𝕺’𝕹𝖊𝖎𝖑 👑🍀

MAC v0.2 - what shipped since launch Comms upgrade, so here is everything that changed under the hood, in one place. Routing The gate used to score a task once and commit. Now it re-scores after classify, so a task that looked cheap but arrives with 400k of context does not get stuffed into an 8k window and silently truncated. Tier ladder is explicit: nano → small → frontier. A task only climbs if the cheaper tier fails the eval, never because the prompt "feels hard". Dedupe Window went from exact-match to hash(intent + normalised args), 30m rolling. That caught the thing I did not expect: the same question arriving from two different agents in the same minute, worded differently. Those used to be two frontier calls. Batching 90s hold window, fan-in on identical intent. Added a bail-out: if anything in the batch is marked irreversible, the whole batch breaks apart and each item routes on its own. Batching is for reads, never for writes. Evals 500-case held-out set, and the promotion rule is now hard-coded instead of being my opinion: a cheaper tier ships only if it stays within 0.4pp of baseline. Shadow → 5% canary → default. Nothing goes straight to default anymore. Logs Every decision is replayable: inputs, the six answers, the route, the confidence, the latency. You can take any decision id and re-run it against the current policy to see if today's config would have made the same call. That is the part I actually use every day. Failure modes Fail-open: controller down means traffic passes untouched, not blocked. Idempotency keys on intake, so retries do not double-charge. Backpressure instead of letting queue depth grow quietly. What this cost [FEES CLAIMED] in creator fees claimed and spent on infra, model credits and the eval harness. [BUYBACK TOTAL] routed to buybacks so far, every one posted with its tx. Measured since v0.1 [TASKS ROUTED] tasks through the controller [% NEVER REACHED A MODEL] never reached a model at all [SPEND BEFORE] → [SPEND AFTER] [QUALITY DELTA] on the held-out set Next → routing logs public, so anyone can replay a decision without asking me → first connector, so this runs on someone else's queue and not only mine → per-intent policies, because support tickets and code review should not share thresholds The models were never the bottleneck. The routing was. $MAC