The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
changed: list<item: string>
child 0, item: string
commands: list<item: string>
child 0, item: string
metrics: struct<typed: struct<n_instances: int64, inclusion_rate: double, bundle_reduction_gmean: double, sta (... 418 chars omitted)
child 0, typed: struct<n_instances: int64, inclusion_rate: double, bundle_reduction_gmean: double, stability_drop_al (... 145 chars omitted)
child 0, n_instances: int64
child 1, inclusion_rate: double
child 2, bundle_reduction_gmean: double
child 3, stability_drop_all_mean: double
child 4, stability_drop_interior_mean: double
child 5, stability_corrupt_mean: double
child 6, basin_depth_drop_mean: double
child 7, basin_depth_drop_max: int64
child 1, labelfree: struct<n_instances: int64, inclusion_rate: double, bundle_reduction_gmean: double, stability_drop_al (... 145 chars omitted)
child 0, n_instances: int64
child 1, inclusion_rate: double
child 2, bundle_reduction_gmean: double
child 3, stability_drop_all_mean: double
child 4, stability_drop_interior_mean: double
child 5, stability_corrupt_mean: double
child 6, basin_depth_drop_mean: double
child 7, basin_depth_drop_max: int64
artifacts: list<item: string>
child 0, item: string
relaxation_rule: string
n_polysemy_pairs: int64
aggregate: struct<typed: struct<n_instances: int64, inclusion_rate: double, bundle_reduction_gmean: double, sta (... 418 chars omitted)
child 0, typed: struct<n_inst
...
list<item: struct<pair_id: int64, family_a: string, family_b: string, share_k: int64, (... 406 chars omitted)
child 0, typed: list<item: struct<pair_id: int64, family_a: string, family_b: string, share_k: int64, both_active_on (... 139 chars omitted)
child 0, item: struct<pair_id: int64, family_a: string, family_b: string, share_k: int64, both_active_on_shared: bo (... 127 chars omitted)
child 0, pair_id: int64
child 1, family_a: string
child 2, family_b: string
child 3, share_k: int64
child 4, both_active_on_shared: bool
child 5, correct_dominates_after_one: bool
child 6, ambiguity_retained_on_missing: bool
child 7, active_on_shared: int64
child 8, families_on_shared: int64
child 1, labelfree: list<item: struct<pair_id: int64, family_a: string, family_b: string, share_k: int64, both_active_on (... 139 chars omitted)
child 0, item: struct<pair_id: int64, family_a: string, family_b: string, share_k: int64, both_active_on_shared: bo (... 127 chars omitted)
child 0, pair_id: int64
child 1, family_a: string
child 2, family_b: string
child 3, share_k: int64
child 4, both_active_on_shared: bool
child 5, correct_dominates_after_one: bool
child 6, ambiguity_retained_on_missing: bool
child 7, active_on_shared: int64
child 8, families_on_shared: int64
branch: string
n_disjoint_families: int64
to
{'experiment': Value('string'), 'branch': Value('string'), 'relaxation_rule': Value('string'), 'aggregate': {'typed': {'n_instances': Value('int64'), 'inclusion_rate': Value('float64'), 'bundle_reduction_gmean': Value('float64'), 'stability_drop_all_mean': Value('float64'), 'stability_drop_interior_mean': Value('float64'), 'stability_corrupt_mean': Value('float64'), 'basin_depth_drop_mean': Value('float64'), 'basin_depth_drop_max': Value('int64')}, 'labelfree': {'n_instances': Value('int64'), 'inclusion_rate': Value('float64'), 'bundle_reduction_gmean': Value('float64'), 'stability_drop_all_mean': Value('float64'), 'stability_drop_interior_mean': Value('float64'), 'stability_corrupt_mean': Value('float64'), 'basin_depth_drop_mean': Value('float64'), 'basin_depth_drop_max': Value('int64')}}, 'polysemy': {'typed': List({'pair_id': Value('int64'), 'family_a': Value('string'), 'family_b': Value('string'), 'share_k': Value('int64'), 'both_active_on_shared': Value('bool'), 'correct_dominates_after_one': Value('bool'), 'ambiguity_retained_on_missing': Value('bool'), 'active_on_shared': Value('int64'), 'families_on_shared': Value('int64')}), 'labelfree': List({'pair_id': Value('int64'), 'family_a': Value('string'), 'family_b': Value('string'), 'share_k': Value('int64'), 'both_active_on_shared': Value('bool'), 'correct_dominates_after_one': Value('bool'), 'ambiguity_retained_on_missing': Value('bool'), 'active_on_shared': Value('int64'), 'families_on_shared': Value('int64')})}, 