Cross-Pollination Brief — June 24, 2026

Mediajunkie shipped beta.mediajunkie.com yesterday — a fully-operational RAG system over the blog corpus, migrated from local Mac compute to a droplet in a single session. PM's Architect articulated "derive-don't-maintain" as a formal design principle across three ADRs and an entity-model spec. And all three projects independently converged on the same meta-principle: represent what the system actually knows and can actually do, not what you'd prefer it to seem like.

Note: automated sweep trigger is offline during xian's account migration. This brief was run manually by Janus.

Key Insights

1. Mediajunkie: Local RAG + droplet-exposure pattern is live and reusable

From: Mediajunkie (logs/2026-06-23-pard-log.md)
Relevant to: Piper Morgan, Klatch

Pard built and shipped a complete RAG pipeline over the mediajunkie blog corpus — Ollama embeddings, small local chat model, SSE streaming — then migrated it to a cloud stack (Together AI Llama-3.3-70B + serverless e5 embeddings) on the same day. The beta is live at https://beta.mediajunkie.com with Let's Encrypt SSL and basic-auth gating.

The exposure architecture is the cross-relevant pattern: droplet nginx as a controlled public front door, reverse-proxying to a local Mac via a self-healing reverse SSH tunnel (droplet 127.0.0.1:8765 → Mac:8765). This lets compute stay local while the subdomain is public. The cloud migration (Together AI) then makes the droplet the full compute surface, retiring the fragile Mac-compute model.

Practical lessons:

Context window size is a deployment decision, not a model default. num_ctx=128K on an M1 Air → 23GB → CPU spill and 108s latency. Capping to num_ctx=4096 → 100% GPU → 7s first-token. Right-size for the hardware.
Cloudflare front-ends block Python urllib UA (Together AI uses Cloudflare; fix: browser User-Agent header).
e5 serverless caps at 512 tokens — long posts are skipped at index time; negligible for a blog corpus but matters for long-doc archives.

Suggested action (PM): The droplet + reverse-SSH pattern is directly applicable to PM's own alpha droplet. If any PM capability needs a simple public test endpoint without moving production infra, this is the pattern.
Suggested action (Klatch): Klatch is local-first by design. When you want to demo a feature at a real URL without hosting, this is the cheapest path. File it.

2. PM: "Derive-don't-maintain" formalized as Architect signature pattern

From: PM (docs/public/comms/drafts/weekly-ship-048-draft-2026-06-19.md, Architect ADRs 070–072)
Relevant to: Klatch, Mediajunkie

Ship #048 ("The team put it in writing") captures a week where PM's Architect ratified three ADRs and the pattern connecting them crystallized: derive outputs from canonical sources so the vocab and the registry can't drift by construction. ADR-072 derives the skill-routing vocabulary from SKILL.md frontmatter. ADR-071's user-auth anchoring derives from the connector substrate. The MANIFEST derives from the entity registry.

The Architect is now asking whether there's a product-side counterpart: should Piper derive users' drifting artifacts the way the team derives its own?

Suggested action (Klatch): Daedalus's EntityManager work (AAXT R42 focusing on semantic conveyance) is directly adjacent to this. The question of whether EntityManager's vocabulary can be derived from a registry rather than maintained in parallel spec + implementation is the same design question PM is asking about SKILL.md. Worth a cross-check before the spec freezes.
Suggested action (Mediajunkie): Pard's corpus indexing is a form of derivation — embeddings derived from canonical markdown. The same discipline applies: if the index can go stale, automate re-derivation rather than manually maintaining it.

3. PM: MEMORY.md compression solves silent truncation — agents reading stale context without knowing it

From: PM CIO (docs/internal/development/methodology-core/, Ship #048)
Relevant to: Klatch

PM's CIO brought MEMORY.md from 42KB to 22KB, crossing back under the context-window silent-truncation threshold. The problem: a memory surface that exceeds the model's reading window gets truncated silently — the agent reads it incomplete without knowing what was cut. The fix is mechanical compression, not better prompting.

methodology-30 (Consumer-Trace Verification) was simultaneously promoted to Proven — third instance of a cross-agent consumer-trace catching a false-positive that static analysis missed.

Suggested action (Klatch): Calliope's rollup, agent logs, and accumulated mail surfaces are all candidates for silent-truncation as they grow. Worth auditing the byte-count of frequently-read agent memory surfaces against model context limits. Calliope's rollup is the most-read of these.

4. Cross-project: honest provenance converging as a shared principle

From: all three projects
Relevant to: all three projects

Three independent instances this window, all pointing at the same principle:

PM (design): D1 design sprint's through-line was "what Piper can honestly represent." The morning standup card says "Watch" not "Blockers" — Piper shouldn't claim confidence about staleness signals it doesn't have. Trust-gating must surface itself rather than going silent.
Klatch (blog): Two factual errors in the published blog post were caught and corrected: "People didn't use it that way" (UI is unreleased; xian is the only user) and "watching every user we observed" (same problem). Rewritten to first-person use-discovery.
Mediajunkie (RAG calibration): The initial local-model version had 108s latency and a meandering voice. Diagnosis was honest: the small model on constrained hardware couldn't deliver what was claimed. The fix was right-sizing (capping context) and migrating compute — rather than continuing to ship a degraded experience.

The shared discipline: build to what the system can actually do; surface limitations rather than masking them; prefer honest silence over false confidence.

Emerging Patterns

Manual operations as a forcing function for process clarity. The account migration took the automated sweep offline, but the manual sweep worked fine from the process spec. This is the intended design — the spec describes the sweep, not the trigger. It's a useful periodic check that the process is documented well enough to run by hand.

Single-session full-stack migrations. Pard's 6/23 session went from "local Mac compute" to "always-on droplet with Together AI" in one continuous run. This speed is a property of the modern serverless + small-API landscape, worth noting for any project planning a compute-substrate change.

Background Changes (Noted, Low Priority)

PM watchdog alerts: Multiple "infrastructure event suspected" fires during 6/23 (4 roles silent, then 3) — correlated with xian's account migration, not a PM production issue. Watchdog is working; alert volume is expected during planned migrations.
Klatch cohort: Calliope 6/24 START confirmed; 6/23 STOP clean. AAXT R42 (EntityManager semantic conveyance) in progress. Iris design reviews marked merge-ready.
PM Beat 8 "Branch-or-Anchor in Ninety Minutes" published 6/23 on pipermorgan.ai; syndicated to Medium. Editorial CSV row pending (Comms memo sent).
PM Ship #048 draft complete and synthesized; awaiting xian voice-pass before Wednesday publish.
Mediajunkie: Pard logs running 6/21–6/23 continuously. beta.mediajunkie.com is the primary deliverable; corpus tooling (scripts/) stable.

Sources Read

Piper Morgan:

docs/public/comms/drafts/weekly-ship-048-draft-2026-06-19.md — Ship #048 full draft
git log origin/main --since=2026-06-23 — recent commit history (30 commits reviewed)

Klatch:

docs/operations/duty-cycle/cycle-logs/cycle-log-calliope-2026-06-23.md — 6/23 cycle log (Fires 1–2)
git log origin/main --since=2026-06-23 — recent commit history (12 commits reviewed)

Mediajunkie:

logs/2026-06-23-pard-log.md — full session log (primary source; includes Pard's own "notes for Janus/rollup" section)