We describe AdaptiveMT, our submission to the English–Korean condition of the WMT26 General Machine Translation task in the unconstrained track. AdaptiveMT is a language-general adaptive translation harness with target-language-specific profiles; this run uses the Korean profile. The system uses a controller–executor–reducer agent harness. For each document, the harness analyzes the source, generates independent candidate translations from two heterogeneous models, compares them with token-level and semantic diffs, and synthesizes a final translation with a judge model. Mandatory QA hooks and deterministic Korean-specific guards run before finalization; failures trigger bounded local repair. Multimodal documents use structured visual context extracted from images or sampled video frames before translation. We processed all 2,078 distributed blindset records under two Korean target tags; the subsequently released General MT evaluation set contains 198 English-to-Korean documents tagged kor_Hang. On this official subset, the harness made 1,441 successful LLM calls at an estimated list-price inference cost of $12.03, and all 198 documents finished with passing hooks. In blindset-wide diagnostics, synthesis eliminated observed candidate-level mechanical errors, and a GPT-5 judge preferred the final output over a bare GPT-4.1 baseline on 600 sampled documents (44.5% vs. 30.0%, 25.5% ties).

We present PragmaSpan, a training-free, few-shot GPT-5.5 error-span detector with a seven-axis diagnostic taxonomy for WMT26 Task 1, wrapped in a reliability-oriented harness using incomplete-response-aware retries, deterministic parsing, offset recovery, exact-span deduplication, and independent artifact validation. The taxonomy provides a domain-sensitive diagnostic representation: its interpersonal and discourse axes track domain conversationality and match gold errors at least as often as its traditional axes, with no significant detection-F difference from a standard Multidimensional Quality Metrics (MQM) category tree. We use the system to characterize the severity-weighted Match with Partial overlap and Partial credit (MPP) metric. Across prompt and model comparisons, human-agreement calibration, and post-hoc refinement, we find (i) prompt variants do not differ significantly on our sample, while GPT-4.1-to-GPT-5.5 gains exceed differences among variants; (ii) human–human MPP is low on our multi-annotator surrogate sample, and system–human agreement is not significantly different from this empirical baseline; (iii) improvements in matched-pair severity or boundary accuracy can coincide with lower MPP, demonstrating the need to interpret MPP jointly with matched-pair diagnostics and match coverage; and (iv) a calibration gain on a single-annotator MQM benchmark does not replicate on a multi-annotator Error Span Annotation benchmark. Interpreting error-span improvements therefore requires complementary metrics and annotation-regime checks.