The rollover ladder, and Haiku for complex tasks

Young, 2026-10-10:

"GO model → GO free → ZEN. ZEN is last, not first — it is the only step that costs real money."

"Expensive models for some complex tasks — Claude Haiku 5.5 at $15/month."

Both are rulings, and the second accepts a risk I had just flagged. Recording them as given, not as I would have phrased them.

1. The ladder — Young's three steps

My proposal had four (GO model → GO model → GO free → ZEN). His has three:

Young's rule is simpler than mine and better for it. I had added a "GO model → GO model" hop to capture the per-model-ceiling problem. He collapses it into step 1, which means the rollover does not need to decide which other GO model — that is a routing decision, and it stays in the router. The ladder is a cost-order rule, not a routing table. Keeping those apart is the right call.

The governing principle, unchanged and now stated by him directly:

ZEN is last, not first — it is the only step that costs real money.

That single sentence settles the ordering question I had carried since the one-workspace correction. A bundled allowance rewards use-within-cap; a PAYG float punishes use. ZEN is where the fleet goes only when there is no free path left.

Recorded consequence for the ledger: the rollover is observable as billing_instrument moving go_allowance → none (GO free) → zen_payg. A jump straight from go_allowance to zen_payg is a defect — it means a free GO model was available and not used.

2. Haiku 5.5 is the complex-task escalation — accepted with a tight ceiling

Young authorises Claude Haiku 5.5 ($15/month) for complex tasks. He is accepting the fragility I flagged in 6F, and he is right to — the ceiling is the point, not the problem.

So the fleet has two different kinds of paid model, and they must not be budgeted alike:

• Volume models carry load. They exhaust by volume — 5,400 requests/day at V2's measured rate. Headroom is wide, so they are safe defaults.
• Haiku is a scalar escalation. At $3 per five hours, it exhausts after roughly 10–20 calls. That is a deliberate design constraint: if Haiku is being used often, the task mix is wrong, not the ceiling.

That reframes my 6F concern. I called the $15 ceiling "fragile as a fleet fallback." Young is not using it as a fallback — he is using it as a deliberately-scarce escalation. Those are opposite roles. Bulk escalation goes to GLM/MiMo; capability escalation goes to Haiku.

The rule that follows: the ledger must record which model drew, not just which instrument. Haiku's $3/5h means a task routed there is a decision, and the ledger's why_paid field is what makes it auditable.

3. What this closes, and the one item it leaves

Closed. No open questions remain in my lane except 6D and 6G.

4. The last two in my lane

6D is the one I would prioritise. Your retention rule — anything touching the vault, credentials or task text must use a zero-retention model — is the strongest rule in the whole design, and right now it is a policy footnote no code reads. That is exactly V2's failure pattern.