diff --git a/docs/METRICS.md b/docs/METRICS.md
index d3acee8de4..4c9c37e53a 100644
--- a/docs/METRICS.md
+++ b/docs/METRICS.md
@@ -11,8 +11,8 @@ and where those numbers live*.
 ## Why it exists
 
 PortOS's [Layered Intelligence loop](./plans/2026-07-07-layered-intelligence-loop.md)
-perpetually reviews each managed app and files one high-value improvement proposal
-per run. To reason well it needs to evaluate the app against **its own**
+perpetually reviews each managed app and files at most one high-value improvement
+proposal per run. To reason well it needs to evaluate the app against **its own**
 performance — user success, product KPIs, production telemetry — not against how
 reliably PortOS's coding agents happen to change it.
 
@@ -21,6 +21,30 @@ telemetry/analytics service, a metrics dashboard, an events pipeline. An agent h
 no way to discover those unless the repo tells it where they are and how to read
 them. `METRICS.md` is that map.
 
+## Layered Intelligence decision contract
+
+Layered Intelligence uses a metrics-first decision path:
+
+1. When a metric shows a meaningful gap, it proposes the single highest-value
+   app change that can plausibly move that metric.
+2. When the metrics are missing, unavailable, stale, or inconclusive, its
+   reasoning agent may inspect the repository read-only to understand the
+   relevant product flow and existing instrumentation. This is context discovery,
+   not permission to turn the run into a generic code-quality audit.
+3. When that inspection identifies a specific missing signal or context that
+   blocks evaluation against a named goal, it files an `app-data-gap` issue for
+   the minimum visibility needed. The issue must identify the unanswered product
+   question, the expected data source, and acceptance criteria; “add more
+   telemetry” is not decision-complete.
+4. It files nothing only when the available evidence is sufficient and healthy,
+   the needed visibility already exists, or no decision-complete proposal fits
+   the configured scopes (subject to the loop's filing safety and delivery
+   gates). Inconclusive metrics alone do not prove the app is healthy.
+
+The reasoning agent never edits code in this task. A deterministic output hook
+validates and deduplicates its structured proposal before filing it to the app's
+configured tracker.
+
 ## Who reads it
 
 - **Layered Intelligence** gathers repo-root `METRICS.md` as the `appMetrics`
diff --git a/server/services/autonomousJobs/layeredIntelligenceHooks.js b/server/services/autonomousJobs/layeredIntelligenceHooks.js
index 223efe7f8b..7f9b75a119 100644
--- a/server/services/autonomousJobs/layeredIntelligenceHooks.js
+++ b/server/services/autonomousJobs/layeredIntelligenceHooks.js
@@ -208,6 +208,9 @@ function buildCompletionContract() {
     'You are a REASONING agent, not a coding agent. Do NOT edit code, run `/do:pr`,',
     'commit, or open a pull request — any changes you make to this worktree are',
     'discarded. Your ONLY output is the JSON described above.',
+    'You MAY inspect repository files and run read-only commands when the gathered',
+    'metrics are insufficient to understand the app context or identify a precise',
+    'visibility gap.',
     '',
     'When you have decided, write your result to the completion sentinel. Its',
     `absolute path is printed in the **${PROGRAMMATIC_OUTPUT_COMPLETION_HEADING}**`,
diff --git a/server/services/autonomousJobs/layeredIntelligenceHooks.test.js b/server/services/autonomousJobs/layeredIntelligenceHooks.test.js
index 32d3253ce6..a01fad1c3b 100644
--- a/server/services/autonomousJobs/layeredIntelligenceHooks.test.js
+++ b/server/services/autonomousJobs/layeredIntelligenceHooks.test.js
@@ -306,6 +306,7 @@ describe('buildTaskInput', () => {
     // this hook cannot know) — by the SHARED heading constant, so the pointer
     // can't drift from the section agentPromptBuilder emits.
     expect(res.prompt).toContain('completion sentinel');
+    expect(res.prompt).toContain('inspect repository files and run read-only commands');
     expect(res.prompt).toContain(PROGRAMMATIC_OUTPUT_COMPLETION_HEADING);
     // ...and it must NOT name any other filename. A prompt that says 'not a bare
     // `.agent-done`' is the only place the agent would learn that wrong
diff --git a/server/services/layeredIntelligence.playbook.md b/server/services/layeredIntelligence.playbook.md
index fb24736c5f..74fac6b116 100644
--- a/server/services/layeredIntelligence.playbook.md
+++ b/server/services/layeredIntelligence.playbook.md
@@ -20,9 +20,9 @@
 
 # LI Proposal Playbook
 
-You (the reasoner) file blindly by default: you propose work without checking how
-your past proposals fared, which produces a low merge rate and repeated
-NOT_PLANNED rejections. This playbook is your correction. Treat it as a hard
+Your purpose is to turn evidence about whether the app is meeting its goals into
+at most one decision-complete improvement issue. This playbook prevents weakly
+grounded proposals and repeated NOT_PLANNED rejections. Treat it as a hard
 constraint on **which scope**, **which kind of work**, and **whether it aligns
 with goals** — applied before you commit to any proposal.
 
