Does a diagnostic correction reach the next incident?

Research brief, checked October 25, 2026. When an agent offers a plausible but wrong production diagnosis, how does an operator turn a discriminating observation into a check that a later agent retrieves and runs? A good answer traces error→human challenge→reviewed rule or instrumentation update→later independent incident→agent read/action→verified repair and original reporter, with scoped production access and human time. Primary sources show separate segments, not this complete chain; do not combine incidents into a fictional success.

Four subquestions and sources. (1) Challenged false diagnosis: Cindy Sridharan’s October 10 firsthand essay rejected two agents’ TLS claim by comparing affected and flags-off production hosts; roughly 400 versus five DNS addresses pointed to wrong client sharding. It shows no future agent using that signal or published post-fix canary. See incident notes. (2) Production correction store: Kiro’s August 21 operator account promotes a per-cell check into a rate-limit skill but gives no before/after diff or later retrieval trace; its raw-comment contamination shows a bad write path. Google SRE’s May 28 account indexes old incidents and reports playbook-updating agents, without a case-level read→outcome trace. Teotti shows a warning that stayed out of the next agent’s planning search. (3) Comparative separation: Wang et al., October 4 found +0.23 action frequency with revised skill preloaded versus +0.12 when the agent had to discover it on selected sandbox tasks; final artifact correctness differed by model. (4) Reporter/authority: Windows extension issue #122 remained open with no public original-reporter retest when checked October 25; it is not a skill-reuse trial. AWS’s September 30 Kiro Crew design describes two incidents explicitly labeled illustrative, not field evidence. Google's permission principles and Kiro's claimed read-only scopes are not published access audits.

Agreement and tension. Operator accounts recommend raw observations, human challenge and a conditional check available to a future investigator, rather than yesterday's root cause as universal rule. Kiro describes corrected lessons injected into later sessions; Wang’s separate skill-loading experiment shows that simply making text discoverable is weaker than preloading it. They are different systems, not contradictory results. Sridharan aspires to fewer interventions, but all three described cases required expert help; the second reviewer endorsed the first reviewer’s mistake without a new control.

Next discriminating test. Store one real warning with failure signature, applicability, source evidence, human disposition, owner and recheck; instrument the next similar task for index hit→body visible→control run→contradiction acted on. Compare the eventual diagnosis with production behavior and released repair with the original reporter’s retest. Count false triggers and pruning work, not only saved investigation minutes. Further generic queries yielded proposals or retrospective claims without the second incident and field endpoint; reopen on a dated public paired-incident trace.