instar 1.3.980 → 1.3.981

package/dist/monitoring/dashboardInsightCollectors.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"dashboardInsightCollectors.d.ts","sourceRoot":"","sources":["../../src/monitoring/dashboardInsightCollectors.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AAEH,OAAO,KAAK,EAAE,oBAAoB,EAAE,MAAM,2BAA2B,CAAC;AACtE,OAAO,KAAK,EAAE,WAAW,EAAE,gBAAgB,EAAiC,MAAM,6BAA6B,CAAC;AAQhH;;;;GAIG;AACH,wBAAgB,yBAAyB,CAAC,MAAM,EAAE,oBAAoB,GAAG,MAAM,gBAAgB,CA0C9F;AAED;;;;;GAKG;AACH,wBAAgB,wBAAwB,CAAC,IAAI,EAAE;IAC7C,oBAAoB,CAAC,EAAE,oBAAoB,GAAG,IAAI,CAAC;CACpD,GAAG,WAAW,EAAE,CAWhB"}
+{"version":3,"file":"dashboardInsightCollectors.d.ts","sourceRoot":"","sources":["../../src/monitoring/dashboardInsightCollectors.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AAEH,OAAO,KAAK,EAAE,oBAAoB,EAAE,MAAM,2BAA2B,CAAC;AACtE,OAAO,KAAK,EAAE,WAAW,EAAE,gBAAgB,EAAiC,MAAM,6BAA6B,CAAC;AAQhH;;;;GAIG;AACH,wBAAgB,yBAAyB,CAAC,MAAM,EAAE,oBAAoB,GAAG,MAAM,gBAAgB,CAwE9F;AAED;;;;;GAKG;AACH,wBAAgB,wBAAwB,CAAC,IAAI,EAAE;IAC7C,oBAAoB,CAAC,EAAE,oBAAoB,GAAG,IAAI,CAAC;CACpD,GAAG,WAAW,EAAE,CAWhB"}

package/dist/monitoring/dashboardInsightCollectors.js

@@ -48,19 +48,47 @@ export function buildLlmActivityCollector(ledger) {
             }
         }
         anomalies.sort((a, b) => (b.severity === 'alert' ? 1 : 0) - (a.severity === 'alert' ? 1 : 0));
-        const overallErrorRate = totalReal > 0 ? totalErrors / totalReal : 0;
+        // MIN_REAL_CALLS is declared at the top of this file as "minimum real calls
+        // before an error-rate is statistically meaningful" and the per-feature loop
+        // above honours it. The AGGREGATE did not: with zero calls it rendered
+        // `Error rate: 0%` and asserted "Routing is healthy — 0 checks ran", so an
+        // empty 24 hours was indistinguishable from a clean one. Same file, same
+        // constant, same reasoning — just applied to the aggregate too.
+        const overallErrorRate = totalReal > 0 ? totalErrors / totalReal : null;
         const metrics = [
             { label: 'LLM calls (24h)', value: String(totalReal) },
             { label: 'Errors (24h)', value: String(totalErrors) },
-            { label: 'Error rate', value: `${Math.round(overallErrorRate * 100)}%` },
+            {
+                label: 'Error rate',
+                // A rate over zero calls is not a rate. Show why it is absent, not a "0%"
+                // a reader cannot tell from a measured zero.
+                value: overallErrorRate === null ? 'no calls yet' : `${Math.round(overallErrorRate * 100)}%`,
+            },
             { label: 'Checks active', value: String(features.length) },
         ];
         const facts = [];
-        if (anomalies.length === 0) {
-            facts.push(`Routing is healthy — ${features.length} checks ran with no check failing a meaningful share of its calls.`);
+        // Order matters, and BAD NEWS COMES FIRST unconditionally. Today an anomaly
+        // cannot occur below the floor (a feature needs >= MIN_REAL_CALLS to flag, so
+        // totalReal < MIN_REAL_CALLS implies none), but that is an arithmetic
+        // coincidence rather than a guarantee — if the constants ever diverge, a
+        // low-volume window must still surface a real failure. The floor gates only
+        // the reassuring conclusion, never a warning.
