What the model needs

Start at the earliest relevant error and inspect what preceded it. Later failures may be consequences, but earlier warnings, reset activity, or another component's log may explain the symptom. Preserve timestamps and the original artifact so the slice can grow.

Choose context by the question, not a universal line or token limit. A short excerpt that omits an acceptance event can be less useful than a larger, organized bundle. Ask the assistant to name missing evidence rather than invent events between the supplied lines.

Building a slice

A useful slice has five parts. Each part earns its place by locating or reproducing the failure.

The block below shows one slice layout as a schematic. Real line numbers and ids differ on every run.

1
2
3
4
5
6

slice layout (schematic, illustrative only):
test and seed: pkt_loop, seed 42
first error: line 8810, TIMEOUT req id 4821 after 500 ms
context: 20 lines before plus 10 lines after line 8810
config: tool flags, RTL revision, last green run
left out: the rest of the 400000-character log and all credentials

Exercise: find the missing event

1
2
3
4
5
6
7
8

SYNTHETIC log excerpt, not an executed run:
run R42, test pkt_loop, seed 42, RTL A, config C
cycle 100 generator: present req 4821
cycle 110 checker: TIMEOUT req 4821
Missing: acceptance observation and consumer activity for cycles 100..110.
Question: does this establish that an accepted request was dropped?
Expected answer: no. Presentation is not necessarily acceptance.
Request the interface acceptance definition and matching monitor events.

Include units and observation conventions. A simulated-cycle deadline, simulated-time deadline, and host wall-clock timeout are different evidence. Keep the checker configuration and approved latency requirement beside the symptom so the assistant cannot quietly substitute one for another.

Common mistakes

Related topics

• Ranked hypotheses with evidence

• Data, privacy and IP