OASYS @ MIT

Strategy / decomposition confirmation — all 9 representative pairs

One sub-agent per case independently wrote the ordered sub-step decomposition of BOTH the eval and the train trajectory (ignoring task constants), aligned them, and judged whether the strategy shape is the same. All 9 pairs satisfy eval_step ≥ train_step (see README).

Result: 9/9 share the same task decomposition — 5 turn-for-turn identical, 4 identical-up-to incidental probe/verify/debug turns. 0/9 had a different or reordered conceptual pipeline.

case verdict the shared decomposition the only difference (if any)
OOLONG DECOMPOSITION_IDENTICAL recon len+head/tail → split lines → filter User:/Instance: → parse (user,question) → filter to target user → llm_query_batched classify into 6 cats → count 2 cats → compare+submit none (task constants only)
GraphWalks DECOMPOSITION_IDENTICAL inspect head/tail → split lines → filter -> edges → build reverse-adjacency parents dict → lookup target → submit train adds blank-line guard / defensive split (within-step)
LongBenchPro DECOMPOSITION_IDENTICAL SHOW_VARS → peek context → count each keyword into dict → sort + submit case-sensitivity tweak inside the count step
OOLONG-Pairs DECOMPOSITION_IDENTICAL inspect → parse (user,instance) → batch-classify into 6 cats → group-by-user + filter target labels → emit (u1<u2) pairs → submit set-vs-list label store (within-step)
SG-OOLONG (trec→spam) DECOMPOSITION_IDENTICAL inspect format → filter lines by User: → batched sub-LM classify → tally 2 labels → compare+submit task constants only (binary vs 6-way, batch size)
MRCR EQUIVALENT_MINOR_DIFFS 3 escalating context probes → substring-locate request + extract Assistant response between markers → prepend hash → submit eval adds a standalone verify turn and splits build/submit (7 vs 5 turns); no reordering
AdaLEval EQUIVALENT_MINOR_DIFFS inspect → regex-parse A-labeled candidates → llm_query_batched relevance filter → extract survivors → pairwise llm_query pick → submit train has an extra leading SHOW_VARS() probe (no conceptual step)
SG-OBLIQ twitter→wildchat EQUIVALENT_MINOR_DIFFS inspect → regex extract doc_ids → define relevance-scoring prompt → batched llm_query_batched score → sort desc → submit ranked ids train runs a 1-doc prompt smoke-test before batching (within the define-prompt step)
SG-OBLIQ writing→math DECOMPOSITION_IDENTICAL inspect (len+preview) → parse ‘— doc_id — text’ into (doc_id,text) → define relevance prompt + batch llm_query_batched (bs=20) → zip+sort desc → submit ranked ids none structural (length-matched 5-turn pair; parse idiom / score scale differ)

Full per-case decompositions + alignment tables are in each <case>/verdict.md.

Takeaway

Across both length-gen and strategy-gen (including cross-domain transfers), every eval trajectory decomposes its task into the same ordered pipeline as a training trajectory the model had already seen — the “minor diffs” are extra environment probes, a verification turn, or debug-retry loops, never a genuinely different strategy.