Version updates

README.md

@@ -81,7 +81,7 @@ cd <your-space>
 git lfs install
 # copy the contents of this DecisionLab folder into it (not the folder itself), then:
 git add .
-git commit -m "DecisionLab 2.3.0"
+git commit -m "DecisionLab 2.3.1"
 git push
 ```
 
@@ -122,7 +122,7 @@ The plain HTTP API works on the Space too, at `https://<your-user>-<your-space>.
 ## Using the lab
 
 1. **Setup** shows each model's status, size, device, load time and how it defines confidence.
-2. **Decision**: pick one of the 47 demos (tabs group them into triage, routing and planning, loop control, guardrails, and agent security (nano set)) or write your own state and questions. Nothing is executed; every demo is a decision only. Questions use the Laya/Jev JSON shape:
+2. **Decision**: pick one of the 47 demos (tabs group them into triage, routing and planning, loop control, guardrails, and agent security) or write your own state and questions. Nothing is executed; every demo is a decision only. Questions use the Laya/Jev JSON shape:
    ```json
    {"team":    {"type": "choice", "instructions": "Which team?", "criteria": {"bug": "Something is broken", "sales": "Pricing"}},
     "urgency": {"type": "score",  "instructions": "How urgent?", "criteria": ["Can wait", "Today", "Right now"]},
@@ -135,7 +135,7 @@ The plain HTTP API works on the Space too, at `https://<your-user>-<your-space>.
 
 ## Demos
 
-47 demos in 5 groups. The tab numbers match the table. Demos 25–47 are the **Agent security (nano set)** group: the states and questions of the Athr_Agent_Sec nano demo set (`app/demos_nano.json`), with no reference answers, so they show each model's answers, confidence and agreement but are not scored for matches.
+47 demos in 5 groups. The tab numbers match the table. Demos 25–47 are the **Agent security** group: the states and questions of the Athr_Agent_Sec nano demo set (`app/demos_nano.json`), with no reference answers, so they show each model's answers, confidence and agreement but are not scored. A scoreboard run of only these demos shows no Agentic Use Score (there is nothing to score it against), just what each model would act on, its speed and the agreement.
 
 | # | Group | Demo | What it tests |
 |---|---|---|---|
@@ -242,7 +242,7 @@ DecisionLab/
 │   ├── security.py    # body limit and security headers (plain ASGI)
 │   ├── validation.py  # checks /api/decide question sets, sizes and model keys (no web framework)
 │   ├── demos.py       # groups, the 47 demos (24 scored + 23 nano), reference answers and high-stakes flags
-│   ├── demos_nano.json  # Agent security (nano set): states and questions only
+│   ├── demos_nano.json  # Agent security: states and questions only
 │   └── static/        # index.html, app.css, app.js, logo.png
 └── tests/             # unittest suite, copied into the image
 ```

app/demos.py

@@ -31,7 +31,7 @@ GROUPS = [
      "blurb": "Inside the agent loop: stop, retry, change course, or verify the work before reporting it."},
     {"id": "guardrails", "family": "Agent decisions", "label": "Guardrails",
      "blurb": "Irreversible actions, prompt injection, fraud, data leaving the company, secrets and permissions."},
-    {"id": "agent_security", "family": "Agent decisions", "label": "Agent security (nano set)",
+    {"id": "agent_security", "family": "Agent decisions", "label": "Agent security",
      "blurb": "Drift, report auditing, action risk, speech consistency and prompt-injection screening, from the "
               "Athr_Agent_Sec nano demo set. No reference answers: compare the models' answers, confidence and agreement."},
 ]
@@ -410,7 +410,7 @@ DEMOS = [
     },
 ]
 
-# Agent security (nano set): only the states and questions of decisionlab_nano_demos.json (operator ruling
+# Agent security: only the states and questions of decisionlab_nano_demos.json (operator ruling
 # 2026-09-30). No reference answers or stakes, so these demos are not scored for matches; every model still answers
 # them, with confidence, agreement and speed.
 _NANO = json.loads((Path(__file__).with_name("demos_nano.json")).read_text(encoding="utf-8"))["demos"]

app/main.py

@@ -27,7 +27,7 @@ MAX_BODY_BYTES = int(os.getenv("MAX_BODY_BYTES", str(256 * 1024)))
 DECIDE_SLOTS = Gate(int(os.getenv("MAX_PENDING_DECIDES", "4")))      # DL-SA-004
 
 # DL-SA-007: no /docs, /redoc or /openapi.json
-VERSION = "2.3.0"
+VERSION = "2.3.1"
 PAGE = stamp_assets((STATIC / "index.html").read_text(encoding="utf-8"), VERSION)
 
