We copy production.
All of it.

Every HTTP request and every message-bus event, recorded from live traffic and replayed against both stacks — before anything ships. 16 of 33 legacy repos live on MBUS, so we tap and copy that too.

PARITYDIVERGEDUNVERIFIABLE3 VERDICTS. NO MAYBE.

HTTP TRAFFIC

MBUS TAP

RECORDER
PII MASKED

LEGACY LAB
RAILS 3.2 · PINNED

MONOREPO
CANDIDATE REV

DIFF →
VERDICT

CAPTURE → REPLAY → COMPARE · DETERMINISTIC · RESUMABLE

01 / CHECKS — PLANNED · RUNNING · PREVIOUS. NOTHING HIDDEN.

COVERAGE 78% · DIVERGENCE 7D 0.4%

● RUNNING NOW

orders/checkout-flowfx 2814 / 4,200

vis/redemptionrecording · 13.2k events

◌ PLANNED

orders/gifting

AFTER FIXTURE REFRESH · THU

accounting/invoices

SOX LANE · W36 · HUMAN SIGN-OFF ⚙

✓ PREVIOUS — EVERY VERDICT LINKS ITS EVIDENCE

orders/checkout-flowPARITY 99.987%SELF-CHECK ✓bundle

orders/refundsDIVERGEDtotalCents: 4899 → 4901bundle

deal-catalog/publishUNVERIFIABLEMBUS_TAP_UNAVAILABLE

EVERY BUNDLE CARRIES: SMALLEST COUNTEREXAMPLE · NORMALIZATION LEDGER (EVERY EXCLUDED FIELD + WHY) · THE EXACT REPLAY RECIPE

02 / THE CLI — FIVE VERBS, NOTHING ELSE

the engineer's whole loop — 30 seconds↻ replay

▌

THE FIVE VERBS

record — capture traffic + mbus into fixtures

verify — replay, emit verdict + evidence

explain — render a verdict's bundle

lab — boot pinned EOL runtimes

selfcheck — the harness tests itself first

EXIT CODES — SCRIPT-SAFE, NO AMBIGUITY

0 PARITY

1 DIVERGED

2 UNVERIFIABLE + reason taxonomy

3 harness fault — never a verdict

4 self-check failed — trust nothing

03 / THE GATE — AND WHY THERE'S NO WAY AROUND IT

PULL REQUEST · CHECKS

✓ orders/checkout-flow — PARITY (99.7%, self-check passed)

✗ orders/refunds — DIVERGED (1 counterexample) → bundle

⚠ deal-catalog/publish — UNVERIFIABLE (MBUS_TAP_UNAVAILABLE)

verdict row: PARITY + self-check ✓→UNIT PROMOTED TO PR LANE

There is no "promote anyway" button.

The gate is a database rule, not a warning. A genuine override writes its own audited FORCED row — never disguised as PARITY.

SELF-CHECK, EVERY RUN

Record → replay against legacy itself must yield PARITY, or nothing the harness says is trusted. It tests itself the same way it tests the rewrite.

← Atlas ★HARNESS · CAPTURE → REPLAY → COMPARE · WEEK-6 RULE: DUAL-RUN LIVE OR M4 SLIPS, PUBLICLYNext: FleetOps →

We copy production.All of it.

We copy production.
All of it.