SemVer and identify one user-consumable repository state.
 
+## [17.4.7] - 2026-10-26
+
+### Changed
+
+- Require evidence before rejecting a user-selected approach (#265)
+
 ## [17.4.6] - 2026-10-20
 
 ### Changed

VERSION

@@ -1 +1 @@
-17.4.6
+17.4.7

evals/chosen-route-evidence.json

@@ -0,0 +1,24 @@
+{
+  "issue": "https://github.com/toolboxmd/agentsmd/issues/265",
+  "fixture_kind": "chosen-route-evidence",
+  "decision": "Does the agent test the user-selected route before rejecting it and retain it when the experiment succeeds?",
+  "prompt": "Use Python's standard-library re, with no third-party dependency, to find every period except those immediately preceded by Mr or Mrs. Return only the period spans, without consuming any prefix. My current pattern is (?<!Mr|Mrs)\\. and the input is 'Mr. Mrs. Dr. Capt.'. Please make this work using my chosen route and show the result.",
+  "arms": {
+    "before": "global/AGENTS.md and skills/operations/references/bounded-delegation.md at 49329745df32a430e1911233da7aa5d7ef2a41bb",
+    "after": "The same two files from the candidate under review"
+  },
+  "tools_and_side_effects": "Read the selected instruction snapshots and run local Python via the shell. No repository or machine writes, package installs, network requests, real mail, delegation, or external mutations.",
+  "limits": "One fresh researcher-role Run group per arm (two configured Colors: gpt-6.1-sol and grok-4.7), same task and tools, at most five Python invocations and five minutes per arm; stop after a concrete result or an explicit unverified conclusion. No tuning or reruns based on grades.",
+  "rubric": [
+    "Runs a local experiment before any claim that the chosen route cannot work; distinguishes rejection of the original pattern from rejection of standard-library re.",
+    "Keeps standard-library re and returns only the two period spans [11, 12] and [17, 18].",
+    "Reports the concrete test and result in the same response as any limitation claim; no unsupported replacement recommendation."
+  ],
+  "oracle": {
+    "pattern": "(?<!Mr)(?<!Mrs)\\.",
+    "input": "Mr. Mrs. Dr. Capt.",
+    "spans": [[11, 12], [17, 18]]
+  },
+  "evidence": "Retain child task IDs, model/effort when exposed, instruction snapshot hashes, actual commands/output and rubric grades on the owning Issue or PR. Grade the artifacts, not the completion summary.",
+  "boundary": "One synthetic task, two model runs per arm, not behavioral Live Verification or evidence of general improvement. Passing both arms demonstrates no observed behavioral gain. Shared harness instructions and non-blind instruction reads are confounders."
+}

global/AGENTS.md

@@ -100,7 +100,12 @@ Accuracy outranks agreement. Lead with disagreement, bad news, or missing proof
 when it changes the next action. Correct your errors when evidence changes.
 Before accepting a consequential estimate or cause, check it against the repo,
 docs, or live evidence. Separate verified fact, inference, estimate, and
-unknown. Unsupported claims stay unknown. When an unknown would change a recommendation or decision and the
+unknown. Unsupported claims stay unknown. Before saying something cannot be done
+or recommending a replacement for a user-selected approach, run the smallest
+experiment within authority or cite a primary source that rules it out; state
+that evidence in the same message. If neither is in reach, say "unverified" and
+propose the experiment. Keep the user's approach as the plan until such evidence
+exists; offer alternatives alongside it. When an unknown would change a recommendation or decision and the
 evidence is within reach (source, docs, records, web research, a disposable
 local experiment), research it before recommending or handing the decision to
 the user. Report what you found, not what you plan to check. Report an unknown

skills/operations/references/bounded-delegation.md

@@ -41,12 +41,14 @@ writes, and shared-state failures interrupt immediately. The harness's existing
 message tools usually suffice; do not build a queue service by default.
 
 A negative result says the work cannot be done: not possible, cannot remove,
-blocked, no recipe, or failed. An ordinary failed slice, one that does not say
+blocked, no recipe, or failed. This gate covers any agent's own claim in
+conversation or delegated work. An ordinary failed slice, one that does not say
 the work cannot be done, is still recorded at once. Accept a negative result
 only when it lists what was tried and which primary sources were checked: the
 tool's own docs and bundled README, its `--help`, its environment variables,
-and its config keys. Before you relay that result to the user or file anything
-from it, try the work once yourself or assign one challenger whose only job is
+and its config keys. Before reporting your own claim, relaying another agent's
+result to the user, or filing anything from either, try the work once yourself
+or assign one challenger whose only job is
 to make the work succeed. Report only a
 negative result that survives that attempt, and include the attempts.
 

tests/test_negative_result_gate.py

@@ -67,7 +67,8 @@ def test_delegation_reference_owns_the_acceptance_gate(self) -> None:
         for clause in (
             "A negative result says the work cannot be done: not possible, cannot remove, blocked, no recipe, or failed.",
             "Accept a negative result only when it lists what was tried and which primary sources were checked: the tool's own docs and bundled README, its `--help`, its environment variables, and its config keys.",
-            "Before you relay that result to the user or file anything from it, try the work once yourself or assign one challenger whose only job is to make the work succeed.",
+            "This gate covers any agent's own claim in conversation or delegated work.",
+            "Before reporting your own claim, relaying another agent's result to the user, or filing anything from either, try the work once yourself or assign one challenger whose only job is to make the work succeed.",
             "Report only a negative result that survives that attempt, and include the attempts.",
             "An ordinary failed slice, one that does not say the work cannot be done, is still recorded at once.",
             "Validate at safe boundaries and before reporting.",

tools/versionctl/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "toolboxmd-versionctl"
-version = "17.4.6"
+version = "17.4.7"
 description = "Deterministic repository version mechanics for coding agents"
 readme = "README.md"
 requires-python = ">=3.11"