+        if (anomalies.length > 0) {
+            facts.push(`${anomalies.length} check${anomalies.length === 1 ? '' : 's'} ${anomalies.length === 1 ? 'is' : 'are'} failing more than usual and worth a look.`);
+        }
+        else if (totalReal === 0) {
+            // Nothing ran. Say so, rather than claiming health from an empty window.
+            facts.push(features.length === 0
+                ? 'No checks have reported in the last 24 hours, so there is nothing to judge yet.'
+                : `${features.length} check${features.length === 1 ? '' : 's'} are configured but none made a call in the last 24 hours, so there is nothing to judge yet.`);
+        }
+        else if (totalReal < MIN_REAL_CALLS) {
+            // The rate above is real data, but the volume is too low to call it healthy —
+            // the same asymmetry the per-feature loop already applies.
+            facts.push(`Only ${totalReal} call${totalReal === 1 ? '' : 's'} in the last 24 hours — too few to say whether routing is healthy.`);
         }
         else {
-            facts.push(`${anomalies.length} check${anomalies.length === 1 ? '' : 's'} ${anomalies.length === 1 ? 'is' : 'are'} failing more than usual and worth a look.`);
+            facts.push(`Routing is healthy — ${features.length} checks ran with no check failing a meaningful share of its calls.`);
         }
         if (busiest && busiest.realCalls > 0)
             facts.push(`The busiest check is ${busiest.feature} (${busiest.realCalls} calls in 24h).`);

package/dist/monitoring/dashboardInsightCollectors.js.map

@@ -1 +1 @@
-{"version":3,"file":"dashboardInsightCollectors.js","sourceRoot":"","sources":["../../src/monitoring/dashboardInsightCollectors.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AAKH,iFAAiF;AACjF,MAAM,gBAAgB,GAAG,IAAI,CAAC;AAC9B,MAAM,gBAAgB,GAAG,GAAG,CAAC;AAC7B,2EAA2E;AAC3E,MAAM,cAAc,GAAG,CAAC,CAAC;AAEzB;;;;GAIG;AACH,MAAM,UAAU,yBAAyB,CAAC,MAA4B;IACpE,OAAO,GAAG,EAAE;QACV,MAAM,OAAO,GAAG,MAAM,CAAC,OAAO,CAAC,EAAE,UAAU,EAAE,EAAE,EAAE,CAAC,CAAC;QACnD,MAAM,QAAQ,GAAG,OAAO,CAAC,QAAQ,IAAI,EAAE,CAAC;QACxC,IAAI,SAAS,GAAG,CAAC,CAAC;QAClB,IAAI,WAAW,GAAG,CAAC,CAAC;QACpB,IAAI,OAAO,GAAkD,IAAI,CAAC;QAClE,MAAM,SAAS,GAAqB,EAAE,CAAC;QAEvC,KAAK,MAAM,CAAC,IAAI,QAAQ,EAAE,CAAC;YACzB,SAAS,IAAI,CAAC,CAAC,SAAS,CAAC;YACzB,WAAW,IAAI,CAAC,CAAC,MAAM,CAAC;YACxB,IAAI,CAAC,OAAO,IAAI,CAAC,CAAC,SAAS,GAAG,OAAO,CAAC,SAAS;gBAAE,