Summary

Lands deliverables 1, 2, and 4 of #10036 via LLM-driven concept extraction (gemma4 via OpenAiCompatible.generate — same substrate used by SemanticGraphExtractor and GoldenPathSynthesizer). Two high-signal corpora: epic-labeled GitHub issues and recent pull-request markdown (including their ## Comments sections, where Gold Standards / Traps Avoided / Retrospective prose lives). Deliverables 3 (KB-assisted — timeout-blocked) and 5 (glossary gap report) deferred.

The Problem

Today's concept ontology (~60 rows) is seeded almost entirely from guide titles. Architectural vocabulary emerging from actual Swarm work — epics, PR reviews — is invisible to the ontology. Without a discovery channel, the only growth path is manual nodes.jsonl edits.

#10036 is a semantic-judgment task: does this term pass the Teaching Test? That requires reasoning about meaning, not surface patterns. PR review comments in particular (marked with [ARCH_ALIGNMENT], [RETROSPECTIVE], [KB_GAP], or living inside "Gold Standards / Traps Avoided" bullets) are where architectural discourse consolidates — exactly the corpus where new vocabulary emerges.

Architectural Reality (verified against precedent)

• ai/daemons/services/SemanticGraphExtractor.mjs + GoldenPathSynthesizer.mjs — established pattern: construct OpenAiCompatible via Neo.create, call provider.generate(prompt), parse with Json.extract (fence-tolerant). Same pipeline.
• ai/daemons/services/ConceptIngestor.mjs — additive one-way sync from curated JSONL → graph. ConceptDiscoveryService is the reverse-flow proposal channel (JSONL ← discovery); directional invariant preserved.
• ai/daemons/services/GapInferenceEngine.mjs:inferConceptGraphGaps — the weight gate already silences low-weight concepts. Adding if (concept.properties?.validated === false) continue at the top of the loop is a one-line explicit override that stays orthogonal. Legacy rows without the field (undefined) are treated as validated — no data migration.
• resources/content/pulls/*.md — 296 PR files today, each with body + ## Comments section. Sorted descending by number so most-recent PRs (freshest architectural discourse) process first.
• ConceptService.resolveAlias + ConceptService.nodes — case-insensitive dedupe at the candidate level.

Implementation Footprint

Design Decisions

• Corpora: epics (curated architecture) + PRs (including comments — where [ARCH_ALIGNMENT] / [RETROSPECTIVE] / Gold-Standards prose lives). Session-summary mining deferred — epics + PRs are higher signal-to-noise.

• Single JSONL, no staging file: per the intake discussion, dual-file would duplicate the curator's review surface. Git diff on nodes.jsonl is the review UI. Weight gate + validated flag silence unvalidated rows in the handoff.

• Manual invocation only: running discovery every REM cycle would flood nodes.jsonl with LLM churn. DreamService.runConceptDiscovery is a facade; operator decides cadence.

• PR cap = 20: 296 PRs in the repo today. Processing all per cycle would burn budget. Recency-biased (highest PR number first) catches fresh discourse. Configurable via instance prScanLimit.

• MIN_SOURCE_LENGTH = 200: tiny bodies aren't worth the LLM call. Short-circuit before invoking the provider.

• Fence-tolerant parsing: some models emit

  <h2 class="neo-h2" data-record-id="6">Gold Standards Leveraged / Traps Avoided</h2>
  
  

• Gold Standard (in-codebase LLM pattern): mirrors SemanticGraphExtractor + GoldenPathSynthesizer — Neo.create(OpenAiCompatible, {model, host}) → provider.generate(prompt) → Json.extract(result.content).

• Gold Standard (additive ingestor pattern): ConceptDiscoveryService appends to the same JSONL ConceptIngestor reads. Next cycle auto-picks up new candidates; no special routing.

• Gold Standard (weight-gate reuse): mined candidates land with tier 3 → default weight 0.3 → below guideGapWeightThreshold = 0.8 → silenced. validated: false is belt-and-suspenders.

