# Data dictionary The two Hugging Face configurations are derived tables of **recorded final-run measurements**, not training examples or new experimental runs. Both use a `test` split to label audit material. This naming does not create an additional holdout. Never randomly split rows and treat their shared environments as independent samples. | Configuration | File | Rows | Unit | |---|---|---:|---| | `parent_runs` (default) | `data/parent_runs.csv` | 480 | One method on one environment in a parent search | | `descendant_runs` | `data/descendant_runs.csv` | 384 | One descendant method on one environment | Each table has the same 46 columns. `experiment` plus `seed` identifies one shared environment. Method comparisons are paired within this group; rows are not independent replicates. There are 24 environments per experiment, 72 environments in total. `main` and `cpu` each contain eight parent and five descendant arms per environment; `revision` has four parent and six descendant arms. The `development` and `revision_development` folders are excluded. ## Identity and objectives | Field | Type | Meaning | |---|---|---| | `record_id` | string | Unique `experiment:phase:seed:arm` identifier | | `experiment` | string | `main` (seeds 1000–1023), `cpu` (2000–2023), or `revision` (3000–3023) | | `phase` | string | `parent` or `descendant` | | `seed` | integer | Environment seed; distinct from the private proposal/child RNG streams | | `arm` | string | Experimental treatment; names preserved from source | | `mode` | string | Lower-level search mode used by the runner | | `n` | integer | Total matrix dimension (28 in the reported final experiments) | | `initial_nnz` | integer | Initial work-repair nonzero entries | | `final_nnz` | integer | Final work-repair nonzero entries; **lower is better** | | `nnz_gain` | integer | `initial_nnz - final_nnz`; positive is improvement | All reported systems have four grades and four retained directions. The common constraint maps `d,Q` are given to methods, not learned. Nonzero count is an exact cost ranking for shared kernel requests under the declared meter; it is not a general intelligence score or an independent task-generalization metric. ## Resources | Field | Type | Meaning | |---|---|---| | `budget` | integer | Configured search-work cap; not necessarily actual work | | `cpu_budget_s` | float/null | Process-CPU allowance where supplied; blank for work-only runs and source records that omit it | | `work` | integer | Actual metered work for this search, including nested auditions where applicable | | `cpu_s` | float | Measured process-CPU seconds for this search | | `wall_s` | float | Measured elapsed seconds for this search | | `parent_training_work` | integer | Acquisition work charged to this descendant; zero for parent records and zero-acquisition controls | | `parent_training_cpu_s` | float | Corresponding measured acquisition CPU cost, or zero | | `total_work_including_parent` | integer | `work + parent_training_work`; one-off, unamortized total | | `total_cpu_including_parent_s` | float | Source total where present; otherwise sum of search and acquisition CPU values | | `kernel_work` | integer | Work for 32 final kernel calls; distinct from search work | | `audit_work` | integer | Separately reported final audit work; excluded from search totals | | `audit_wall_s` | float | Separately reported final audit elapsed time | | `operator_bytes` | integer | Python work-operator object size in the original runtime; not whole-process peak memory | The abstract word-RAM meter is **not** hardware FLOPs or CPU instructions. Actual timing is host dependent. In the CPU audit, candidate-granularity stopping permits small allowance overshoots. The full source JSON includes work-category ledgers, process peak RSS, and additional detail. For parent rows, the `parent_training_*` values are zero and the total equals that row's own search cost. For descendant rows, acquisition belongs to the specific parent treatment. Shared acquisition costs are repeated for counterfactual descendant treatments: **do not sum these totals across arms** and call that the overall cost of one deployment. No amortization assumption has been introduced. For some original zero-acquisition control rows, CPU totals were absent; the exported totals are explicit arithmetic derivations, not additional measurements. ## Search counts and exact checks | Field | Type | Meaning | |---|---|---| | `attempts` | integer | Proposals attempted | | `completed` | integer | Completed proposal computations | | `strict_accepts` | integer | Strictly improving accepted candidates | | `neutral_accepts` | integer | Accepted equal-score moves | | `unchanged` | integer | Completed proposals leaving the represented operator unchanged | | `worse_rejected` | integer | Worse-scoring rejected proposals | | `invalid` | integer | Proposals rejected by exact validity checks | | `budget_interrupts` | integer | Work-limit interrupts recorded by the runner | | `cpu_interrupts` | integer | CPU-limit interrupts recorded by the runner | | `tasks` | integer | Final kernel requests (32 per record) | | `task_failures` | integer | Full-kernel correctness failures (zero in all final records) | | `cross_ablation_failures` | integer | Failures after intentionally removing the two-sided term; **not full-method failures** | | `operator_digest` | string | Original truncated operator fingerprint; provenance aid, not an integrity/security hash | | `zero_progress` | boolean | Source zero-progress indicator | Omitted source defaultdict counters among `attempts`, `completed`, `strict_accepts`, `neutral_accepts`, `unchanged`, `worse_rejected`, `invalid`, `budget_interrupts`, and `cpu_interrupts` export as zero. They are not missing measurements. Counters need not partition all attempts; source runners track other event types as well. ## Paired-role fields and provenance | Field | Type | Meaning | |---|---|---| | `generator_nnz` | integer/null | Final discovery repair nonzeros where recorded | | `generator_changed` | boolean/null | Whether final discovery repair differs from initial repair | | `generator_accepts` | integer/null | Accepted discovery-repair updates where recorded | | `auditions` | integer/null | Nested paired generator evaluations where recorded | | `audition_work` | integer/null | Metered nested-evaluation work, already included in `work` | | `nested_cpu_s` | float/null | Nested-evaluation CPU time, already included in `cpu_s` | | `pair_certificate_failures` | integer/null | Failed pair witness checks where recorded | | `source_file` | string | Relative path to authoritative original JSON | | `source_pointer` | string | JSON Pointer to the exact source record, e.g. `/records/0` | Outside the specified zero-counter and total-cost derivations, absent values remain empty CSV cells and should be interpreted as null, **not zero**. The viewer exports deliberately omit event histories, nested ledgers, and some auxiliary fields. Those remain intact in the original JSON and CSV files. ## Treatment names Main/CPU parent arms: `recursive`, `frozen`, `boundary`, `coordinate`, `rebuild_coordinate`, `no_normalization`, `uncoupled`, `no_update`. Main/CPU descendant arms: `learned_parent`, `initial_parent`, `initial_parent_full_budget`, `coordinate_full_budget`, `rebuild_full_budget`. Revision parent arms: `dual`, `collapsed`, `static`, `rebuild_coordinate`. Revision descendant arms: `dual_generator`, `initial_generator`, `collapse_same_training`, `collapsed_trained`, `initial_full_budget`, `rebuild_full_budget`. `collapse_same_training` uses the work operator from the identical dual-parent training. It is the paired role-collapse intervention; `collapsed_trained` is a separately trained collapsed control. ## Rebuild without rerunning experiments ```bash python scripts/build_hf_tables.py --check ``` Omit `--check` to regenerate the two tables byte for byte from the original final-run JSON. This command does not rerun learning or change source data. The script sorts by experiment, seed and arm, keeps stable field order, uses UTF-8 and LF newlines, and asserts row counts, unique IDs and work ledgers. If files are intentionally changed, rebuild the release checksum manifest before using the uploader. Preserve the original provenance manifest.