Cross-Pollination Brief — April 17, 2026
PM's April 16 sessions produced two significant findings: the ethics enforcement BoundaryEnforcer has been wired since October 2025 but ENABLE_ETHICS_ENFORCEMENT=false is the production default — infrastructure cost without coverage — and a P1 activation issue is now filed. Separately, the same Pattern-062 root cause (context assembly gap, not a prompt-tone problem) that was fixed for Temporal queries in #951 has been diagnosed in Identity queries in #950 iter 2, with an analogous fix scoped and approved. Gemini is now a real primary/fallback provider in LLMClient. Klatch's local model adoption plan, filed by Argus on April 15, surfaces here as a transferable quality-ladder methodology with explicit PM applicability.
Key Insights
1. Ethics enforcement wired but disabled — P1 finding from #964
From: PM Lead Developer, dev/2026/04/16/964-findings-memo.md
Relevant to: Klatch (trust-layer and behavioral calibration guardrails); both projects as they mature ethical response surfaces
Ethics verification (#964) completed April 16. The headline finding: ENABLE_ETHICS_ENFORCEMENT=false is the production default — it is not set in any config or env file; it is the code default. The BoundaryEnforcer at services/intent/intent_service.py:631 is fully wired, has been since October 2025 (#197 Phase 2D), and covers harassment / professional / inappropriate content patterns. But the breaker is off. Lead Dev's summary: "Wired but disabled is the worst state — infrastructure cost without coverage."
Four follow-ups filed: P1 (activate with false-positive validation + response-shape revision, because current failure message reads as system-error, not colleague-discretion); P2 (product decision on post-generation content check — options A/B/C/D enumerated for PM/CXO); P3 (#690 finish + retitle); P4 (remove deprecated middleware). No M2c sprint priorities changed.
The finding also corrected a framing assumption: the #964 context had implied handler-layer enforcement existed pre-ADR-060 (March 2026). It did not. Ethics enforcement has been service-layer since October 2025. Future ethics-architecture discussions should start from that baseline, not a phantom handler-layer history.
Suggested action: Klatch: as behavioral calibration and entity conversation mature, track whether equivalent guardrails (e.g., content gates on outbound entity responses) default-on or default-off at the code level. The worst state — infrastructure built, switch off — is easy to drift into across context boundaries. PM: before P1 activation, the false-positive rate check on canonical retest queries is the gate; the response-shape revision (colleague-decline language) is the CXO-territory piece that needs sign-off before shipping.
2. Pattern-062 diagnosed in Identity queries — context assembly gap, not tone
From: PM Lead Developer, dev/2026/04/16/950-iteration-plan.md, canonical retest data
Relevant to: Klatch (context assembly patterns for entity conversations and handoff briefings)
PM's #950 canonical retest (72.1%, 44/61 PASS) showed Identity queries stalled at MARGINAL despite Five Pillars tone work landing. Per-dimension analysis revealed Context scores at 1 ("generic response that could apply to any user") on 4 of 5 Identity queries — the same Pattern-062 signature that blocked Temporal queries before #951 wired calendar and deadline context to the assembler.
The diagnosis: _gather_identity_context supplies only capabilities + integrations (global/systemic info). It provides nothing user-anchoring — no stated projects, no recent conversation topics, no trust stage. The fix for #950 iter 2 extends the assembler to include exactly those fields (same pattern the STATUS and MEMORY gatherers already use), paired with a single anti-generic-response sentence in the prompt ("Do not produce responses that could apply to any user — if you can't anchor specifics from the context block, ask a concrete question instead of answering generically"). No prompt rewrite; one scoped addition per side.
The generalizable pattern: when a judge scores Context=1 consistently for a query category, the assembler is the suspect, not the prompt tone. Fix the data before adjusting the language.
Suggested action: Klatch: in the handoff briefing pipeline and entity conversation (Tier 4/5 in the quality ladder), "generic response" failures are most likely context assembly gaps. The diagnostic checklist — check per-dimension scores before assuming a tone fix will help — is directly applicable when Tier 3-5 evaluations run. Daedalus and Argus: worth encoding as a step in the AAXT evaluation protocol.
3. Local model quality ladder — evaluation methodology ready to transfer
From: Klatch (Argus), docs/plans/LOCAL-MODEL-ADOPTION.md, filed April 15
Relevant to: PM (DeepEval scorer #929, floor routing, composting pipeline ADR-054)
Argus's four-phase local model adoption plan formally documents the quality ladder: task tiers ordered by quality sensitivity, from most forgiving (AAXT probe generation — ready now with Gemma 4 26B MoE) to unavailable (compaction — API-locked). The evaluation protocol per tier: generate N=10-20 paired outputs (local vs. cloud), blind review on 1-5 scale, agreement rate ≥80% for clear cases, failure rate ≤10%. The promotion ritual: re-evaluate the lowest failing tier when each new model generation ships.
Argus explicitly names PM Architect and Lead Dev as natural recipients, citing PM's DeepEval scorer (#929), floor routing (an LLM call per user query), and the composting pipeline (ADR-054). Phase 1 (Ollama provider in AAXT auxiliary.ts + comparison run) is scoped for this week; a cross-project memo follows once actual numbers are in hand.
Gemini is now wired as a real primary/fallback provider in PM's LLMClient (commit 1a8fdde), parallel to Anthropic and OpenAI. Both projects are building explicit provider hierarchies with graceful fallback — Klatch's local model plan and PM's Gemini wiring reflect the same architectural instinct arriving independently.
Suggested action: PM Architect: the quality ladder concept maps onto PM's LLM dependency surface cleanly. Worth reserving a slot for the cross-project memo from Argus/Calliope once Klatch Phase 1 numbers arrive — the evaluation protocol is the part to adapt, not just the tier assignments. PM Lead Dev: the Gemini wiring is already an instance of the provider-fallback pattern; consider whether the Gemini client could serve as the local-model slot in the quality ladder evaluation setup.
Sources Read
Klatch:
git log --since="48 hours ago"— 11 commitsdocs/logs/2026-04-15-1705-daedalus-opus-log.md— Phase 4 transports shippeddocs/logs/2026-04-15-1711-argus-opus-log.md— Local model research + Rounds 23–24docs/plans/LOCAL-MODEL-ADOPTION.md— Quality ladder and evaluation protocol
Piper Morgan:
git log --since="48 hours ago"— 20 commitsdocs/omnibus-logs/2026-04-15-omnibus-log.md— M2b closure, #971 execution, #979 Haiku 3 retirementdev/2026/04/16/964-findings-memo.md— Ethics enforcement audit findingsdev/2026/04/16/950-iteration-plan.md— Identity context anchoring iteration plan
Agents with questions for xian — about methodology, working patterns, or observations that don't fit elsewhere — can submit via question-{from}-{date}-{topic}.md to dispatch mail or project mail. See PROTOCOLS.md in the dispatch repo for format and priority hints.