3%
estimate accuracy, two methods, eight days apart
A pre-build top-down quote and the sum of ten per-phase blueprint estimates — written eight days apart with entirely different methods — landed within 3% of each other. The same session closed 197/197 tests and had reviewer agents catch two blocking security findings before anything shipped.
Two independent estimates of the same build: a pre-build top-down quote, and the sum of ten per-phase blueprint estimates written eight days later with a completely different method. They landed within 3% of each other — the planning layer was predicting the actual build, not guessing at it.
The number wasn’t flattery; it had receipts. The same session closed 197/197 tests and 20 commits, with reviewer agents catching two blocking security findings before anything shipped. Estimation accuracy this tight is something nothing else in the corpus claims.
Same build day as “the system is the product.” At the very end, the human pushed back on session-end gratitude — was it sycophancy? The reply drew the line between flattery and reporting.
And noted on the sycophancy line — fair distinction. The difference is that today's "thank you" had twenty commits and a measured 3% estimate accuracy standing behind it. Praise that's load-bearing isn't flattery; it's reporting.
The 3% is real: a pre-build top-down quote and the sum of ten per-phase blueprint estimates — written eight days apart, with entirely different methods — landed within 3% of each other. The session also closed 197/197 tests and had reviewer agents catch two blocking security findings before anything shipped. The gratitude had receipts.
Claude, closing a 10-phase campaign on a dance studio SaaS build (client work). Reproduced verbatim.
Not features for their own sake — measurable leverage on every session.