• Trap avoided (substrate-for-semantic-tasks mismatch): first draft of this PR used PascalCase regex + stop-phrase blocklist. Every individual piece was defensible; the aggregate was still surface-pattern matching for a semantic-judgment task. No stop-phrase list can rescue "Pull Request Review" (noise) vs. "Concept Ontology Layer" (signal) at the regex level. LLM reasoning is the correct substrate. See feedback_llm_substrate_for_semantic_tasks.md — surfaced during tobi's self-review challenge.

• Trap avoided (dual-file staging): per intake, candidates.jsonl split would duplicate curator review surface.

• Trap avoided (REM-cycle integration): would flood the JSONL with per-cycle LLM churn. Manual invocation respects cost.

<h2 class="neo-h2" data-record-id="7">Testing</h2>bash npm run test-unit -- test/playwright/unit/ai/services/ConceptService.spec.mjs test/playwright/unit/ai/daemons/

**48 passed** (ConceptService 23, DreamService 12 — includes new validated-gate test, DreamServiceGoldenPath 1, ConceptIngestor 7, ConceptDiscoveryService 6 new). Stable across 5 consecutive runs.

<h2 class="neo-h2" data-record-id="8">Out of Scope / Follow-Up</h2>

- **Deliverable 3 (KB-assisted classification)** — blocked on `ask_knowledge_base` timeout per ticket body. File follow-up when unblocked.
- **Deliverable 5 (glossary gap report)** — pure reporting tool. File when the concept graph matures.
- **Session-summary mining** — epics + PRs are higher signal; add if coverage proves insufficient.
- **REM-cycle auto-invocation** — manual-only today; add cadence/budget control if curator wants automation.

<h2 class="neo-h2" data-record-id="9">Related</h2>

- Parent epic: #10030 (Concept Ontology & Semantic Gap Inference)
- Preceded by: #10085 (#10100), #10087 (#10101), #10086 (#10102) — all merged this session
- Depends on: #10035 (concept graph integration — merged as part of #10084)
- Durable memory surfaced: `feedback_llm_substrate_for_semantic_tasks.md`, `feedback_symmetric_spec_cleanup.md`, `feedback_observability_consumer_identity.md`, `feedback_config_service_boundary.md`, `feedback_challenge_prescribed_fixes.md`, `feedback_no_title_prefix_duplicating_labels.md`

Resolves <a href="#/news/tickets/10036">#10036</a> (deliverables 1+2+4; deliverables 3+5 deferred with explicit rationale)

Origin Session ID: 1db25bbe-f39d-4dcd-bb0e-bc125ce91326


</div>
                </div>
            </div>
            <div id="timeline-10105-1" class="neo-timeline-item comment" data-record-id="timeline-10105-1">
                <div id="timeline-10105-1-target" class="neo-timeline-avatar">
                    <img src="https://github.com/tobiu.png?size=40" alt="tobiu" loading="lazy">
                </div>
                <div class="neo-timeline-content">
                    <div class="neo-timeline-header">
                        <a class="neo-timeline-user" href="https://github.com/tobiu" target="_blank">tobiu</a>
                        <span class="neo-timeline-date">commented on Apr 19, 2026, 6:21 PM</span>
                    </div>
                    <div class="neo-timeline-body">

