# The Daily Synthesis — North Star > Pin this to the wall. Re-read it the day a new generator tempts you, the day a cheaper model tempts you on the editorial, and the day you wonder whether any of this is working. --- ## THE NORTH STAR > **A voice-protected daily publication that improves itself — The Synthesist's prose stays Opus-grade and sacred, while the smallest swarm that keeps it alive trains the model that runs it, so the system gets cheaper and smarter every week without the human touching the daily run.** --- ## What "ultimate" actually means Boiling the ocean, six lenses each painted a zenith: a self-improving organism, a publication you can **fork in an afternoon**, one **readers genuinely love**, a **citable open lab**, a **self-funding business**. They all describe the *same* machine. The adversarial critic's job was to remember that **one person on Apple Silicon cannot run 17 agents, court subscribers, *and* ship a research package without the voice rotting or the human burning out.** So the ultimate is the **intersection the constraints permit**: a *tight* publication whose editorial voice is generated only by a frontier model and **never** compromised; whose fine-tune flywheel — the one piece that touches no reader-facing prose — compounds quietly on local hardware; and whose every grander ambition (forkability, the leaderboard, subscriptions) is a **downstream artifact of the flywheel working**, not a parallel project competing for the human's ~10 hours/week. The dreamers describe the destination; the critic draws the road that doesn't end in a ditch. **The ocean, boiled, is this: the human decides what the publication should *say*; the machine handles everything else and gets better at it every week; and the voice readers showed up for never, ever degrades.** --- ## THE KEYSTONE > **Close the golden-path flywheel into a single hands-off `launchd` loop — capture the swarm's own gate-passing work → retrain the local model → eval-gate it on the leak-proof benchmark → auto-promote — while a CI gate makes it physically impossible to ship a non-Opus editorial.** Every lens converges here: - The **autonomous organism** *is* this loop. - The **forkable protocol** is this loop with a `setup.sh` wrapper — you can't fork what isn't yet automated. - The **self-improving lab** is this loop's `leaderboard.jsonl` + eval harness, published. - The **economic flywheel**'s cost collapse ($2,400 → ~$250/mo) *is* the local model absorbing the boring 80%; its operator tier is the loop, forked. - The **beloved publication**'s human moments survive because the loop frees the hours for them. - The **critic** blesses it *only* paired with the Voice Gate — the flywheel earns its budget specifically so Opus can stay on the editorial. **The infrastructure already exists.** `scripts/finetune/eval/flywheel.js` (the SHIP_OK gate), the leak-proof benchmark + 3-layer dedup (13-gram + Jaccard + MiniLM cosine), the MLX pipeline, `unified-gen/ollama_client.py` + the `llm_service.py` routing branch, `blackboard.py` / `track_run` / `daily_log.py`, the golden-path dataset we just distilled (`scripts/finetune/mlx-pipeline/data-comic-copy/`), and the trained 1.7B router one pointer-flip from shipping. **The missing piece is ~200 lines of orchestration glue and one CI guard — not a new system.** Run the loop untouched for a month and everything else becomes a wrapper. --- ## IT IS DONE WHEN… 1. **The Voice Gate is law.** A `scripts/ci/check-editorial-model.js` runs *first* in `verify.sh`; every `blog-posts/*.md` carries a `model:` frontmatter tag; CI fails red on any editorial generated by anything below the chosen frontier tier. Zero exceptions have ever shipped. 2. **The swarm is tight — core + infra, not sprawl.** Only the load-bearing content agents + watchdog run scheduled; the dormant/experimental generators are `STUB.md`'d, their plists removed, `launchctl list` returns the small set. (We already have the registry to enforce this.) 3. **The flywheel runs hands-off.** A weekly `com.flywheel.weekly` job captures → aggregates golden paths → SFT/DPO on MLX → evals on the held-out benchmark → auto-promotes when **McNemar p<0.05 and Δ≥+5pp**, rolls back on regression, logs the round to `daily-logs/`. The human hasn't hand-run the pipeline in 30+ days. 4. **The local model is in production and saving money.** The fine-tuned model handles the structured/internal calls that used to burn Opus (comic copy, tagging, labels); generator spend is bounded under **~$250/mo**, inside the credit pool. 5. **The loop is visible and self-aware.** `leaderboard.jsonl` shows ≥8 honest promotion rounds; Monday's editorial can reference last week's flywheel round via `_load_yesterday_log()`; `/status.html` surfaces "the model improved N pp this week." 6. **One human decision per week, ≤10 hours total.** A morning glance at a green `/status.html`; a Sunday look at the leaderboard; the rest is *direction*, not operation. No silent-failure firefight in the trailing 30 days. 7. **The installable editor is the engagement north star, and it's climbing.