'n_disjoint_families': Value('int64'), 'n_polysemy_pairs': Value('int64'), 'family_results_typed': List({'family': Value('string'), 'variant': Value('string'), 'walk_len': Value('int64'), 'full_dominant': Value('string'), 'inclusion': Value('bool'), 'bundle_size_full': Value('int64'), 'bundle_reduction': Value('float64'), 'stability_drop_all': Value('float64'), 'stability_drop_interior': Value('float64'), 'stability_corrupt': Value('float64'), 'basin_depth_drop': Value('int64'), 'per_step_drop': List({'dropped_step': Value('int64'), 'dominant_after': Value('string'), 'n_active_after': Value('int64'), 'n_families_after': Value('int64'), 'survived': Value('bool')})})}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 478, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2818, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2355, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2380, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2369, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2297, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
changed: list<item: string>
child 0, item: string
commands: list<item: string>
child 0, item: string
metrics: struct<typed: struct<n_instances: int64, inclusion_rate: double, bundle_reduction_gmean: double, sta (... 418 chars omitted)
child 0, typed: struct<n_instances: int64, inclusion_rate: double, bundle_reduction_gmean: double, stability_drop_al (... 145 chars omitted)
child 0, n_instances: int64
child 1, inclusion_rate: double
child 2, bundle_reduction_gmean: double
child 3, stability_drop_all_mean: double
child 4, stability_drop_interior_mean: double
child 5, stability_corrupt_mean: double
child 6, basin_depth_drop_mean: double
child 7, basin_depth_drop_max: int64
child 1, labelfree: struct<n_instances: int64, inclusion_rate: double, bundle_reduction_gmean: double, stability_drop_al (... 145 chars omitted)
child 0, n_instances: int64
child 1, inclusion_rate: double
child 2, bundle_reduction_gmean: double
child 3, stability_drop_all_mean: double
child 4, stability_drop_interior_mean: double
child 5, stability_corrupt_mean: double
child 6, basin_depth_drop_mean: double
child 7, basin_depth_drop_max: int64
artifacts: list<item: string>
child 0, item: string
relaxation_rule: string
n_polysemy_pairs: int64
aggregate: struct<typed: struct<n_instances: int64, inclusion_rate: double, bundle_reduction_gmean: double, sta (... 418 chars omitted)
child 0, typed: struct<n_inst
...
list<item: struct<pair_id: int64, family_a: string, family_b: string, share_k: int64, (... 406 chars omitted)
child 0, typed: list<item: struct<pair_id: int64, family_a: string, family_b: string, share_k: int64, both_active_on (... 139 chars omitted)
child 0, item: struct<pair_id: int64, family_a: string, family_b: string, share_k: int64, both_active_on_shared: bo (... 127 chars omitted)
child 0, pair_id: int64
child 1, family_a: string
child 2, family_b: string
child 3, share_k: int64
child 4, both_active_on_shared: bool
child 5, correct_dominates_after_one: bool
child 6, ambiguity_retained_on_missing: bool
child 7, active_on_shared: int64
child 8, families_on_shared: int64
child 1, labelfree: list<item: struct<pair_id: int64, family_a: string, family_b: string, share_k: int64, both_active_on (... 139 chars omitted)
child 0, item: struct<pair_id: int64, family_a: string, family_b: string, share_k: int64, both_active_on_shared: bo (... 127 chars omitted)
child 0, pair_id: int64
child 1, family_a: string
child 2, family_b: string
child 3, share_k: int64
child 4, both_active_on_shared: bool
child 5, correct_dominates_after_one: bool
child 6, ambiguity_retained_on_missing: bool
child 7, active_on_shared: int64
child 8, families_on_shared: int64
branch: string
n_disjoint_families: int64
to
