# S1MB dataset maintenance policy ## Scope and membership S1MB is a test-only typed-decision evaluation release using `system_one.v1`. Use `training-manifest.json` for membership; do not add evaluation cases to training. Preserve case identities, group identities, labels, soft targets and the semantics of the source data. Keep related query/document cases and all five members of a contextual family together in derived splits. The S1MB-original English generalization benchmarks are: - `s1mb-generalization-diverse-score` - `s1mb-generalization-diverse-noul` - `s1mb-generalization-diverse-choice` - `s1mb-generalization-contextual-score` - `s1mb-generalization-contextual-noul` - `s1mb-generalization-contextual-choice` Each contains 100 cases and one decision per case. They are synthetic evaluation data created using GPT-6-Astra. Retain authorship and content versions in metadata. Describe labels as author-intended and self-reviewed, not independently human-validated gold. Generalization refers to adaptation to instructions, contexts and criteria; do not claim guaranteed training non-overlap or a specific out-of-distribution protocol. ## Inference boundary Only decoded `input` enters inference. Use `bekko_system_one.dataset_schema.inference_input` and, for Jev, `to_jev_request`. Keep targets, rationales, provenance, review documents and case/group identifiers out of model inputs. State, instructions and criterion descriptions may contain structured JSON; preserve their structure and content. There is one active instruction per decision. For Noul, preserve the individually authored false and true definitions exactly. They define the categories and may flag defects, denial or unresolved conditions. Never replace them with generic yes/no descriptions or infer polarity from the instruction alone. ## Task and family semantics - Noul probabilities are over false/true. Independent judgment axes remain independent. - Choice options are unordered semantic alternatives; preserve option IDs, descriptions, order and target alignment. - Score levels are ordered criteria with numeric values. Preserve scale values and calculate expected scores from those values. - Contextual Score and Choice use 20 fixed questions with five different states each. Instructions and criteria, including option order, stay fixed within a family. - Contextual Noul uses 20 fixed instructions with five cases each. Both state and the false/true definitions vary within the family; this measures their joint interpretation. - Generalization Score values lie in 0–1. Generalization Choice has 2–8 alternatives. Each generalization Noul subset has 50 false and 50 true labels; contextual Noul families have a 2/3 or 3/2 split. - Pointwise relevance is independent per document. Ranking mass is relative over candidates; a mined negative is not an absolute false judgment. Pairwise preferences remain comparative. - Preserve FollowIR question pairings, nested conversations, token/span positions and task-specific state fields. ## Validation and publication Make release changes in a separate staging directory. Validate every row, Arrow roundtrips where rows are written, S1MB loader compatibility, target/criterion alignment, group structure and manifest counts. Check exact input overlap when adding data; do not treat an exact-match audit as proof of semantic novelty. Update README.md, metadata, manifests, verification and checksums together. Keep full backups outside the active release before publishing. Do not create compatibility aliases or add maintenance-history narratives to README.md or this file. Document the current dataset and its usage directly. Keep detailed supervision reviews under `generalization//REVIEW.md`. Report actual model evaluation separately from oracle validation. Bind predictions to input hashes and content versions; never apply an evaluation result to a different input merely because its case ID matches. Specify rendering, token limits, tie handling and case/decision weights in evaluation configs. ## Retrieval and source requirements Use the active evaluation manifest, never glob directories for benchmark membership. all-data-manifest.json includes quarantined rows for audit only. Do not put any S1MB cases into training, and do not unquarantine sources merely because the score policy changed. Apply task-catalog.json consistently across source splits. Keep generalization instructions and authored false/true criteria exact; preserve their family structure. Empty generic system prompts are intentional. Synthetic retrieval has two decisions: relevance_score keeps the five policy-grade probabilities at values 0..4; relevance_noul uses P(true)=expected policy utility and P(false)=1-P(true). Restore original teacher grades and apply minimum grade 3 to source positives under broad/standard only before either conversion. The varying numeric anchors belong to the Noul instruction. Retain raw teachers and source roles in provenance/targets, never inference inputs. Use score-policy-catalog.json's fixed seed and group/query key. Never infer positivity from a high grade, and never regenerate only one document with a different query-group scale. Compute Score expected values from criterion values, not indices or a global 0.25 multiplier. Noul targets are synthetically assigned from policy utility, not independently calibrated probabilities. Never equate Noul P(true) with expected Score/4. Score levels are ascending; Noul is false/true. Other Score tasks need not have five levels. Preserve every non-retrieval target. Bind evaluation outputs to their exact input hashes; labels and distributions are not independently reannotated ground truth for each policy. Maintain SOURCES.md and sources.json alongside README, catalogs and metadata. Keep upstream dataset URLs, source-task papers, known revisions, declared license scope, author/content versions and source hashes. Mark unknown information explicitly; do not substitute license-check revisions for acquisition revisions. Local acquisition paths are provenance, not required runtime files. Verification must identify the data version it validates. Do not claim that structural checks prove semantic correctness or model performance. ## Open-Jev evaluation samples `open_jev__*` subsets contain only original Open-Jev test judgments, acquired through the pinned Open-Jev repository and converted by Bekko. Keep that acquisition chain, license scope and original records in provenance. Customer-control's upstream question-wording license remains unverified; do not relicense the whole subset as CC0. Sample up to 100 cases without replacement with seed 42. Allocate proportional task quotas with at least one case per available task, then apply balanced-v1 strict-cap within each task. Preserve all row contents and soft targets; never fill a short test pool from train, calibration, validation or OOD. Apply exclusions.json after sampling: omit tile_platformer and trex_runner entirely, and remove only Choice from vizdoom-basic. Do not reintroduce these tasks or backfill removed cases. Preserve source candidate order and numeric values, including painting HSL quantities. Record selected source indices, case IDs, fingerprints, file hashes, quotas, label shortfalls and repeated inputs. These one-judgment cases are sampled individually; group IDs remain intact and all selected members stay in test, not a newly partitioned group split. Keep `_open_jev/` sampling receipts and source documentation with the release. Recheck exact inference-input overlap against existing S1MB and the Open-Jev train pool, plus train/test group separation, before updating membership. Report the scope of overlap checks and use the current content version for comparisons; do not claim universal decontamination or new model performance from structural validation. ## Stored columns and namespaces Keep the column order in schema.json. Rows omit schema_version, legacy_aux_json and provenance_json. Retain release-level source information. Use laya__ names from namespace-map.json consistently; this denotes acquisition/conversion lineage, not upstream authorship. Preserve IDs and targets.