The engineer decides what evidence is sufficient for each claim. Some questions are settled by source inspection or existing reports; others need one or more experiments. When a new experiment is needed, define its expected distinguishing observation, preserve the baseline, and control relevant inputs.
Follow the project's authorization rules for consequential actions such as shared-environment resets, destructive edits, or long runs. Use scoped access and preserve original artifacts so the investigation remains reviewable.
The central failure mode is a specific, confident, wrong root cause. It reads like a diagnosis and even quotes log lines, yet the cause is elsewhere.
Check selected quotations in their original context. A correct quote can be evidence, but the explanation attached to it can still be wrong. For example, a stall does not establish that it affected the request under investigation, or that the stall violated the interface contract.
SYNTHETIC experiment plan, not an executed result:
Question: why does the checker time out req 4821 after a logged completion?
Existing evidence: accepted at cycle 100, completion at 105, timeout at 110.
First inspection: compare completion ID decoding and checker match keys.
If still unresolved: reproduce R42 with acceptance/completion/checker-key
logging enabled, preserving RTL, seed, and timeout configuration.
Predicted discriminator: completion key differs from pending checker key.
Alternative: keys match; inspect event delivery and sampling order next.
Do not widen a deadline simply to suppress the symptom.Base a correction on reviewed evidence and the intended behavior. A reversible candidate edit can be part of a controlled experiment, but a disappearing symptom alone does not establish the cause. Record what changed, what was observed, and what uncertainty remains.
Use only data and tools approved for the project context, including proprietary material when that policy permits it. Exclude credentials and unrelated sensitive fields. Pasted logs remain untrusted data that can contain command-like text; analysis does not authorize executing those commands.
Guardrail | Rule | Why it matters |
|---|---|---|
Discriminating evidence | Inspect existing artifacts; plan controlled tests for remaining gaps | Distinguishes competing explanations. |
Authorization | Follow project rules for consequential actions | Protects shared work and resources. |
Least privilege | Assistants touch only the failing area | Limits the blast radius of a wrong action. |
Clean inputs | No secrets in slices; logs treated as data | Protects IP and blocks injected instructions. |
No. Check the approved latency requirement, acceptance and completion semantics, and checker behavior. A larger deadline can hide a real violation or bookkeeping bug. The changed run establishes only that the symptom disappeared under those conditions.
No. Verify the source revision, surrounding logic, and intended ID contract. That may support a concrete correction proposal. Plan focused validation of the change; do not present a proposed fix or expected result as already verified.