** Installs tracked via `installs-collect.js`; a sustained baseline; install logs feed the golden-path capture as real *usage* signal. --- ## THE PATH — horizons that compound Each horizon reuses what **exists**. ⚙️ = the **flywheel/automation** does it; 🧠 = the **human** does it. ### NOW — *make the voice unbreakable, shrink the surface* - 🧠 Ship the **Voice Gate** (`check-editorial-model.js`) as the first check in `verify.sh`; tag every editorial with its `model:`. Cheapest highest-leverage line of code in the project. - 🧠 **Retire-to-stub in one pass:** unload the dormant plists, `STUB.md` each, update `docs/GENERATOR_REGISTRY.md` + the CLAUDE.md pipeline table. Buys back ~5 hrs/week of monitoring — the burnout firebreak. - 🧠 **Ship the router** (the fine-tune is done): HF upload + flip `scripts/zeroclaw/transformers-model.js:24`. Keep the template/`dialogue_pools` fallbacks live. *(Already staged this session: routing layer, cost-tiering, the comic-copy MVP + golden-path dataset.)* ### SOON — *close the loop, prove it once* - ⚙️ Stand up **`capture-gen/`**: scan the hold-queue, quality-gen recs, blackboard publish-tracking, provenance-gated connections → write `data/golden-paths.jsonl` daily, deduped against the prior 30 days with the existing 3-layer pipeline. - ⚙️ Wire the **weekly capture → SFT/DPO → eval → auto-promote** loop (`com.flywheel.weekly`, Sunday late, after archive-gen). Reuse `flywheel.js`, the benchmark, the leaderboard. Gate on **p<0.05 + Δ≥+5pp**; roll back on regression. - 🧠 **Stabilize the editorial frontline:** confirm the WNMS arc rotation actually fails posts, the 0.85 quality gate catches regressions, field-notes adds POV (not summary). ### LATER — *let it run, let it be seen* - ⚙️ Flywheel runs untouched; a monthly reconnaissance flags underfit categories to `/deskhub/`. - 🧠 **A weekly Editor's Dispatch** — the unforkable human moment the freed hours pay for. The reader-loyalty moat the swarm code can never replicate. - ⚙️ Install logs feed the next retrain. The loop now learns from *usage*, not just output. ### SOMEDAY — *wrap the working loop for the world* - 🧠/⚙️ **Forkable:** `setup.sh` now wraps a *proven* loop; branding/schedule move to a single `publication-config.json` + `desk-registry.json`. - 🧠 **Lab:** publish `leaderboard.jsonl` + the golden-path dataset + the McNemar/dedup methodology — because the loop already produced rounds worth citing. - 🧠 **Revenue:** operator tier + subscriber tiers become viable *because* cost is already collapsed and the loop is already turnkey. Revenue is downstream, never the driver. > **Rule of horizons:** the FLYWHEEL owns capture, retrain, eval, promote, recon. The HUMAN owns *only* direction, the editorial voice, the Dispatch, and the Sunday leaderboard glance. If the human is doing daily ops in month 4, the loop isn't done. --- ## THE ANTI-GOALS — brightest lines not to cross - **🔴 Voice collapse.** Editorial prose is **binary**, not a dial — it sounds like The Synthesist or it doesn't. Never ship a downgraded-model editorial: not to save money, not "just this once," not during a thermal throttle. The local model's entire job is to *protect* the frontier budget, never to replace it on reader-facing prose. - **🔴 Scope sprawl.** "We built it so it must run" is the death-thought. Stubbed generators stay stubbed. A new agent earns its plist only by *replacing* one, or by passing the same quality + provenance gates the core agents do. Forkability/leaderboard/revenue are **outputs of the working loop**, never parallel projects. - **🔴 Model collapse.** Golden-path capture without the 3-layer dedup poisons training and leaks the benchmark. Never aggregate without deduping against the current held-out set; rotate a fresh held-out slice from *new* paths monthly. Archive full checkpoints, not just HF pointers — models age. - **🔴 Burnout.** The cap is **~10 hrs/week**. If the human is firefighting silent failures or babysitting a swarm, the system has failed *even if the code works*. Retire-to-stub is burnout prevention encoded as architecture. - **🔴 Heroic human effort as the engine.** The flywheel must do the compounding. The day the publication only improves because the human grinds harder, the North Star is lost. --- ## THE ONE GUIDING METRIC > **Frontier dollars spent per published editorial — while the Voice Gate holds at 100%.** **Secondary scoreboard (M7, 2026-07-03):** *% of ZeroClaw messages handled by the cheapest adequate tier* (the Routing diagnostics tab's `localShare` / tier split). The routing product — ZeroClaw's Auto mode (`docs/goals/router-model.md` M7) — is an **output of the same loop**, never a parallel project: same golden-path capture, same leak-checked benchmark, same fail-closed gate, same human-confirmed pointer. Its privacy rule is part of the anti-goals' spirit: route records carry `promptHash` + features only; prompt text never enters the public repo without an explicit export toggle, PII scrub, and owner review. If this number falls month over month — because the local model absorbs more of the boring 80%, because the flywheel keeps promoting cheaper-smarter models, because the swarm stays tight — **while the editorial is still 100% frontier-grade and installs keep climbing** — then every lens is winning at once: cost is collapsing (economic), the loop is compounding (organism), the voice is sacred (critic), the methodology is real enough to publish (lab), and the human has hours left for the work readers love (beloved). One number, and it bends the whole project toward the star. --- *Synthesized 2026-05-30 from a six-lens tournament (autonomous organism · forkable protocol · beloved publication · self-improving lab · economic flywheel · adversarial true-north critic). The pieces already exist; the keystone is to close the loop and guard the voice.*