OAAO,GAAG,EAAE,OAAO,EAAE,CAAC,CAAC,OAAO,EAAE,SAAS,EAAE,CAAC,CAAC,SAAS,EAAE,CAAC;YAC1G,IAAI,CAAC,CAAC,SAAS,IAAI,cAAc,IAAI,CAAC,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;gBAClD,MAAM,IAAI,GAAG,CAAC,CAAC,MAAM,GAAG,CAAC,CAAC,SAAS,CAAC;gBACpC,IAAI,IAAI,IAAI,gBAAgB,EAAE,CAAC;oBAC7B,SAAS,CAAC,IAAI,CAAC;wBACb,IAAI,EAAE,GAAG,CAAC,CAAC,OAAO,eAAe,IAAI,CAAC,KAAK,CAAC,IAAI,GAAG,GAAG,CAAC,kBAAkB,CAAC,CAAC,MAAM,OAAO,CAAC,CAAC,SAAS,UAAU;wBAC7G,QAAQ,EAAE,IAAI,IAAI,gBAAgB,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,OAAO;qBACvD,CAAC,CAAC;gBACL,CAAC;YACH,CAAC;QACH,CAAC;QACD,SAAS,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,QAAQ,KAAK,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,QAAQ,KAAK,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;QAE9F,MAAM,gBAAgB,GAAG,SAAS,GAAG,CAAC,CAAC,CAAC,CAAC,WAAW,GAAG,SAAS,CAAC,CAAC,CAAC,CAAC,CAAC;QACrE,MAAM,OAAO,GAAoB;YAC/B,EAAE,KAAK,EAAE,iBAAiB,EAAE,KAAK,EAAE,MAAM,CAAC,SAAS,CAAC,EAAE;YACtD,EAAE,KAAK,EAAE,cAAc,EAAE,KAAK,EAAE,MAAM,CAAC,WAAW,CAAC,EAAE;YACrD,EAAE,KAAK,EAAE,YAAY,EAAE,KAAK,EAAE,GAAG,IAAI,CAAC,KAAK,CAAC,gBAAgB,GAAG,GAAG,CAAC,GAAG,EAAE;YACxE,EAAE,KAAK,EAAE,eAAe,EAAE,KAAK,EAAE,MAAM,CAAC,QAAQ,CAAC,MAAM,CAAC,EAAE;SAC3D,CAAC;QACF,MAAM,KAAK,GAAa,EAAE,CAAC;QAC3B,IAAI,SAAS,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;YAC3B,KAAK,CAAC,IAAI,CAAC,wBAAwB,QAAQ,CAAC,MAAM,oEAAoE,CAAC,CAAC;QAC1H,CAAC;aAAM,CAAC;YACN,KAAK,CAAC,IAAI,CAAC,GAAG,SAAS,CAAC,MAAM,SAAS,SAAS,CAAC,MAAM,KAAK,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,GAAG,IAAI,SAAS,CAAC,MAAM,KAAK,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,KAAK,4CAA4C,CAAC,CAAC;QACjK,CAAC;QACD,IAAI,OAAO,IAAI,OAAO,CAAC,SAAS,GAAG,CAAC;YAAE,KAAK,CAAC,IAAI,CAAC,wBAAwB,OAAO,CAAC,OAAO,KAAK,OAAO,CAAC,SAAS,iBAAiB,CAAC,CAAC;QAEjI,OAAO,EAAE,KAAK,EAAE,OAAO,EAAE,SAAS,EAAE,SAAS,EAAE,IAAI,CAAC,GAAG,EAAE,EAAE,CAAC;IAC9D,CAAC,CAAC;AACJ,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,wBAAwB,CAAC,IAExC;IACC,MAAM,KAAK,GAAkB,EAAE,CAAC;IAChC,IAAI,IAAI,CAAC,oBAAoB,EAAE,CAAC;QAC9B,KAAK,CAAC,IAAI,CAAC;YACT,EAAE,EAAE,cAAc;YAClB,KAAK,EAAE,cAAc;YACrB,GAAG,EAAE,cAAc;YACnB,OAAO,EAAE,yBAAyB,CAAC,IAAI,CAAC,oBAAoB,CAAC;SAC9D,CAAC,CAAC;IACL,CAAC;IACD,OAAO,KAAK,CAAC;AACf,CAAC"}
