Put the Answer Before the Explanation: 1.8 Seconds Down to 0.56, and What It Costs
Part 5 of 6. If the model writes its explanation first, nothing is usable until it finishes. If the decision comes first, a screen or a downstream step can start right away. I tested both orders, with and without a cap on the explanation, on 500 tickets. The order roughly tripled how fast the answer appeared but barely changed total time, and the cap saved less than I predicted.
Bart Porcincula · Published
Finding
With Qwen3.5-9B returning three decisions plus notes as JSON on 500 tickets, putting the decisions before the notes made the answer usable in a median 0.56 s instead of 1.80 s while total response time was 1.53 s against 1.85 s; capping the notes cut output by 1.6x rather than the predicted 2x, and cost 5.2 points of queue accuracy when the notes came first.
Decisions-first versus notes-first structured output (Oct 2026)
A 2x2 comparison on 500 support tickets: the three decision keys before or after a free-text notes field, and the notes uncapped or capped at about 40 words. Output is constrained to a fixed JSON shape. Timing is measured on the server's stream. Accuracy differences are paired bootstrap 95% intervals against the notes-first uncapped arm. Three repeats at one request at a time produced identical answers; a fourth run at four requests at a time gives throughput.
| Measurement | Value | Note |
|---|---|---|
| Time until the decisions are complete (median): notes first, uncapped | 1.80 s | the reasoning has to finish before the answer appears |
| Time until the decisions are complete (median): decisions first | 0.56 s | 31% of the notes-first time |
| Total response time (median): notes first vs decisions first | 1.85 s vs 1.53 s | decisions-first is mostly a perceived-latency win, not a throughput one |
| Output tokens per ticket (mean): notes first uncapped / capped | 117.6 / 72.0 | a 1.6x cut, short of the 2x predicted; the fixed JSON keys are a cost cap cannot touch |
| Queue accuracy: notes-first capped vs uncapped | 26.4% vs 31.6% | -5.2 pt [-8.6, -1.8] |
| Queue accuracy: decisions-first uncapped vs notes-first uncapped | 31.4% vs 31.6% | -0.2 pt [-2.8, +2.4], no cost |
| Incident yes/no accuracy: decisions first vs notes first | 75.8% vs 77.8% | -2.0 pt [-3.4, -0.8], the one question where answering first cost something |
| Throughput at four requests at once (tickets/min) | decisions-first capped 120.4; notes-first uncapped 78.4 | notes-first capped 109.3; decisions-first uncapped 92.0 |
| Unparsed or truncated outputs | 0 |
The short version
Ask a model for a structured reply with three decisions and a short explanation, and you choose which comes first in the format. That choice decides when your code or your user can act. All six use the same setup: the open Qwen3.5-9B model compressed to 4 bits, run through llama.cpp on one 16 GB RTX 5060 Ti, at temperature 0 (the setting that makes a model as repeatable as it gets), on 500 customer-support tickets. Each experiment's questions, thresholds and analysis were written down before it was run.
- Decisions first makes the answer usable about three times sooner: median 0.56 seconds against 1.80 when the notes come first.
- It barely changes how long the whole response takes (1.53 s against 1.85 s). The model still writes the same notes. You are changing when you see the useful part, not how much work is done.
- Capping the explanation saves less than you would expect. I predicted halving the output. It fell 1.6 times, because the fixed structure around the decisions costs tokens that a cap cannot touch.
- The cap hurt when the reasoning came first: queue accuracy fell 5.2 points. With the answer first, the cap cost nothing measurable.
- Answering first is not free. On the yes/no incident question it was 2.0 points worse. Queue and priority showed no benefit from reasoning first.
Takeaway: if a person or a downstream step is waiting, put the decisions at the front of the format and treat the explanation as a log for later. Expect a latency win, not a throughput win, and check the one question type that might genuinely benefit from thinking out loud.
The four arms
| Arm | Order | Notes |
|---|---|---|
| Notes first | explanation, then the three decisions | uncapped |
| Notes first, capped | the same | explanation limited to about 40 words |
| Decisions first | three decisions, then the explanation | uncapped |
| Decisions first, capped | the same | explanation limited to about 40 words |
The output is held to a fixed JSON shape, so the model cannot wander off format, and I measured the time to the point where all three decision fields are complete.
Results (one request at a time)
| Arm | Decisions ready | Done | Output tokens | Queue | Priority | Incident |
|---|---|---|---|---|---|---|
| Notes first | 1.80 s | 1.85 s | 117.6 | 31.6% | 39.8% | 77.8% |
| Notes first, capped | 1.16 s | 1.20 s | 72.0 | 26.4% | 37.8% | 76.6% |
| Decisions first | 0.56 s | 1.53 s | 97.5 | 31.4% | 40.6% | 75.8% |
| Decisions first, capped | 0.56 s | 1.07 s | 64.2 | 31.0% | 40.6% | 75.8% |
At four requests at a time, tickets per minute were 78.4, 109.3, 92.0 and 120.4 in the same order. So the combination of answer first and a cap is the fastest in practice, and the cap is where most of the throughput comes from.
Reading it
Why would reasoning first help at all? In theory, writing the explanation first gives the model a chance to think before it commits. On these three questions it did not help queue or priority. It helped the yes/no incident question by about two points. That is small, and it is on the easiest question.
The cap is a trade you can see. When the explanation comes first and is cut short, the model has less room to work its way to the answer, and queue accuracy drops. When the answer comes first, the cap only shortens a note that no longer feeds the decision.
What I would not claim
- Absolute accuracy is low (queue about 32%) because the questions are hard for this model and the labels are noisy. The comparison is between arms.
- The three repeated runs gave identical answers, so they are one deterministic result, not three independent samples. Only the timing varied, by under 2%.
- Timings are server-side. A real client adds network and display time to both orders equally.
- One model, one dataset, one key order, one cap wording.
Part 5 of 6. Previous: the phrase check that stops matching. Next: does an AI grader that sees the answer agree with it.
Cite this research
Bart Porcincula. "Put the Answer Before the Explanation: 1.8 Seconds Down to 0.56, and What It Costs." Published 2026-10-29; updated 2026-10-29. https://porsync.com/research/answer-first-schema-faster-first-useful-token
Permanent article URL