Design in Product social media card
← Back to Hub substantive

Cross-Pollination Brief — April 18, 2026

Three threads closed in the last 24 hours. PM's ethics activation path cleared when CXO delivered voice guidance for Mode-2 decline responses — #992 ETHICS-ACTIVATE is now implementation-ready after weeks as a blocked P1. PM formally closed M2c with routing at 95.1% and quality at 72.1%, a +9.8 point improvement since the M2a baseline; M2d (MUX lifecycle) is next. And both Klatch and PM added DECISIONS.md files on the same day — a convergent infrastructure move explicitly designed to make the daily brief more reliable.

Key Insights

1. Ethics activation unblocked — CXO voice guidance received, #992 ready for implementation

From: PM Lead Developer + CXO, dev/2026/04/16/964-findings-memo.md, CXO inbox memos Apr 16 evening Relevant to: Klatch (behavioral guardrails, entity response surfaces); both projects as ethical response architecture matures

When #964 FLOOR-ETHICS-VERIFY closed April 16, the P1 finding was ENABLE_ETHICS_ENFORCEMENT=false in production — BoundaryEnforcer wired since October 2025 but switch off by default. Activation was blocked on CXO voice input for Mode-2 decline copy (how Piper says no). That input arrived the same evening.

CXO delivered three structured decline templates, anti-patterns to avoid, a Colleague Test auto-fail rule (Tone=0 = auto-fail, not MARGINAL), and a false-positive ceiling: 2-3% before beta, validated against the canonical retest corpus. The acceptance criteria are now fully specified in #992. The path to production: implement the voice layer, run canonical probes, validate false-positive rate, then flip the flag.

CXO also weighed in on the P2 question (post-generation response content check) — non-binding Option A preference but deferred the architectural decision to PM.

Suggested action: Klatch: as entity conversations deepen and Phase 4 transports mature, a guardrail that exists but defaults-off is a liability. Track the state of any behavioral gates (not just ethics — content shape, response scope, floor prohibitions) as a named bit, not implicit code behavior. "Is enforcement on?" should be checkable at a glance. PM: the Colleague Test Tone=0 auto-fail criterion is new and affects scoring across the whole rubric, not just ethics probes. Flag this for Argus before the next AAXT run against the full canonical suite.

2. Evaluation instrument taxonomy settled — Colleague Test stays R/C/T, fabrication separate

From: PM CXO (via cross-pollination), Architect; docs/end-of-day mail sweep Apr 16, #994, #995 Relevant to: Klatch (AAXT design, evaluation methodology); both projects

Three evaluation-methodology decisions landed April 16 evening, all traceable to cross-pollination inputs from Klatch's Architect:

  • #994 AAXT scorer vocabulary: Adopt the six-failure-mode taxonomy (filed as a PR for the AAXT scoring protocol). CXO endorsed. If the vocabulary is still mutable in the current scoring spec, now is the time to adopt it — it becomes the shared language for failure analysis across projects.

  • #995 Standalone fabrication probe set: Architect proposed 5–10 probes across 5 absence categories as a regression fence for the #960 guardrail. CXO endorsed. This becomes the fabrication instrument — structurally separate from the Colleague Test rubric.

  • Colleague Test rubric stays R/C/T: CXO explicitly ruled against adding a fabrication dimension to the Colleague Test. The argument: fabrication and conversational quality are distinct failure modes that need distinct instruments. Mixing them muddies both. Rubric stays three-dimensional; fabrication gets its own harness.

The generalizable principle: evaluation instruments should be specialized. Combining orthogonal concerns (tone quality vs. knowledge hallucination) into a single rubric optimizes for compactness but degrades diagnostic resolution. This is the same "separate the concerns" principle that Pattern-062 diagnosis applied to context assembly.

Suggested action: Klatch AAXT team (Argus, Daedalus): before the next major AAXT run, confirm whether the current scorer implements the six-failure-mode vocabulary and whether a standalone fabrication probe set exists. If PM is building #995 in parallel, there may be a shared probe set worth coordinating on — same absence categories, same failure taxonomy.

3. DECISIONS.md added to both repos on the same day — coordination infrastructure

From: Klatch (commit 04a0f36), PM (commit f8fdc29), both Apr 18 Relevant to: Both projects; cross-pollination sweep reliability

Both Klatch and PM added DECISIONS.md files today. The format: append-only, one line per decision, DATE | DECISION | PARTICIPANTS, with major ADRs as the full artifact and this as the greppable index underneath. The stated motivation on both commits: "anti-zombie brief checks" — giving the daily sweep something machine-readable to scan for decisions that crossed repo boundaries.

This is a convergent infrastructure move that emerged independently in both projects at the same time. The Klatch DECISIONS.md already has six entries (April 10–17); the PM version has four. Both capture decisions that were previously scattered across commit messages, memos, and session logs.

Suggested action: Both projects: commit to a session discipline of appending to DECISIONS.md whenever a project-level call is made — especially decisions that affect the other project or surface methodological choices. The brief can then do a targeted grep rather than full log scan. Longer term: the format could be the seed of a lightweight ADR index.

4. M2c closed — routing 95.1%, quality 72.1%, M2d next

From: PM Lead Developer, docs/internal/planning/m2-structure.md, Apr 16 Relevant to: Klatch (project milestone context; floor + quality arc)

M2b (test infrastructure: 5/5) and M2c (conversational depth: 6/6) both formally closed April 16. M2c closed at canonical retest Run 5: 95.1% routing (vs. 93% at M2a), 72.1% quality (+9.8 points from the 62.3% M2a baseline). The gap to the aspirational 80% quality ceiling is primarily a fresh-account context problem tracked in #989 — users with no projects or history can't receive anchored responses, so Context scores stall at 1.

M2d (MUX lifecycle) is next. Fourteen follow-up issues filed during M2c are organized by family (context assembler, ethics, LLM infrastructure, testing, hygiene) — none block M2d start.

Suggested action: Klatch: the fresh-account ceiling is a variant of the cold-start onboarding gap surfaced in Klatch's own Alpha tester work (April 2026 brief). Both projects have the same structural problem: high-quality responses require anchoring context, but new users have none. Worth comparing approaches — PM's solution involves explicit user-anchoring fields in the context assembler; Klatch's will likely involve channel bootstrapping and entity initialization.

Sources Read

  • Klatch: DECISIONS.md (new, Apr 18)
  • Klatch: docs/mail/memo-janus-to-calliope-apr15-brief-redacted-2026-04-16.md
  • PM: docs/internal/planning/m2-structure.md (M2b/M2c closure, Apr 16)
  • PM: commit body a6b3a6d — #964 ethics verification findings summary
  • PM: commit body d9f9b3f — #950 Five Pillars floor prompt evolution
  • PM: commit body 52059b3 — end-of-day mail sweep, #993–995 filed, CXO voice guidance
  • PM: commit body 4f5ee41 — canonical retest iter 2 (44/61 PASS, 72.1%)
  • PM: DECISIONS.md (new, Apr 18)
  • PM: docs/omnibus-logs/2026-04-15-omnibus-log.md

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.

← Back to Hub