+{"version":3,"file":"dashboardInsightCollectors.js","sourceRoot":"","sources":["../../src/monitoring/dashboardInsightCollectors.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AAKH,iFAAiF;AACjF,MAAM,gBAAgB,GAAG,IAAI,CAAC;AAC9B,MAAM,gBAAgB,GAAG,GAAG,CAAC;AAC7B,2EAA2E;AAC3E,MAAM,cAAc,GAAG,CAAC,CAAC;AAEzB;;;;GAIG;AACH,MAAM,UAAU,yBAAyB,CAAC,MAA4B;IACpE,OAAO,GAAG,EAAE;QACV,MAAM,OAAO,GAAG,MAAM,CAAC,OAAO,CAAC,EAAE,UAAU,EAAE,EAAE,EAAE,CAAC,CAAC;QACnD,MAAM,QAAQ,GAAG,OAAO,CAAC,QAAQ,IAAI,EAAE,CAAC;QACxC,IAAI,SAAS,GAAG,CAAC,CAAC;QAClB,IAAI,WAAW,GAAG,CAAC,CAAC;QACpB,IAAI,OAAO,GAAkD,IAAI,CAAC;QAClE,MAAM,SAAS,GAAqB,EAAE,CAAC;QAEvC,KAAK,MAAM,CAAC,IAAI,QAAQ,EAAE,CAAC;YACzB,SAAS,IAAI,CAAC,CAAC,SAAS,CAAC;YACzB,WAAW,IAAI,CAAC,CAAC,MAAM,CAAC;YACxB,IAAI,CAAC,OAAO,IAAI,CAAC,CAAC,SAAS,GAAG,OAAO,CAAC,SAAS;gBAAE,OAAO,GAAG,EAAE,OAAO,EAAE,CAAC,CAAC,OAAO,EAAE,SAAS,EAAE,CAAC,CAAC,SAAS,EAAE,CAAC;YAC1G,IAAI,CAAC,CAAC,SAAS,IAAI,cAAc,IAAI,CAAC,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;gBAClD,MAAM,IAAI,GAAG,CAAC,CAAC,MAAM,GAAG,CAAC,CAAC,SAAS,CAAC;gBACpC,IAAI,IAAI,IAAI,gBAAgB,EAAE,CAAC;oBAC7B,SAAS,CAAC,IAAI,CAAC;wBACb,IAAI,EAAE,GAAG,CAAC,CAAC,OAAO,eAAe,IAAI,CAAC,KAAK,CAAC,IAAI,GAAG,GAAG,CAAC,kBAAkB,CAAC,CAAC,MAAM,OAAO,CAAC,CAAC,SAAS,UAAU;wBAC7G,QAAQ,EAAE,IAAI,IAAI,gBAAgB,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,OAAO;qBACvD,CAAC,CAAC;gBACL,CAAC;YACH,CAAC;QACH,CAAC;QACD,SAAS,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,QAAQ,KAAK,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,QAAQ,KAAK,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;QAE9F,4EAA4E;QAC5E,6EAA6E;QAC7E,uEAAuE;QACvE,2EAA2E;QAC3E,yEAAyE;QACzE,gEAAgE;QAChE,MAAM,gBAAgB,GAAG,SAAS,GAAG,CAAC,CAAC,CAAC,CAAC,WAAW,GAAG,SAAS,CAAC,CAAC,CAAC,IAAI,CAAC;QACxE,MAAM,OAAO,GAAoB;YAC/B,EAAE,KAAK,EAAE,iBAAiB,EAAE,KAAK,EAAE,MAAM,CAAC,SAAS,CAAC,EAAE;YACtD,EAAE,KAAK,EAAE,cAAc,EAAE,KAAK,EAAE,MAAM,CAAC,WAAW,CAAC,EAAE;YACrD;gBACE,KAAK,EAAE,YAAY;gBACnB,0EAA0E;gBAC1E,6CAA6C;gBAC7C,KAAK,EAAE,gBAAgB,KAAK,IAAI,CAAC,CAAC,CAAC,cAAc,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,gBAAgB,GAAG,GAAG,CAAC,GAAG;aAC7F;YACD,EAAE,KAAK,EAAE,eAAe,EAAE,KAAK,EAAE,MAAM,CAAC,QAAQ,CAAC,MAAM,CAAC,EAAE;SAC3D,CAAC;QACF,MAAM,KAAK,GAAa,EAAE,CAAC;QAC3B,4EAA4E;QAC5E,8EAA8E;QAC9E,sEAAsE;QACtE,yEAAyE;QACzE,4EAA4E;QAC5E,8CAA8C;QAC9C,IAAI,SAAS,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;YACzB,KAAK,CAAC,IAAI,CAAC,GAAG,SAAS,CAAC,MAAM,SAAS,SAAS,CAAC,MAAM,KAAK,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,GAAG,IAAI,SAAS,CAAC,MAAM,KAAK,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,KAAK,4CAA4C,CAAC,CAAC;QACjK,CAAC;aAAM,IAAI,SAAS