29s median, 182s at p95: the latency tail that rewrote the product

The ratio is 6.3x ^[13], but the shape under it is the part that decides anything. Sorted, the run has a fat body between 15s and 50s, a scatter in the 60-115s band, then five samples piled at roughly 182s and a single one at 210.3s ^[9]. That last sample is a 180s client timeout with a successful retry behind it, which means the retry itself came back in about 30.3s ^[16], near the median of 29s ^[2]. Part of the tail is not slow work at all. It is work that stalls, gets abandoned, and then runs at ordinary speed on the second attempt.

That distinction matters because it tells you tuning the prompt will not fix the tail. Fan-out then multiplies whatever the tail is. Porchlight's developer had sketched a design where a visitor announces a change and the server fans out across every open departure, so 18 departures means 18 turns: roughly nine minutes at p50, closer to an hour at p95 ^[10]. Even the best case in that design, one turn, is described as a coin flip between ten seconds and three minutes ^[19].

The deadline that survived the redesign is the honest number. WEB_DEADLINE_MS is set to 100 seconds ^[12], which sits well below the measured p95 of 182s ^[2]. Counting only the samples the writeup places explicitly, six of the 67 exceed 100s ^[14], so about one live turn in eleven ^[14] is designed to be given up on. That is a decision to hand a visitor a fast failure instead of a slow success, with everything else served from a precompute pass that writes each verdict to liveCache.json with a capturedAt stamp the UI shows ^[11].

On the quality side, the agent resolved 14 departures across 54 captured judgements, 11 of them scoring 0.00 against a twenty-line keyword baseline that ships in the repo, while refusing 35 non-matching pairs at 0.90 confidence or better ^[8][6]. So 79 percent of its wins are invisible to the dumb control ^[17], which is the whole argument for the agent being on the critical path rather than decorating it. The seed example makes the mechanism concrete: a member complaining about "quick clips" replacing long sit-downs shares no salient token with an announcement that "deep-dive interviews are back", so the keyword check returns false and no stopword list will rescue it ^[7].

One arithmetic note the writeup does not close: 14 resolutions plus 35 refusals is 49, leaving five of the 54 judgements in neither bucket ^[15]. With precision being the property that makes it safe to email people who already left once ^[18], those five are exactly the ones a buyer would ask about.

All of this is self-reported by the project's developer from a single batch run ^[1], and the run was taken after the architecture existed rather than before ^[3]. The control cost twenty lines ^[6]. The latency capture cost one batch ^[1]. The redesign cost the architecture.