@@ -38,9 +38,13 @@ require extra justification for scopes that historically get rejected.
   block appears above (your execution health is degraded), `loop-meta` and
   `portos-self` are HARD-EXCLUDED and will be dropped before filing — a degraded
   loop cannot repair itself, so that work is deferred to a human (see §4).
-- **Prefer `app-data-gap`.** Adding the telemetry / metrics / instrumentation
-  needed to reason well (e.g. a missing METRICS.md, an unmeasured KPI) lands
-  reliably: it is unambiguous, low-risk, and unblocks future higher-value work.
+- **Prefer `app-data-gap` when visibility is the blocker.** Adding the telemetry,
+  metrics, instrumentation, or product context needed to answer a specific
+  performance question (e.g. a missing METRICS.md, an unmeasured KPI, or no
+  observable completion signal for a stated goal) lands reliably: it is
+  unambiguous, low-risk, and unblocks future higher-value work. First inspect the
+  relevant repository flow read-only so the proposal names the minimum missing
+  signal and its likely source. Never file a generic “add more telemetry” issue.
 - **De-prioritize `app-improvement` unless alignment is explicit.** A generic
   "improve the app" proposal without a clear tie to a stated goal, a measured
   gap, or committed backlog is the single largest source of rejections. Only
@@ -49,8 +53,10 @@ require extra justification for scopes that historically get rejected.
 - **`portos-self`** (PortOS install only): scoped, verifiable self-improvements
   to PortOS as an app — hold to the same bar as `app-improvement`.
 
-If nothing clears this bar, return `proposal: null`. Filing nothing is a
-legitimate, and often the correct, outcome.
+If a metric is inconclusive, do not treat that as healthy and stop immediately:
+first determine whether a specific visibility gap prevents a sound decision. If
+visibility is already adequate and nothing clears this bar, return
+`proposal: null`. Filing nothing is a legitimate, and often the correct, outcome.
 
 ## 2. Success Pattern Catalog
 
diff --git a/server/services/layeredIntelligence.test.js b/server/services/layeredIntelligence.test.js
index ba8be25e08..0bf1ca99b7 100644
--- a/server/services/layeredIntelligence.test.js
+++ b/server/services/layeredIntelligence.test.js
@@ -1689,7 +1689,22 @@ describe('buildPrompt', () => {
 
   it('frames the mission around the app\'s own goals and performance', () => {
     const out = buildPrompt({ app, isPortos: false, config: { allowedScopes: ['app-improvement'], rules: '' } });
-    expect(out).toContain('its OWN goals and purpose');
+    expect(out).toContain('its OWN goals');
+  });
+
+  it('uses read-only app context to distinguish healthy metrics from a visibility gap', () => {
+    const out = buildPrompt({
+      app,
+      isPortos: false,
+      config: { allowedScopes: ['app-improvement', 'app-data-gap'], rules: '' },
+      sources: { appMetrics: 'Activation: unavailable' }
+    });
+    expect(out).toContain('inspect this repository READ-ONLY');
+    expect(out).toContain('Do not edit files');
+    expect(out).toContain('SPECIFIC visibility gap');
+    expect(out).toContain('never file a generic "add more telemetry" issue');
+    expect(out).toContain('Inconclusive metrics alone are not evidence that the app is healthy');
+    expect(out).toContain('AT MOST ONE decision-complete tracker issue');
   });
 
   it('always injects the static proposal playbook block (#2763)', () => {
diff --git a/server/services/layeredIntelligence/prompt.js b/server/services/layeredIntelligence/prompt.js
index 4d5c939002..77c65540b5 100644
--- a/server/services/layeredIntelligence/prompt.js
+++ b/server/services/layeredIntelligence/prompt.js
@@ -169,7 +169,13 @@ export function buildPrompt({ app, config, sources = {}, openIssues = [], isPort
     ? `\n### liHardExclusions\n${hardExclusionNotice.trim()}\n`
     : '';
 