,KAAK,CAAC,EAAE,CAAC;YAC3B,yEAAyE;YACzE,KAAK,CAAC,IAAI,CACR,QAAQ,CAAC,MAAM,KAAK,CAAC;gBACnB,CAAC,CAAC,iFAAiF;gBACnF,CAAC,CAAC,GAAG,QAAQ,CAAC,MAAM,SAAS,QAAQ,CAAC,MAAM,KAAK,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,GAAG,8FAA8F,CAC9J,CAAC;QACJ,CAAC;aAAM,IAAI,SAAS,GAAG,cAAc,EAAE,CAAC;YACtC,8EAA8E;YAC9E,2DAA2D;YAC3D,KAAK,CAAC,IAAI,CACR,QAAQ,SAAS,QAAQ,SAAS,KAAK,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,GAAG,oEAAoE,CACxH,CAAC;QACJ,CAAC;aAAM,CAAC;YACN,KAAK,CAAC,IAAI,CAAC,wBAAwB,QAAQ,CAAC,MAAM,oEAAoE,CAAC,CAAC;QAC1H,CAAC;QACD,IAAI,OAAO,IAAI,OAAO,CAAC,SAAS,GAAG,CAAC;YAAE,KAAK,CAAC,IAAI,CAAC,wBAAwB,OAAO,CAAC,OAAO,KAAK,OAAO,CAAC,SAAS,iBAAiB,CAAC,CAAC;QAEjI,OAAO,EAAE,KAAK,EAAE,OAAO,EAAE,SAAS,EAAE,SAAS,EAAE,IAAI,CAAC,GAAG,EAAE,EAAE,CAAC;IAC9D,CAAC,CAAC;AACJ,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,wBAAwB,CAAC,IAExC;IACC,MAAM,KAAK,GAAkB,EAAE,CAAC;IAChC,IAAI,IAAI,CAAC,oBAAoB,EAAE,CAAC;QAC9B,KAAK,CAAC,IAAI,CAAC;YACT,EAAE,EAAE,cAAc;YAClB,KAAK,EAAE,cAAc;YACrB,GAAG,EAAE,cAAc;YACnB,OAAO,EAAE,yBAAyB,CAAC,IAAI,CAAC,oBAAoB,CAAC;SAC9D,CAAC,CAAC;IACL,CAAC;IACD,OAAO,KAAK,CAAC;AACf,CAAC"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "instar",
-  "version": "1.3.980",
+  "version": "1.3.981",
   "description": "Coherence infrastructure for self-evolving AI agents — on the Claude Code or Codex subscription you already have.",
   "type": "module",
   "main": "dist/index.js",

package/src/data/builtin-manifest.json

@@ -1,8 +1,8 @@
 {
   "$schema": "./builtin-manifest.schema.json",
   "schemaVersion": 1,
-  "generatedAt": "2026-07-26T04:26:53.883Z",
-  "instarVersion": "1.3.980",
+  "generatedAt": "2026-07-26T04:54:53.613Z",
+  "instarVersion": "1.3.981",
   "entryCount": 202,
   "entries": {
     "hook:session-start": {

package/upgrades/1.3.981.md

@@ -0,0 +1,81 @@
+# Upgrade Guide — vNEXT
+
+<!-- assembled-by: assemble-next-md -->
+<!-- bump: patch -->
+
+## What Changed
+
+**The LLM Activity dashboard panel claimed "Routing is healthy" when nothing had run — and in one
+case displayed `Error rate: 50%` directly above that claim.**
+
+`dashboardInsightCollectors.buildLlmActivityCollector` computed
+`overallErrorRate = totalReal > 0 ? totalErrors / totalReal : 0`, rendered it as `Error rate: 0%`, and
+pushed a health fact whenever the anomaly list was empty:
+
+    "Routing is healthy — 0 checks ran with no check failing a meaningful share of its calls."
