Orders enters shadow traffic

Every checkout now runs on both stacks in parallel. Parity is at 99.987% and climbing — nothing cuts over until it holds 99.999%. What this means for you: checkout stays safe, and you'll see it proven live on Harness.

Under the hood, every request against /tps — orders' primary API — now fires twice: once against the Rails 3.2 monolith that has served checkout since 2017, and once against the new Encore.ts orders service running the same rules in TypeScript. The legacy response is the only one that ever reaches a customer. The Encore.ts response goes straight into Harness, which diffs every field, every rounding decision, every timestamp, and files a verdict. Nobody outside this program should be able to tell the difference — that's the entire point of the exercise.

Harness has been quietly widening its net since Decision 0021 shipped: MBUS shadow-taps now cover 16 of the 33 repos folded into the monorepo, streaming live diffs instead of end-of-day batch comparisons. A run exits 0 for a clean match, climbs through 1-3 for shrinking categories of acceptable noise (timestamp skew, floating-point drift under a cent, field ordering), and exits 4 the moment it can't explain a divergence — which means trust nothing from that run and rerun everything, per the self-check policy Robert posted back in W31.

The rounding question is the one that actually stalled this milestone by two days. WorkLedger unit U-88412 — the refund-rounding port — hit a real disagreement between legacy's half-up rounding and the bankers' rounding Decision 0021 mandates program-wide, and four agents sat blocked on orders while it worked through Atlas. That question, and the two it spun off on MBUS coverage and SOX evidence, are linked below; if you have an opinion on rounding, this is the thread.

None of this ships anywhere near production without SOX sign-off. The accounting/invoices port went through its own human review lane in week 33 — first time the gate actually held, no agent could self-promote past it — and the same lane now watches every dollar figure that crosses shadow traffic on checkout. That's deliberate: Harness proves the two stacks agree, but a human still has to agree the proof is good enough.

Parity sits at 99.987% today. The remaining gap is almost entirely edge cases in international tax rounding and a handful of legacy discount-stacking rules nobody had written down anywhere except in Rails code from 2014. Cutover doesn't happen until the gate reads 99.999% for a full week — no exceptions, no manual overrides. Watch the number on Harness; it updates itself.

Fourteen comments came in within an hour of this posting, which tells you how many people have opinions about rounding. Three questions spun off into Q&A already — read them, and if you know the answer to any of them, Atlas is listening.