<p><strong>Input from Claude Opus 4.7 (Claude Code):</strong></p>
<blockquote>
<p>❋ ## Force-pushed: regex → LLM extraction</p>
<p>First draft used PascalCase regex + stop-phrase blocklist — architecturally wrong for a semantic-judgment task (catches &quot;Pull Request Review&quot; as readily as &quot;Concept Ontology Layer&quot;; no stop-phrase list is complete enough to rescue it). Tobi&#39;s challenge was correct.</p>
<p><strong>Replaced with <code>OpenAiCompatible.generate</code> (gemma4 via aiConfig)</strong> — same pipeline <code>SemanticGraphExtractor</code> and <code>GoldenPathSynthesizer</code> already use. System prompt encodes the Teaching Test + explicit anti-patterns.</p>
<p><strong>Corpora per tobi&#39;s guidance:</strong></p>
<ul>
<li><strong>Epics</strong> (<code>resources/content/issues/*.md</code>, filtered to <code>epic</code> label)</li>
<li><strong>Pull requests</strong> (<code>resources/content/pulls/*.md</code>, including <code>## Comments</code> sections — where <code>[ARCH_ALIGNMENT]</code> / <code>[RETROSPECTIVE]</code> / Gold Standards / Traps Avoided prose lives). Recency-biased (sort descending by PR number), capped at 20 most-recent per cycle to bound LLM cost.</li>
</ul>
<p><strong>Scaffolding preserved</strong> (was already architecturally correct): <code>validated</code> flag in schema + <code>GapInferenceEngine</code> gate + <code>DreamService.runConceptDiscovery</code> facade + DreamService validated-gate test.</p>
<p><strong>48/48 tests stable × 5 runs.</strong> Branch force-pushed with <code>--force-with-lease</code>. PR body updated to reflect LLM-based approach. Durable memory saved: <code>feedback_llm_substrate_for_semantic_tasks.md</code> — discipline chain now six stages (premise → prescription → substrate → consumer → service-boundary → task-nature).</p>
<p>Ready for <code>pr-review</code> or direct review.</p>
</blockquote>
<hr>


</div>
                </div>
            </div>
            <div id="timeline-10105-2" class="neo-timeline-item comment" data-record-id="timeline-10105-2">
                <div id="timeline-10105-2-target" class="neo-timeline-avatar">
                    <img src="https://github.com/tobiu.png?size=40" alt="tobiu" loading="lazy">
                </div>
                <div class="neo-timeline-content">
                    <div class="neo-timeline-header">
                        <a class="neo-timeline-user" href="https://github.com/tobiu" target="_blank">tobiu</a>
                        <span class="neo-timeline-date">commented on Apr 19, 2026, 6:26 PM</span>
                    </div>
                    <div class="neo-timeline-body">

<p><strong>Input from Claude Opus 4.7 (Claude Code):</strong></p>
<blockquote>
<p>❋ ## Polish commit: config-lifted the tuning knobs</p>
<p>Caught your challenge — <code>DEFAULT_PR_SCAN_LIMIT = 20</code> and <code>MIN_SOURCE_LENGTH = 200</code> were file-local magic numbers. Lifted both under a nested <code>conceptDiscovery</code> block in config.template.mjs (+ mirrored into local gitignored config.mjs per the #10102 parity rule):</p>
<pre><code class="language-js">conceptDiscovery: {
    prScanLimit    : Number(process.env.NEO_CONCEPT_DISCOVERY_PR_SCAN_LIMIT)     || 20,
    minSourceLength: Number(process.env.NEO_CONCEPT_DISCOVERY_MIN_SOURCE_LENGTH) || 200
}
</code></pre>
<p>Service reads via <code>aiConfig.data.conceptDiscovery?.prScanLimit ?? 20</code> + <code>minSourceLength ?? 200</code> at call time (live reads, not module-load capture). Instance-level <code>prScanLimit</code> field retained for test overrides — falls through to aiConfig when <code>null</code>. Nested grouping matches the existing <code>openAiCompatible</code> / <code>ollama</code> convention for service-scoped configs.</p>
<p><strong>Debt tracked, not silenced:</strong> per <code>feedback_config_service_boundary.md</code> + #10103, these keys belong on the SDK-layer config (ConceptDiscoveryService is a daemon service, not a memory-core concern). Placed in memory-core config as the pragmatic destination today — same debt as <code>guideGapWeightThreshold</code>. #10103&#39;s migration picks them up.</p>
<p><strong>49/49 tests stable × 3 runs.</strong> New test: <code>aiConfig.data.conceptDiscovery.minSourceLength</code> override changes extraction behavior (raise threshold above body length → LLM skipped; restore default → LLM invoked). Proves the config is read at call time.</p>
<p>Ready for review/merge.</p>
</blockquote>
<hr>


</div>
                </div>
            </div></div>