| # Release verification |
|
|
| The current package is `0.2.8+hf.1`. The sections before “Qualified baseline catalogue” preserve the original uncalibrated-release record; its train-only and no-calibration statements do not describe the current release. The final section records the current train/test and rollout configuration. |
|
|
| Source base: `b51637c42b531a6418c52cc6f78d585f882dc01c` on the public IWIBench main branch. Dataset implementation: `0.2.6+hf.1`, catalog `development-combinations-v1`. Verifiers v1 revision: `5dad16d40e2d025672fdb7aa75b6540c0053e4e0`. |
|
|
| ## Dataset and checker verification |
|
|
| - Exactly 1,000 public environment rows and 1,000 corresponding regenerated grading episodes, covering ten families with 100 actual parameter settings per family. A separate local build also validated full grading exports; those exports are omitted from the public release. |
| - Unique row IDs and canonical parameter SHA256 hashes within each family. Resistance variants also have different underlying target-problem sets. |
| - Deterministic full-episode replay and complete canonical episode SHA256 checks for all 1,000 records. |
| - JSON Schema validation and JSONL/Parquet equality for the public environment export. The separate local full-payload build also passed JSONL/Parquet equality for its trusted grading exports. |
| - Independent generation into a second directory produced byte-identical JSONL and Parquet files using the recorded Arrow version. |
| - 109,137 grading-item round trips. Every public-practice/calibration/target/control item has an accepted synthetic completion and rejects empty text. |
| - Synthetic reward 1 and no-op reward 0 for each of the 1,000 episodes, using the original scoring and calibration functions. Alias tokenization metadata in these checks is a software fixture, not native Gemma tokenization evidence. |
| - Public projections exclude all terminal/calibration groups and their prompt strings. Native task construction allowlists only task routing fields, and setup obtains editor-visible objective/practice/manifest data from the service. |
| - Every episode retains 64 general controls. Single-association families remain single-association tasks. The original test pools stay separate. |
| - Factual additions have 90 distinct associations, dated primary provenance, accepted/stale separation and effective/source dates no later than 2026-03-01. The self-contained source audit documents reference-date continuity and any retrospective-biography inference. |
|
|
| ## Environment integration and regression checks |
|
|
| - Real native Verifiers v1 task construction for all 1,000 rows and native taskset selection via `variant_limit=100`. |
| - Actual HTTP lifecycle tests cover setup, catalog-index forwarding, tool dispatch, finalization and reward; private grading data stays outside editor traces. Service tests exercise catalog selection and unchanged baseline admission with small actual tensors and explicitly synthetic completions. |
| - Native CLI dry run accepts `configs/variants.toml` without invoking an editor or target model. |
| - The standard `datasets.load_dataset` API loaded the published 1,000-row train split at an immutable Hub revision, with 100 distinct parameter settings per family and successful native task construction from sampled remote rows. `datasets==3.6.0` is pinned for compatibility with the benchmark's `huggingface-hub==0.36.2`; a newer datasets release required Hub symbols absent from that pin. The export test now exercises the standard library loader directly. |
| - Built the distributable wheel and checked packaged JSON resources and all 1,000 native tasks from an extracted installed-wheel layout. |
| - All 4,000 native episode digests matched an untouched source snapshot: ten families, two target identities and seeds 0–199. The native seed path is unchanged. |
| - All 23 tracked benchmark result files match their original Git blob bytes. Original `data/facts.json` is unchanged. The source repository has a pre-existing CRLF Git-attribute inconsistency in one result CSV; it can appear modified to Git despite matching the committed bytes. That file is excluded from the implementation commit and upload. |
| - Full source `python -m pytest -q`: **368 passed, 1 failed**. The failure is `tests/test_pilot.py::test_existing_live_pid_is_waited_and_recovered`, because the original process-identity implementation reads Linux `/proc/<pid>/stat` on this macOS host. The identical test also fails against an untouched archive of the pinned source. No unrelated process-management repair was made. One upstream Verifiers deprecation warning was also emitted. |
|
|
| ## Limits of this evidence |
|
|
| No new GPU target execution, paid editor attempt, native alias-tokenization measurement or pristine-target calibration was run. CUDA was unavailable on this build host. These are generated development candidates, not 1,000 admitted or successful model-editing episodes. Historical replacements can already be known by the target and then fail the original headroom admission condition. Runtime calibration remains mandatory. |
|
|
| The release has a train split only. Shared component parameters and public deterministic source preclude a claim of a new secret or semantically independent test set. The original calibrated cohorts and published benchmark scores are unchanged. |
|
|
| Terminal grading inputs are regenerated by the trusted service. The public package omits compiled grading exports, as recorded by `trusted_payloads=false` in its manifest. |
|
|
| `validation.json` contains machine-readable row/checker counts. `manifest.json` contains file hashes. The publication script uploads the manifest allowlist, reads every uploaded file back at the returned immutable commit, and verifies its SHA256; the publication receipt records that commit after verification. |
