CSE Roadmap

Common tier · about 4 hours · 12 parts

Part 4 of 12 · 2 min read

Each extra point of quality costs more time than the last

Getting from a rough demo to a good product is fast; getting from good to reliable is where most of the real effort goes.

Think of sanding a piece of furniture: the first few passes remove visible roughness quickly, but getting from "pretty smooth" to "glass smooth" takes far more passes for far less visible improvement each time.

In plain words: the usefulness threshold — the quality bar a system has to clear before it's worth putting in front of users — is a schedule decision as much as a quality decision. That's because closing the "last mile" between a working demo and a genuinely reliable product costs disproportionately more time the closer you get to perfect.

A worked example. One reported real-world anchor has a team reaching 80 percent of its target experience in one month, then needing four more months to pass 95 percent. Extrapolating from that anchor, and assuming the remaining error halves at a constant rate, reaching 90 percent takes about 3 months from the start. 95 percent takes 5 months, 98 percent takes 7.6 months, and 99 percent takes 9.6 months. The jump from 95 to 98 percent — just 3 points — costs nearly as much additional time as the entire climb from 0 to 95 percent did.

The key lines. From code/mini-scope/lastmile.py:

# from code/mini-scope/lastmile.py
def time_to_quality(q, t1=1.0, q1=0.80, t2=5.0, q2=0.95):
    if not q < 1:
        raise ValueError("quality must be below 1")
    h = halving_time(t1, q1, t2, q2)
    return t1 + h * log2((1 - q1) / (1 - q))

Read it as: the formula works in terms of 1 - q — the remaining error — not q itself. Framing it this way makes the compounding cost visible. Going from 80 percent to 90 percent halves the error from 0.20 to 0.10, and going from 98 to 99 percent also halves it, from 0.02 to 0.01. It's the same relative difficulty, even though the absolute quality gain shrinks each time.

Usefulness threshold, last mile

The usefulness threshold is how good a system must be — on quality, latency, cost — before it's worth putting in front of users. The last mile is the slow final stretch of improvement between a working demo and a genuinely reliable product, where that threshold decision actually bites.

Example: the 95-to-98-percent jump costing nearly as much time as 0-to-95-percent did is the last mile in one number. It's exactly why picking the usefulness threshold is really picking how much of that expensive final stretch you're committing to before you've even started.