{'experiment': Value('string'), 'branch': Value('string'), 'relaxation_rule': Value('string'), 'aggregate': {'typed': {'n_instances': Value('int64'), 'inclusion_rate': Value('float64'), 'bundle_reduction_gmean': Value('float64'), 'stability_drop_all_mean': Value('float64'), 'stability_drop_interior_mean': Value('float64'), 'stability_corrupt_mean': Value('float64'), 'basin_depth_drop_mean': Value('float64'), 'basin_depth_drop_max': Value('int64')}, 'labelfree': {'n_instances': Value('int64'), 'inclusion_rate': Value('float64'), 'bundle_reduction_gmean': Value('float64'), 'stability_drop_all_mean': Value('float64'), 'stability_drop_interior_mean': Value('float64'), 'stability_corrupt_mean': Value('float64'), 'basin_depth_drop_mean': Value('float64'), 'basin_depth_drop_max': Value('int64')}}, 'polysemy': {'typed': List({'pair_id': Value('int64'), 'family_a': Value('string'), 'family_b': Value('string'), 'share_k': Value('int64'), 'both_active_on_shared': Value('bool'), 'correct_dominates_after_one': Value('bool'), 'ambiguity_retained_on_missing': Value('bool'), 'active_on_shared': Value('int64'), 'families_on_shared': Value('int64')}), 'labelfree': List({'pair_id': Value('int64'), 'family_a': Value('string'), 'family_b': Value('string'), 'share_k': Value('int64'), 'both_active_on_shared': Value('bool'), 'correct_dominates_after_one': Value('bool'), 'ambiguity_retained_on_missing': Value('bool'), 'active_on_shared': Value('int64'), 'families_on_shared': Value('int64')})}, 'n_disjoint_families': Value('int64'), 'n_polysemy_pairs': Value('int64'), 'family_results_typed': List({'family': Value('string'), 'variant': Value('string'), 'walk_len': Value('int64'), 'full_dominant': Value('string'), 'inclusion': Value('bool'), 'bundle_size_full': Value('int64'), 'bundle_reduction': Value('float64'), 'stability_drop_all': Value('float64'), 'stability_drop_interior': Value('float64'), 'stability_corrupt': Value('float64'), 'basin_depth_drop': Value('int64'), 'per_step_drop': List({'dropped_step': Value('int64'), 'dominant_after': Value('string'), 'n_active_after': Value('int64'), 'n_families_after': Value('int64'), 'survived': Value('bool')})})}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Basin Retrieval: Structural Compatibility as a First-Stage Search-Space Compression Operator
This repository contains the code, data, and results for the paper Basin Retrieval: Structural Compatibility as a First-Stage Search-Space Compression Operator.
The paper makes one claim: a semantically indifferent structural retrieval operator can compress a relational search space into a bounded basin of compatible candidates while preserving the target, leaving identification to a downstream stage. On held-out relational data it achieves 34× compression with perfect target inclusion (1.0), on project-history content 32×, and on temporal traces up to 554×, again with perfect inclusion.
This is the first of three planned papers. It presents only the compression
operator and its envelope. Two companion papers — Semantic Fragmentation in
Structural Retrieval (why semantic labels hurt first-stage retrieval) and
Emergent Evidence Graphs from Structural Basin Retrieval (why a retrieved
basin naturally induces a reasoning graph) — live as drafts in episteme/paper/
and draw on the same archived code and data in this repository.
Repository layout
basin-retrieval/
├── paper/
│ └── basin-retrieval.md the paper
├── code/
│ ├── core/ the encoding + retrieval operator
│ │ ├── signature.py canonical first-occurrence recurrence signatures
│ │ ├── graph_dataset.py typed relational graph data model
│ │ ├── generator.py synthetic relational-family generator
│ │ ├── relaxation.py prefix-consistency basin retrieval
│ │ ├── matcher_relaxation.py DP/LCS alignment
│ │ └── identity_regimes.py identity regime comparisons
│ ├── basin_retrieval/ → 34× compression result (self-contained)
│ ├── behavioral_relevance/ → 32× compression on LDGR history
│ ├── payload_graph/ → 0.80 content-graph vs node-bag control
│ └── topology/ → 23.9× / 554× / granularity-sweep results
├── data/
│ ├── ldgr_history/ extracted LDGR observations + artifacts (JSONL)
│ └── ldgr_benchmarks/ LDGR event-log dbs (for topology experiments)
├── results/ result JSONs (primary evidence)
│ ├── topology/ 554× / granularity-sweep evidence of record
│ └── dual_lookup/ 0.567 naive-union evidence of record
└── reports/ human-readable analysis
How LDGR was used in this research
LDGR is a minimal durable investigation loop backed by SQLite — a tool our team uses to record research as permanent event logs. In this paper it played two distinct and deliberate roles, and both are worth stating because they are what keep the 32× and 554× results grounded in real rather than synthetic content.