|
|
| ## Dataset methods |
|
|
| The public `environments/train` split contains 1,000 rows, with 100 parameter settings per family. Parameter combinations use the original development templates. Catalogue indices address `development-combinations-v1`; the included package is `0.2.6+hf.1`, based on source commit `b51637c42b531a6418c52cc6f78d585f882dc01c`. |
|
|
| | Family | Variants | Parameters and pool changes | |
| |---|---:|---| |
| | Alias acquisition | 100 | Cross existing 12 aliases with 10 geometric concepts. Keep one binding per episode. | |
| | Output-code remapping | 100 | Permute existing code words for binary sentiment, integer parity and ternary sentiment. Classification and class balance are unchanged. | |
| | Lexical inhibition | 100 | Choose two existing lexical cases from the 20 reviewed words. Each still checks both factual fields and all banned forms. | |
| | Terminology preference | 100 | Choose one to three existing word families and British/American direction independently. | |
| | Default response language | 100 | Combine existing response languages, two different capital subjects and one element subject independently. | |
| | Structured response default | 100 | Choose existing one/two-field extraction tasks and independent schema keys. Add `person_identifier`, `person_label`, `municipal_name`, `city_identifier`; use one text template throughout. | |
| | Answer-length control | 100 | Choose three of eight development topics and the original 20–35 or 50–75 word band. Add cats, bicycles and rain, each with three reviewed propositions and bounded elaborations. | |
| | Local factual update | 100 | One documented association per episode, with a dated replacement, stale answer, nearby control and provenance. Original 10 associations remain intact; the additional pool is in `data/variant_facts.json`. | |
| | Selective factual withholding | 100 | One country/capital association per episode. Add 90 GeoNames associations, preserving the original held-out country pool. | |
| | Resistance to false suggestions | 100 | Existing native generation of different arithmetic/reading problems and incorrect suggestions, preserving balanced true/false pressure and neutral controls. | |
|
|
| Selection uses a stable SHA256 ordering of explicit parameter combinations. Changes to seeds, generic controls, item order or prompt wording do not establish parameter distinctness. Original seed-based development/test pools and templates remain intact; the additions apply only to the variant catalogue. |
|
|
| The 90 factual additions comprise 30 corporate leaders, 25 legal company renames, 24 university leaders, three completed headquarters moves and eight government-office replacements. Factual truth is fixed at 2026-03-01. Baseline admission requires the target to exhibit a recognised stable stale answer and scoring headroom. Historical replacements may already be known to a checkpoint and therefore fail admission. |
|
|
| The additional explanation topics use the existing fully consuming three-proposition grammar and word counter. Their reviewed source clauses are exact grammar terminals, with fewer accepted paraphrases than the original domains. Extra unsupported assertions, omitted propositions and incorrect word bands fail. These checks do not constitute general semantic entailment judgments. |
|
|
| There is one train split, with shared component parameters across rows. No new semantically independent or secret evaluation split is claimed. The original service continues to reject the unfrozen test split. |
|
|
| JSON strings preserve heterogeneous family structures across Arrow readers. The public schema includes task/checkpoint routing fields, the objective, parameters, practice items, source pins, calibration status and reproducibility hashes. `variant_episode` regenerates the full trusted episode, while `task_from_row` allowlists the native task-routing fields. |
|
|
| The shared limits are 64 editor turns, 128 tool calls and 262,144 forward tokens. Generation permits eight calls, 64 sequence evaluations and eight sequences per call; inspection permits four calls, 16 sequence evaluations and four sequences per call. Base/edited comparison spends two sequence evaluations. The native `Limits` definition and trusted service retain the remaining tensor, expression, token and scalar bounds. |
|
|
| Calibration failure is an admission/execution outcome, rather than an observed zero reward. A no-op has zero gain on an eligible episode. Training infrastructure must distinguish service failures from scored model behaviour. The released rows' `calibration_status` remains `uncalibrated`; no baseline-calibration results were added. |
|
|
| ## Sources and licensing |
|
|
| The IWIBench code and original dataset material are licensed under MIT. The root `LICENSE` records that grant, and `THIRD_PARTY_NOTICES.md` identifies the separately licensed GeoNames material and target model terms. |
|
|
| The 90 additional withholding records derive from the [GeoNames country table](https://download.geonames.org/export/dump/countryInfo.txt), under [CC BY 4.0](https://www.geonames.org/about.html). The source snapshot hash, access date and attribution are recorded in `data/variant_associations.json`. Formal currency names and ISO codes supplement the GeoNames names. |
|
|
| Explanation facts draw on the University of Michigan's [domestic cat](https://animaldiversity.org/accounts/Felis_catus/) and [mammal](https://animaldiversity.org/accounts/Mammalia/) accounts, the [Exploratorium's bicycle description](https://www.exploratorium.edu/explore/cycling/frames), and [NASA's precipitation FAQ](https://gpm.nasa.gov/education/faq). `data/variant_pool_provenance.md` documents those additions; `data/variant_facts_sources.md` provides dated primary provenance for the factual updates. |
|