+
+With an empty 24-hour window that sentence claimed health from no observations, its own denominator
+("0 checks ran") contradicting its claim. The sharper failure: the per-feature anomaly loop correctly
+requires `f.realCalls >= MIN_REAL_CALLS` before flagging a check, but the health sentence had no such
+condition — so a window with 2 calls where 1 errored rendered `Error rate: 50%` beside "Routing is
+healthy". Both halves were individually defensible; the verdict was simply never told about the
+threshold the alarm was already using.
+
+`MIN_REAL_CALLS = 5` is declared at the top of that same file, commented "minimum real calls before an
+error-rate is statistically meaningful". This applies it to the aggregate:
+
+- `overallErrorRate` is `null` at zero calls; the metric renders **`no calls yet`** rather than a `0%`
+  indistinguishable from a measured zero. The call count and check count still travel beside it.
+- Zero calls → a fact saying there is nothing to judge yet (distinguishing "no checks reported at all"
+  from "checks configured but idle").
+- Below the floor → the real rate is still shown, but the verdict is withheld: "Only N calls in the
+  last 24 hours — too few to say whether routing is healthy."
+- **Anomalies are evaluated FIRST and unconditionally**, so no evidence caveat can ever displace a
+  warning. Today an anomaly cannot occur below the floor, but that is an arithmetic coincidence
+  between two constants rather than a guarantee, and the ordering makes it structural.
+
+Verified before/after on the built collector: nothing-ran went from `0%` + "Routing is healthy" to
+`no calls yet` + "nothing to judge yet"; 2 calls with 1 error went from `50%` + "Routing is healthy" to
+`50%` + "too few to say"; 10 clean calls still reads "Routing is healthy" unchanged; 6 calls with 3
+errors still surfaces the failing check.
+
+A grammar fix falls out of the plural handling: "1 checks are failing" → "1 check is failing".
+
+## What to Tell Your User
+
+Nothing to do. On most installs this panel ships dark, so there is nothing visible.
+
+Where it runs, the panel is now honest about a quiet window: it says there is nothing to judge yet
+rather than reporting a flawless zero. Read that as "nothing is wrong and nothing is known" — it is
+deliberately not an alarm. Genuine failures are still reported exactly as promptly as before, and are
+now always reported ahead of any note about low call volume.
+
+## Summary of New Capabilities
+
+- The LLM Activity panel distinguishes "no calls observed" from "no errors observed": the error rate
+  reads `no calls yet` instead of `0%`, with the call count beside it.
+- A health verdict now requires enough calls to support it, reusing the file's own existing
+  `MIN_REAL_CALLS` threshold rather than a new one.
+- A real failure is surfaced ahead of any low-evidence caveat, by branch order rather than by luck.
+
+## Evidence
+
+**Reproduction (before):** build the collector over an empty ledger and read the snapshot. Observed:
+`Error rate: 0%` and the fact `"Routing is healthy — 0 checks ran with no check failing a meaningful
+share of its calls."` Then seed 2 calls with 1 error: observed `Error rate: 50%` together with
+`"Routing is healthy — 1 checks ran…"`.
+
+**Observed after,** same four scenarios against the built collector:
+
+| scenario | before | after |
+|---|---|---|
+| nothing ran (0 calls, 0 checks) | `0%` + "Routing is healthy — 0 checks ran…" | `no calls yet` + "No checks have reported in the last 24 hours, so there is nothing to judge yet." |
+| 2 calls, 1 error (below floor) | `50%` + "Routing is healthy — 1 checks ran…" | `50%` + "Only 2 calls in the last 24 hours — too few to say whether routing is healthy." |
+| 10 clean calls | `0%` + "Routing is healthy" | `0%` + "Routing is healthy" — unchanged |
+| 6 calls, 3 errors | `50%` + "1 checks are failing…" | `50%` + "1 check is failing more than usual and worth a look." |
+
+**Tests:** `tests/unit/dashboard-insight-collectors-unobserved.test.ts` — four cases covering
+no-rate-and-no-health-claim at zero calls, real-rate-with-verdict-withheld below the floor,
+health-claimed-once-supported, and a real failure surfacing ahead of any caveat. The existing
+`tests/e2e/insights-alive.test.ts` seeds 6 calls above the floor and passes unchanged. 7 tests pass
+across both files; `npx tsc --noEmit` clean.

