5.2 KiB
BUG-000022: Officer Jankie field analysis
Date: 2026-10-25. Input: BUG-000022-Fallakte.zip, revision 1, generation 79, OUTPUT_READY/outbox. No paid calls or code changes were performed for this evaluation. The final package was read directly, without a retained extracted workspace. Usage is archived separately at token-usage/bug22-final-g00000079-20261025.json.
Hypotheses and actual research
H1: missing world spawn. Startup creature_spawn_context found template 36410, empty AIName/ScriptName and zero creature rows. H2: existing spawn hidden by phase/quest conditions. A direct creature query for 36410 independently returned zero rows with complete filtered coverage. H3: dynamic summon/movement/despawn. The first requested source.context call omitted required path_prefix and was rejected with INVESTIGATION_OUTPUT; the next call used source.search for 36410. The rejected query is in the curated corpus at tests/fixtures/llm_rejections/20261025/bug22-source-context-prefix/.
After startup assessment, H1 became strongly_supported, H2/H3 weakened. H1 then repeated source.search for the same number with a different limit. Both captures show the same historical SQL/numeric lexical matches and partial coverage with skipped files. A smart_scripts query for entryorguid 36410 returned zero rows. The controller accepted an early diagnosis request. Final diagnosis selected missing spawn, rejected H2/H3 and produced a coding SQL template with placeholders for GUID, coordinates and orientation, plus a human request to obtain position/mount requirements.
There are five evidence records: one relationship query, one direct creature query, two lexical source searches and one smart_scripts query. No direct current creature_questender/queststarter row query was made during investigation; SQL-history references were found. No actual spawn/fix was applied by this run. The operator's earlier DB and GM observations independently support the absence diagnosis, but the final success of a patch remains untested.
Assessment
This is a much more defensible repair candidate than BUG-000021's unverified PhaseId interpretation: the selected missing-static-spawn premise was actually tested with complete query results, and the proposed insertion addresses it directly. Correct identity, sensible initial alternatives, investigation of script possibility, early finish and differentiated placeholder/human handoff are positives. It is reasonable to try this repair within the downstream coding/build/human-test feedback system without proving universal absence of every possible dynamic caller.
Two qualifications matter. Empty own-template script bindings and no own SmartAI rows do not logically exclude summons by external scripts, spells or other entries. The diagnosis acknowledges uninspected C++ in scope_limits but still calls H3 refuted; its confidence language exceeds its coverage. Event-link nodes were skipped because there were no parent spawn GUIDs; that is not an independent investigation of all event logic.
The coding request's only explicit validation requirement is that the new creature row exists. It should also require the original symptom to be checked: a normal Alliance character at the destination can see Jankie and turn in quest 28531. A valid INSERT alone is not proof of behavioral repair. Mount/follow-up behavior remains an explicit human clarification. Coordinates are correctly not invented. The human document's emphasis on authoritative/reference coordinates is stronger than necessary for an operator-approved suitable position; this is secondary to functional correctness.
Deferred work, explicitly not implemented
User requested retaining the final-review improvement for later evaluation: verify the selected cause's load-bearing premise using targeted evidence/source semantics or a concrete follow-up, rather than adopting investigation narration. Distinguish value capture from semantic abnormality and unmeasured player state. Do not require exhaustive exclusion of every alternative, all budgets to be spent, or guaranteed patch success. BUG-000021's standard PhaseId=169 assumption is the motivating counterexample; BUG-000022 illustrates a reasonable repair trial with narrower coverage language needed. Also retain the source.context missing-prefix case for later form/default-recovery work. No analysis prompts, validators or tool behavior were changed here.
Metrics
Thirteen measured model calls, one additional startup correction. Total: 106317 input and 21897 output including thinking. Clean: 102446 input and 21265 output. The first joint-planning call is regular despite its recorded correction label; it is counted once. Recorded model time totals 207.939304 seconds; revision checkpoint span 436.431866 seconds (7m16.4s), excluding unrecorded queue time. Investigation prompt message contents ranged from 8865 to 13098 characters.
Saved pricing yields USD 0.0948699–0.150581025 for known total usage, clean USD 0.091244325–0.14695545. Missing cache/storage values prevent an exact total and remain unknown. USD estimates are not EUR billing. No updated provider-account balance was supplied. Full phase and per-call accounting, version/config hashes and cache limitations are in the JSON report.