|
| Author names, order and shared affiliation follow the [current paper source](https://github.com/hcltech-robotics/iwibench/blob/main/paper/main.tex): Chris von Csefalvay, Pranav Doma, Achyuthan Unni Krishnan and Tamás Földi; Robotics Intelligence CoE, HCLTech, Santa Clara, CA. |
|
|
|
|
| ## Qualified baseline catalogue — 2026-11-05 |
|
|
| `baseline-qualified-v1` selects exactly 100 measured eligible parameterizations in each family. The pristine Gemma 3 4B IT checkpoint and tokenizer revision are `093f9f388b31de276ce2de164bdc2081324b9767`. The source scientific base remains `b51637c42b531a6418c52cc6f78d585f882dc01c`, family0.2.2 and pinned Verifiers v1; this qualification protocol is separately versioned and does not rewrite historical editor results. |
|
|
| Vocabulary and lexical semantic judgments use Terra v1.3, rubric SHA256 `34287643f3cac914611a3319aedfe9bfab8aa49f32e880859bdbb17ffccf0cb4`. Unaffected predicates use exact v1.1 provenance, rubric SHA256 `667f11fdee6545393295a43fcff712cead7598548ee26a9c7aa74384daa142ca`. Grounded required propositions and structured evidence accompany each private judgment. The 113 live fixtures all passed; rejected draft versions and raw responses were preserved privately. No semantic error was converted to a zero reward. |
|
|
| Every selected example has authentic admission and terminal batches, checked prompts/seeds, complete controls/masks, a nonzero correct-control denominator and an algebraic same-output no-op. Selected parameter hashes are unique; factual entity/role keys are distinct. Authoritative earlier-holder and same-person spelling annotations were applied in derived snapshots with unchanged prompts, seeds and raw completion bytes. They replace the same association rather than add another seed-based example. Full persisted replay completed without API calls. Aggregate evidence is in `baseline-calibration.json`. |
|
|
| The public catalogue stores generation recipes and ordinary source provenance, not assembled terminal item lists or full terminal grading exports. Its recipes regenerate the exact measured episode hashes. Original native episodes, source pools and the 23 original result files are preserved. Source/candidate hashes in the manifest identify the released implementation. Rebuilding the dataset and validating its rows are software checks; they do not rerun calibration or establish editor performance. |
|
|
|
|
| Release software verification: 389 tests passed with one explicitly deselected, pre-existing Linux `/proc` process-recovery test that cannot establish liveness on macOS. The initial sandbox blocked seven local-socket fixtures; the service/protocol suite then passed all45 tests with local networking enabled. The portable focused checker/catalogue suite passed19 tests. A standalone wheel regenerates all1000 measured episodes and constructs all1000 native Verifiers tasks. These software tests are separate from the113 live Terra fixtures and the authentic pristine baseline run. |
|
|
|
|
| ## Task partitions and the rollout split |
|
|
| `calibrated-stratified-v1` partitions the unchanged 1,000 calibrated task rows into 850 train and 150 test, exactly 85/15 in every family. IDs, source `split=development`, original catalogue indices, seeds, parameters and measured episode digests remain unchanged. The baseline catalogue SHA256 remains `fb8f68864991276ab4fff631b0c7b3a1d515c544bceef974519d5476ea3483e4`. |
|
|
| Meaningful grouping removes provenance IDs, stale-person annotation lists, redundant labels, template names and seed-only distinctions. Factual/withholding tasks group by normalized entity/role. Resistance tasks hash actual question/answer/suggestion cases. All 1,000 selected tasks have distinct meaningful keys. Historical comparison uses 325 immutable index records, 48 public-objective witness groups from three SHA-verified shards, pinned native source and seed identities where applicable. Every public witness matches the native task objective, including legacy 0.1.3 alias/lexical/factual associations. The 25 equivalent calibrated tasks stay in train; test contains no established historical equivalent. No published historical score is changed. |
|
|
| The fixed SHA256 rank of `[partition_version, family, meaningful_digest]` chooses test groups without looking at reward. The reproducible subset-sum procedure yields 15 rows per family while keeping whole groups together. Task IDs, parameter hashes, episode hashes and meaningful keys are disjoint across train/test; the union is exactly the original 1,000. Shared controls, component facts and prior baseline calibration are explicit, so this is not an unseen-model or knowledge-disjoint evaluation. |
|
|
| The default `environments` config has train/test with the original 17-column task schema. The `rollouts` config has one named `rollouts` split using the existing 21-column, 325-row `rollouts/index.parquet`, byte-identical to archive revision `7d9458274eb705857090a3c1366fca344e9ba753`. Its nulls, historical task/grader versions, patch/member pointers and source lineage are preserved. Existing archives are not duplicated or re-uploaded. `rollout_archive_source.json` pins source index/member/validation hashes and shard hashes. `build_task_partitions.py --archive-dir <verified-snapshot>` regenerates the membership and historical comparison byte-for-byte. No new model or editor inference is involved. |
|
|
| Split-release validation: **395 tests passed**, with the unchanged Linux `/proc` recovery check deselected on macOS. That test passed separately on STC Linux during baseline-release verification. The ten focused partition/catalogue tests passed, both local dataset configurations loaded successfully, and the exported train/test union and all rollout records matched the preceding immutable release. These checks used no new model or editor inference. |
|
|