package/upgrades/side-effects/dashboard-error-rate-unobserved.md

@@ -0,0 +1,141 @@
+# Side-Effects Review — the dashboard showed "Error rate: 50%" beside "Routing is healthy"
+
+**Version / slug:** `dashboard-error-rate-unobserved`
+**Date:** `2026-07-26`
+**Author:** `Echo (instar-dev agent)`
+**Second-pass reviewer:** `author-applied lenses — see Phase 5 (reduced independence, disclosed)`
+
+## Summary of the change
+
+The LLM Activity insight collector computed `overallErrorRate = totalReal > 0 ? totalErrors/totalReal : 0`
+and rendered it as `Error rate: 0%`, then asserted a health fact whenever no anomaly had been flagged:
+
+    "Routing is healthy — 0 checks ran with no check failing a meaningful share of its calls."
+
+Two failures fell out of that. With an empty 24-hour window the panel claimed health from no data —
+and the sentence's own denominator ("0 checks ran") contradicted its claim. Worse, with a genuine
+failure below the anomaly floor the panel showed **`Error rate: 50%` beside "Routing is healthy"**,
+because the per-feature loop correctly required `f.realCalls >= MIN_REAL_CALLS` before flagging while
+the health sentence had no such condition.
+
+`MIN_REAL_CALLS = 5` is declared at the top of this same file, commented "minimum real calls before
+an error-rate is statistically meaningful". The fix applies that existing, already-honoured constant
+to the aggregate: the rate is `null` (rendered "no calls yet") at zero calls, a sub-floor window says
+so plainly, and **a real failure is surfaced ahead of any evidence caveat**.
+
+## Decision-point inventory
+
+| point | classification | note |
+|---|---|---|
+| the `facts` branch order | `invariant` | Deterministic. Anomalies are checked FIRST and unconditionally, so no caveat can displace a warning. |
+| the sub-floor caveat | `invariant` | A comparison against the file's existing `MIN_REAL_CALLS`. Gates only the reassuring conclusion. |
+| `Error rate` rendering | `invariant` | `null` at zero calls → a string saying why, never a computed default. |
+
+## 1. Over-block
+
+**What legitimate inputs does this change reject that it shouldn't?**
+
+Nothing is blocked — this is a read-only dashboard collector with no gating authority. The analogue
+is **withholding a health claim from a genuinely healthy window**: a window with 1–4 clean calls now
+reads "too few to say" instead of "healthy". Bounded, and the real rate is still displayed beside the
+call count so the reader can judge for themselves. At 5+ calls the claim returns unchanged
+(demonstrated: 10 clean calls → "Routing is healthy", byte-identical to before).
+
+## 2. Under-block
+
+**What failure modes does this still miss?**
+
+- The floor is a call count, not a confidence interval. 5 clean calls yields "healthy" though a real
+  interval would still be wide. Deliberate: this reuses the file's existing constant rather than
+  inventing a second, differently-tuned threshold in the same file.
+- `busiest` is still reported whenever `realCalls > 0`, which is honest (it is a count, not a rate).
+- The 24h window itself is not surfaced as a caveat when the process started minutes ago — a
+  short-uptime window looks the same as a genuinely quiet one. Out of scope here; noted rather than
+  silently skipped.
+
+## 3. Level-of-abstraction fit
+
+**Is this at the right layer?**
+
+Yes. The defect is in how this collector renders its own aggregate, so the fix belongs in it. No
+shared helper was extracted: the sibling instance in `CoherenceGate` (PR #1648) has different types
+and a different output shape, and abstracting over two instances would be the premature-generalisation
+mistake this project's own spec review caught elsewhere.