1. The project's own history as memory content (data/ldgr_history/)
The research that produced this paper was itself run through LDGR. Every observation (hypothesis, finding), artifact (code file, report, result JSON), and decision (pivot, validation, stop) was recorded as it happened. We then used that self-recorded history as the memory content for the 32× behavioral result and the 0.80 content-graph control.
This is deliberate and a little recursive: the substrate retrieves the project's own recorded findings. It tests the operator on natural relational content — real observations, real artifact descriptions, real report chunks — rather than on generated families. The corpus is 32 items stratified by topic (deletion, matcher, identity, noise, polysemy, typed, phase 0), drawn from 17 observations and 40 artifacts.
The data/ldgr_history/ directory is a complete, field-faithful extraction of
the source LDGR database into JSONL — 17 observations and 40 artifacts with
zero field mismatches against the source. The behavioral experiment reads
these directly; no database or LDGR installation is required to reproduce the
32× result.
2. LDGR event logs as a temporal-trace corpus (data/ldgr_benchmarks/)
For the temporal-trace results (23.9× on synthetic motifs, 554× on long-window
workflow logs), we used LDGR event logs from many separate LDGR-tracked
software-development runs. LDGR records every workflow event as a row
(entity_type:event_type over observation, artifact, decision, run).
We do not feed raw text into the operator. We tokenize each event into a
coarse, content-safe category tag — for example observation:add:failure,
artifact:add:report, decision:record:continue, run:end:pass. The
coarsening is rule-based: observations are categorized by keywords in their
body (failure / constraint / result / implementation / data / note), artifacts
by their path (validator / report / result / implementation / patch), decisions
by their rationale (stop / pivot / blocker / validated / completed), runs by
status (pass / fail / partial). These tagged events are then windowed into
state-transition sequences and indexed by the same canonical recurrence
signature used for the relational graph.
This tokenization choice is itself part of the paper's thesis. The finest categorical detail (the raw text of each observation) would fragment the recurrence basins; coarse category tags preserve recurrence density. That is the granularity sweep result stated the other way around: coarse tokens gave best reuse (0.948), full categorical detail the worst (0.691).
data/ldgr_benchmarks/ bundles 28 of these event-log databases. The full
128-db corpus is available on Hugging Face.
Why both roles matter
Together the two roles cover the two regimes the paper needs to defend: the behavioral result shows the operator compresses natural relational content (real project history); the temporal result shows it scales on genuinely sequential, high-volume data. Neither uses synthetic families alone. LDGR is what lets us claim the compression envelope holds on real workloads rather than only on the generated corpus.
Quick reproduction
Python 3.12, no external dependencies beyond the standard library.
# 34× compression on synthetic relational families (self-contained, ~5s)
python3 code/basin_retrieval/bench_relaxation.py
# 32× compression on LDGR project history (self-contained via JSONL snapshot, ~10s)
python3 code/behavioral_relevance/bench_behavioral_relevance.py
# content-graph vs node-bag on rewired decoys (self-contained, ~20s)
python3 code/payload_graph/bench_payload_graph_refinement.py
# 23.9× synthetic motif compression (self-contained)
python3 code/topology/set_valued_prediction_experiment.py
Seed for all experiments: 20260706.
See REPRODUCING.md for the full claim-to-evidence map, including which
experiments are fully self-contained and which require the LDGR benchmark
corpus.
What this is not
This is a compression operator, not a semantic retrieval system, reasoning engine, or memory architecture. It does not rank, does not interpret content, and does not claim to be a complete retrieval solution. It compresses the candidate space so a downstream stage can afford to process what survives.
License
MIT.
- Downloads last month
- 48