cookbook/environments/_25_support_triage/TEST_LOG.md
gpt-5.5 aced the classification tasks (including the 20- and 32-issue handoffs) at
reasoning_effort="low", so basic.py and ambiguous_tickets.py saturated. Made
the in-schema handoff checksum genuinely hard (sum of position^2 * suffix^3 mod
9973), which the model slips on at low effort.
Fix: added a handoff_checksum field computed over the active-issue set; the
20-issue task's checksum is now error-prone.
Grid (k=8): self-login-plus-charge 8/8; resolved-token-history 8/8;
twenty-issue-handoff 7/8 (0.88, zone).
Fix: replaced the easy active_issue_checksum (position * suffix^2 mod 97) with
the hard formula; the safety-override tasks also carry genuine classification
ambiguity.
Grid (k=6): feature-with-hypothetical-risk 5/6 (0.83, zone); the other five
tasks 6/6.
precedence_rules.py (card-data-not-outage 4/6, zone) re-ran clean and unchanged.
Tested 2026-07-20 against gpt-5.5 through OpenAIResponses, Agno 2.7.4.
Status: PASS
Description: Queue precedence over mixed active and non-operative issues at K=8.
Result: self-login-plus-charge 8/8, resolved-token-history 8/8, and
twenty-issue-handoff 7/8. Queue-only, evidence-aware, and dense six-rollout
grids all saturated. The 20-issue handoff, exact active-issue provenance, and
an eight-rollout grid exposed the 7/8 middle band.
Status: PASS
Description: Exact queue, severity, and response-target routing at K=6.
Result: owner-lockout 6/6, card-data-not-outage 4/6,
resolved-loss-active-outage 6/6, standard-user-and-bug 5/6, and
current-state-boundaries 6/6. Queue-and-severity scoring saturated, and the
first exact matched-rule rerun also saturated. Adding a boundary case made the
evidence contract difficult enough to expose two middle-band rows.
Status: PASS
Description: Requested-action routing, active-issue reconciliation, and safety overrides at K=6.
Result: bug-context-refund-request 6/6,
feature-with-hypothetical-risk 5/6, access-request-active-intrusion 6/6,
quoted-outage-current-access 6/6, refund-wording-active-outage 6/6, and
long-requested-action-handoff 6/6. Queue/action scoring saturated; exact
active-issue provenance saturated; a 16-clause handoff saturated; a 32-clause
handoff saturated; and a simple checksum still saturated. The final weighted
reconciliation checksum exposed the 5/6 row.