+
+## 4. Signal vs authority compliance
+
+**Required reference:** [docs/signal-vs-authority.md](../../docs/signal-vs-authority.md)
+
+**Does this change hold blocking authority with brittle logic?**
+
+No. Pure presentation, zero authority: no branch blocks, delays or alters anything, and the collector
+is consumed only by the insights read surface. There is no misclassification path that can deny
+something, because nothing is denied. The change strictly reduces the panel's capacity to mislead.
+
+## 4b. Judgment-point check (Judgment Within Floors standard)
+
+No judgment point is introduced. Every branch is a deterministic comparison on counted calls against
+a constant already present in the file. The one asymmetry is deliberate and stated in a code comment:
+anomalies are evaluated first and unconditionally, so thin evidence can never suppress a warning —
+only the reassuring conclusion requires evidence behind it.
+
+## 5. Interactions
+
+- `overallErrorRate` is a local; verified by grep that nothing outside this file reads it.
+- `InsightMetric.value` is typed `string`, so "no calls yet" needs no type change and no consumer
+  performs arithmetic on it.
+- The existing E2E (`tests/e2e/insights-alive.test.ts`) seeds 6 calls with 3 errors — above the floor,
+  so it still routes to the anomaly branch and still asserts `/50%/`. Re-run: passes unchanged.
+- A visible grammar fix falls out of the plural handling: "1 checks are failing" → "1 check is failing".
+- No persistence, no config, no migration: every value is computed per call from the ledger.
+
+## 6. External surfaces
+
+The `/insights` payload for the `llm-activity` page changes in two places: `Error rate` may read
+"no calls yet" instead of "0%", and the first fact may be a caveat rather than a health claim. Both
+are display strings already typed as free text. The feature ships dark on the fleet (dev-agent gated,
+dryRun), so exposure is limited to development agents today.
+
+## 6b. Operator-surface quality
+
+This is the whole point of the change. An operator glancing at the panel during a quiet window
+previously read "Error rate: 0% · Routing is healthy" and would reasonably conclude the routing layer
+was fine. They now read "no calls yet" and "nothing to judge yet". And in the sub-floor case they no
+longer see a 50% error rate sitting next to the word "healthy" — the panel stops contradicting
+itself. None of the new wording is alarming: "nothing to judge yet" says nothing is wrong AND nothing
+is known.
+
+## 7. Multi-machine posture (Cross-Machine Coherence)
+
+**Posture: `unified` by construction — no new state.** The collector derives everything, per call,
+from the per-process metrics ledger that already existed. Each machine reports its own LLM activity,
+which is correct: a call was made by a specific process. No replication path is needed because nothing
+is stored, and no `machine-local-justification` marker is required because no new machine-local state
+is introduced.
+
+## 8. Rollback cost
+
+Trivial: revert one commit touching one source file plus its new test. No migration, no persisted
+state, no config key. A consumer written against the new strings sees the old ones again, and the old
+strings' failure mode is the one documented above.
+
+## Phase 5 — Second-pass review (independent reviewer subagent)
+
+**Disclosure, per Truthful Provenance:** no independent reviewer subagent was spawned — a standing
+instruction in this session prohibits spawning subagents unless the operator requests it. The review
+lenses were applied by the author. That is **reduced independence**, recorded as such rather than
+presented as a concurring second pass.
+
+What author-applied review actually caught and changed:
+
+1. **The first draft ordered the branches wrongly**, putting the sub-floor caveat ahead of the
+   anomaly message. Today no anomaly can occur below the floor (a feature needs `>= MIN_REAL_CALLS`
+   to flag, so `totalReal < MIN_REAL_CALLS` implies none) — but that is an arithmetic coincidence,
+   not a guarantee. Reordered so anomalies win unconditionally, and the reasoning is in a comment so
+   a future edit to either constant cannot silently reintroduce suppression.
+2. **My own first test asserted the wrong thing** and failed on correct output: it matched
+   `/routing is healthy/i`, which the caveat legitimately contains ("too few to say whether routing
+   is healthy"). Corrected to assert the absence of the AFFIRMATIVE claim. Worth recording because
+   this project's Tier 1 is about instruments that mis-measure — including my tests.
+3. The consumer check was run rather than assumed, establishing that `overallErrorRate` is local and
+   that `InsightMetric.value` is already a string.