evolve: guard the author summary channel against filler echoes

mcp/src/lab/eval/steps/evolve-loop.ts

@@ -73,6 +73,25 @@ function num(v: unknown): number | undefined {
   return typeof v === "number" && Number.isFinite(v) ? v : undefined;
 }
 
+/**
+ * Authors occasionally end schema mode on a bare text turn, echoing filler
+ * ("placeholder", "") into `summary`. Junk there poisons every later
+ * briefing — the EXPLORE directive's "pick an approach that is none of the
+ * above" is only checkable against real summaries — and the report the
+ * human reads. Replace it with an honest marker instead of passing it
+ * through. (The version echo has its own fallback in the gen workflows;
+ * this is the summary-channel counterpart.)
+ */
+const NO_SUMMARY =
+  "(no usable approach summary reported by this generation's author — read this version's YAML diff to see what it changed)";
+function usableSummary(v: unknown): string | undefined {
+  if (typeof v !== "string") return undefined;
+  const t = v.trim();
+  if (t.length < 8) return undefined;
+  if (/^(placeholder|todo|tbd|n\/?a|none|null|summary|unknown)[.!]?$/i.test(t)) return undefined;
+  return t;
+}
+
 function indent(s: string, pad: string): string {
   return s
     .split("\n")
@@ -266,7 +285,7 @@ export default defineStep({
           genRunId: String((journaled["runs"] as AnyRec[] | undefined)?.[0]?.["runId"] ?? ""),
           version: typeof journaled["version"] === "string" ? (journaled["version"] as string) : undefined,
           fitness,
-          summary: typeof journaled["summary"] === "string" ? (journaled["summary"] as string) : undefined,
+          summary: usableSummary(journaled["summary"]) ?? NO_SUMMARY,
           digestText: typeof journaled["digestText"] === "string" ? (journaled["digestText"] as string) : undefined,
           explore: journaled["directive"] === "explore",
         };
@@ -350,7 +369,7 @@ export default defineStep({
         version: typeof out["version"] === "string" ? (out["version"] as string) : undefined,
         fitness,
         allPassCount: num(digest["allPassCount"]),
-        summary: typeof out["summary"] === "string" ? (out["summary"] as string) : undefined,
+        summary: usableSummary(out["summary"]) ?? NO_SUMMARY,
         changes: out["changes"],
         missingSecrets: out["missingSecrets"],
         digestText: typeof digest["text"] === "string" ? (digest["text"] as string) : undefined,

mcp/src/lab/gaia/evolve-smoke.ts

@@ -190,7 +190,9 @@ async function main() {
           output: {
             candidate: input.candidateName,
             version: `v${g + 1}`,
-            summary: `approach ${g}`,
+            // gen 0 echoes filler (the observed schema-mode failure); the
+            // loop must replace it with the honest no-summary marker.
+            summary: g === 0 ? "placeholder" : `approach ${g}`,
             authorCost: 1,
             digest: { fitness: rates[g], text: `digest ${g}`, results: [{ cost: 2 }] },
           },
@@ -224,7 +226,16 @@ async function main() {
     assert.ok(genCalls[1].briefing.includes('the seeded produce workflow "gaia-produce"'));
     // margin 0 semantics: a TIE (0.4 vs baseline 0.4) does not become best
     assert.ok(genCalls[1].briefing.includes("BEST SO FAR: the baseline itself"));
-    console.log("✔ eval/evolve-loop: climbs gaia `fitness`, accuracy naming, margin-0 tie handling");
+    // junk-summary guard: gen 0's "placeholder" echo must NOT reach the next
+    // briefing or the report — both carry the honest no-summary marker.
+    assert.ok(!genCalls[1].briefing.includes("placeholder"));
+    assert.ok(genCalls[1].briefing.includes("no usable approach summary"));
+    assert.equal(
+      loopOut.generations[0].summary.includes("no usable approach summary"),
+      true,
+    );
+    assert.equal(loopOut.generations[1].summary, "approach 1");
+    console.log("✔ eval/evolve-loop: climbs gaia `fitness`, accuracy naming, margin-0 tie handling, junk-summary guard");
 
     console.log("\nALL GAIA EVOLVE VALIDATION CHECKS PASSED");
   } finally {

mcp/src/lab/gaia/workflows/gaia-evolve-gen.yaml

@@ -279,3 +279,11 @@ params:
     final healthy publish, a summary of the APPROACH this generation took
     (the next generation reads it to know what has been tried), the list of
     changes, and any missing secrets.
+
+    Your structured answer is harvested from your FINAL message — the turn
+    where you stop calling tools. Fill every field with REAL values then:
+    never filler like "placeholder" or an empty summary. That summary is
+    the only record of this generation's approach the next generation and
+    the human report ever see; a generation whose approach is unreadable
+    gets retried by later generations, wasting their budget. If you run
+    out of steps, still describe truthfully whatever you actually shipped.

mcp/src/lab/harvey/workflows/harvey-evolve-gen.yaml

@@ -295,3 +295,11 @@ params:
     final healthy publish, a summary of the APPROACH this generation took
     (the next generation reads it to know what has been tried), the list of
     changes, and any missing secrets.
+
+    Your structured answer is harvested from your FINAL message — the turn
+    where you stop calling tools. Fill every field with REAL values then:
+    never filler like "placeholder" or an empty summary. That summary is
+    the only record of this generation's approach the next generation and
+    the human report ever see; a generation whose approach is unreadable
+    gets retried by later generations, wasting their budget. If you run
+    out of steps, still describe truthfully whatever you actually shipped.