 

app/static/app.css

@@ -416,3 +416,8 @@ input[type="range"] { height: 28px; }
   .repo a { padding: 12px 0; }
   .search input { font-size: 16px; }
 }
+
+/* ---------- runs with no reference answers: no score, a plain note instead */
+.unscored-note { grid-column: 1 / -1; border: 1px dashed var(--line); border-radius: var(--radius-m); padding: 18px 22px; }
+.unscored-note h3 { font: 500 22px var(--display); margin: 0 0 6px; }
+.unscored-note p { margin: 0; color: var(--graphite); max-width: 70ch; }

app/static/app.js

@@ -467,7 +467,7 @@ function groupBreakdown(rows, thr) {
     const w = best(Object.fromEntries(MODELS.map((m) => [m.key, L[m.key].rate])));
     return `<tr><td>${esc(groupOf(id)?.label ?? id)}<small>${esc(groupOf(id)?.family ?? "")}</small></td>${MODELS.map((m) => {
       const x = L[m.key];
-      if (!x.n) return `<td><b>–</b><small>no reference answers</small></td>`;   // e.g. the nano set: states and questions only
+      if (!x.n) return `<td><b>–</b><small>no reference answers</small></td>`;   // e.g. Agent security: states and questions only
       return `<td class="${w.has(m.key) ? "win" : ""}"><b>${x.hits} / ${x.n}</b><small>${x.cw} confident mistake${x.cw === 1 ? "" : "s"}</small></td>`;
     }).join("")}</tr>`;
   }).join("");
@@ -484,8 +484,41 @@ function renderScoreHead() {
   $("score-head").innerHTML = `<th scope="col">Demo</th><th scope="col">Question</th><th scope="col">Reference</th>${modelHeads()}<th scope="col">Agree</th>`;
 }
 
+// A run with no reference answers (e.g. only the Agent security demos) cannot be scored: the Agentic Use Score
+// would fall back to the same defaults for every model. Show what CAN be compared instead.
+function renderUnscoredSummary() {
+  const { rows, times } = board;
+  const thr = threshold();
+  const card = (m) => {
+    const recs = rows.map((r) => r.ans[m.key]).filter(Boolean);
+    const acts = recs.filter((rec) => conf(rec) >= thr).length;
+    const med = median(times[m.key] || []);
+    const facts = [
+      ["Questions answered", `${recs.length} / ${rows.length}`],
+      ["Median model time", med != null ? `${Math.round(med)} ms` : "–"],
+      [`Acts on (conf ≥ ${thr.toFixed(2)})`, `${acts} / ${recs.length}`],
+      ["Defers", `${recs.length - acts} / ${recs.length}`],
+    ];
+    return `<div class="sum-card side-${esc(m.side)}">
+      <h3>${esc(m.name)}</h3>
+      <dl>${facts.map(([k, v]) => `<div><dt>${k}</dt><dd>${v}</dd></div>`).join("")}</dl>
+    </div>`;
+  };
+  const answered = rows.filter((r) => MODELS.every((m) => r.ans[m.key]));
+  const allAgree = answered.filter((r) => new Set(MODELS.map((m) => r.ans[m.key].choice)).size === 1).length;
+  const sum = $("summary");
+  sum.hidden = false;
+  sum.innerHTML = `<div class="unscored-note">
+      <h3>Not scored</h3>
+      <p>The Agentic Use Score compares each answer with a reference answer, and the questions in this run have none. So there is no score or ranking here; compare the models below by what they would act on at your threshold, their speed, and how often they agree.</p>
+    </div>`
+    + MODELS.map(card).join("")
+    + `<p class="sum-both">All ${MODELS.length} models give the same top answer on ${allAgree} of ${answered.length} questions. Confidence uses the measure selected above; changing it or the threshold updates every number here.</p>`;
+}
+
 function renderSummary() {
   const { rows, times } = board;
+  if (!rows.some((r) => r.ref !== undefined)) return renderUnscoredSummary();
   const thr = threshold();
   const S = Object.fromEntries(MODELS.map((m) => [m.key, modelStats(rows, m.key, times[m.key], thr)]));
   const ratio = (x, y) => (y ? x / y : null);

tests/test_demos.py

@@ -47,6 +47,8 @@ class DemoSetTest(unittest.TestCase):
         """Operator ruling 2026-09-30: only the states and questions of decisionlab_nano_demos.json are added."""
         nano = [d for d in DEMOS if d["group"] == "agent_security"]
         self.assertEqual(len(nano), 23)
+        label = next(g["label"] for g in GROUPS if g["id"] == "agent_security")
+        self.assertEqual(label, "Agent security")                      # operator, 2026-10-01: no "(nano set)"
         self.assertEqual(sum(len(d["questions"]) for d in nano), 34)
         for d in nano:
             with self.subTest(demo=d["id"]):