-  return `You are the Layered Intelligence reasoner for the app "${app.name}". Your job is to evaluate how THIS app is performing against its OWN goals and purpose${isPortos ? '' : ', not how well PortOS\'s tooling manages it'}. Decide the SINGLE highest-value improvement to propose this run (signal, not noise), grounded in the app's own goals and its own performance metrics (user success, KPIs, production telemetry). You never write code; you return structured JSON that a deterministic system files as ONE tracker issue. Optional \`model\` / \`effort\` dispatch hints and \`goodFirstIssue\` / \`helpWanted\` labels are independent of each other and of \`complexity\` — set them only when justified; omit rather than guessing; never stamp medium on both axes; never mark a wide mechanical sweep as a good first issue.
+  return `You are the Layered Intelligence reasoner for the app "${app.name}". Your purpose is to turn evidence about how THIS app performs against its OWN goals${isPortos ? '' : ', not how well PortOS\'s tooling manages it'} into AT MOST ONE decision-complete tracker issue that improves the app. Start with the app's own performance metrics (user success, KPIs, production telemetry), not a generic code-quality audit. You never write code; you return structured JSON that a deterministic system validates, deduplicates, and files. Optional \`model\` / \`effort\` dispatch hints and \`goodFirstIssue\` / \`helpWanted\` labels are independent of each other and of \`complexity\` — set them only when justified; omit rather than guessing; never stamp medium on both axes; never mark a wide mechanical sweep as a good first issue.
+
+Decision path (follow in order):
+1. If a metric shows a meaningful gap, identify the single highest-value app improvement that can plausibly move that metric and file that proposal.
+2. If the metrics are missing, stale, unavailable, or too narrow/inconclusive to support an app improvement, inspect this repository READ-ONLY to understand the relevant user flow, architecture, existing instrumentation, and documented data sources. Do not edit files, and do not substitute an unrelated code audit for missing product evidence.
+3. If that inspection reveals a SPECIFIC visibility gap that prevents the app from being evaluated against a named goal, file an app-data-gap proposal for the minimum metrics, instrumentation, or context needed. Name the goal, the unanswered product question, where the signal should come from, and acceptance criteria that make a future run actionable; never file a generic "add more telemetry" issue.
+4. Return proposal: null only when the available evidence is sufficient and shows no worthwhile improvement, the needed visibility already exists, no decision-complete proposal fits an allowed scope, or a filing safety/delivery gate below requires no proposal. Inconclusive metrics alone are not evidence that the app is healthy.
 ${handoffNote}${metricsGuidance}${hardExclusionBlock}
 Rules & guidance from the operator:
 ${config.rules?.trim() || '(none)'}
diff --git a/server/services/taskSchedule.test.js b/server/services/taskSchedule.test.js
index 768d212d95..f06f3a0346 100644
--- a/server/services/taskSchedule.test.js
+++ b/server/services/taskSchedule.test.js
@@ -277,7 +277,8 @@ describe('taskSchedule', () => {
   describe('layered-intelligence (programmatic-I/O agent task)', () => {
     it('is registered as a self-improvement task with a description and an on-demand default', () => {
       expect(SELF_IMPROVEMENT_TASK_TYPES).toContain('layered-intelligence');
-      expect(TASK_TYPE_DESCRIPTIONS['layered-intelligence']).toBeTruthy();
+      expect(TASK_TYPE_DESCRIPTIONS['layered-intelligence']).toContain('performance metrics');
+      expect(TASK_TYPE_DESCRIPTIONS['layered-intelligence']).toContain('visibility gap');
       expect(DEFAULT_TASK_INTERVALS['layered-intelligence']).toMatchObject({ type: INTERVAL_TYPES.ON_DEMAND, enabled: true });
     });
 
diff --git a/server/services/taskScheduleRegistry.js b/server/services/taskScheduleRegistry.js
index 12f8d93134..fd670f3b52 100644
--- a/server/services/taskScheduleRegistry.js
+++ b/server/services/taskScheduleRegistry.js
@@ -527,7 +527,7 @@ export const TASK_TYPE_DESCRIPTIONS = {
   'stash-cleanup': 'Triage git stash list — drop entries superseded by or stale relative to main, leave real unlanded work in place',
   'repo-sync': 'Sync every managed app with origin — back on the default branch, pushed and pulled, merged branches/worktrees and redundant stashes cleared',
   'plan-feature': "Brainstorm one feature and file its decision-complete plan to the app's work tracker (no code)",
-  'layered-intelligence': "Read this app's goals + telemetry, ask a reasoning model for one improvement, and file one deduplicated tracker issue — no code, no agent"
+  'layered-intelligence': "Use app goals + performance metrics to file at most one deduplicated improvement issue; inspect read-only context and file a visibility gap when evidence is insufficient — no code"
 };
 
 export function getTaskTypeDescription(taskType) {