Hugging Face's logo

Datasets:
Bakobiibizo
/
episteme-evidence

Dataset card Files
xet
 Community
Dataset Viewer
Duplicate
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
baseline_mean_expected_concept_recall_without_trajectory: double
baseline_mean_expected_top_concept_recall_without_trajectory: double
boundary: string
mean_expected_concept_recall: double
mean_expected_top_concept_recall: double
mean_semantic_lift_from_trajectory_constraints: double
mean_top_concept_lift_from_trajectory_constraints: double
mean_top_margin_lift_from_trajectory_constraints: double
mean_top_score_lift_from_trajectory_constraints: double
mean_trajectory_evidence_count: int64
query_count: int64
query_records: string
schema_version: string
semantic_records: string
trajectory_motif_accuracy: double
trajectory_records: string
composition_records: string
context_resolution_accuracy: int64
additive_transition_counts: list<item: null>
  child 0, item: null
mean_first_additive_transition_competitor_count: null
additive_ambiguity_counts: list<item: int64>
  child 0, item: int64
continuous_refinement_records: string
ambiguous_surface_multi_basin_rate: int64
tie_policy_mean_expected_utility_on_ties: double
mean_first_additive_ambiguity_competitor_count: int64
mean_first_replacement_transition_competitor_count: int64
supports_projection_hypothesis: bool
perturbation_records: string
composition_success_rate: int64
tie_policy_lift_vs_forced_choice_on_ties: double
ambiguous_surface_count: int64
replacement_transition_counts: list<item: int64>
  child 0, item: int64
forced_choice_mean_expected_utility_on_ties: double
context_resolution_records: string
replacement_ambiguity_counts: list<item: int64>
  child 0, item: int64
ambiguity_records: string
continuous_refinement_final_accuracy: int64
threshold: double
mean_first_replacement_ambiguity_competitor_count: int64
to
{'additive_ambiguity_counts': List(Value('int64')), 'additive_transition_counts': List(Value('null')), 'ambiguity_records': Value('string'), 'ambiguous_surface_count': Value('int64'), 'ambiguous_surface_multi_basin_rate': Value('int64'), 'boundary': Value('string'), 'composition_records': Value('string'), 'composition_success_rate': Value('int64'), 'context_resolution_accuracy': Value('int64'), 'context_resolution_records': Value('string'), 'continuous_refinement_final_accuracy': Value('int64'), 'continuous_refinement_records': Value('string'), 'forced_choice_mean_expected_utility_on_ties': Value('float64'), 'mean_first_additive_ambiguity_competitor_count': Value('int64'), 'mean_first_additive_transition_competitor_count': Value('null'), 'mean_first_replacement_ambiguity_competitor_count': Value('int64'), 'mean_first_replacement_transition_competitor_count': Value('int64'), 'perturbation_records': Value('string'), 'replacement_ambiguity_counts': List(Value('int64')), 'replacement_transition_counts': List(Value('int64')), 'schema_version': Value('string'), 'supports_projection_hypothesis': Value('bool'), 'threshold': Value('float64'), 'tie_policy_lift_vs_forced_choice_on_ties': Value('float64'), 'tie_policy_mean_expected_utility_on_ties': Value('float64')}
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
              baseline_mean_expected_concept_recall_without_trajectory: double
              baseline_mean_expected_top_concept_recall_without_trajectory: double
              boundary: string
              mean_expected_concept_recall: double
              mean_expected_top_concept_recall: double
              mean_semantic_lift_from_trajectory_constraints: double
              mean_top_concept_lift_from_trajectory_constraints: double
              mean_top_margin_lift_from_trajectory_constraints: double
              mean_top_score_lift_from_trajectory_constraints: double
              mean_trajectory_evidence_count: int64
              query_count: int64
              query_records: string
              schema_version: string
              semantic_records: string
              trajectory_motif_accuracy: double
              trajectory_records: string
              composition_records: string
              context_resolution_accuracy: int64
              additive_transition_counts: list<item: null>
                child 0, item: null
              mean_first_additive_transition_competitor_count: null
              additive_ambiguity_counts: list<item: int64>
                child 0, item: int64
              continuous_refinement_records: string
              ambiguous_surface_multi_basin_rate: int64
              tie_policy_mean_expected_utility_on_ties: double
              mean_first_additive_ambiguity_competitor_count: int64
              mean_first_replacement_transition_competitor_count: int64
              supports_projection_hypothesis: bool
              perturbation_records: string
              composition_success_rate: int64
              tie_policy_lift_vs_forced_choice_on_ties: double
              ambiguous_surface_count: int64
              replacement_transition_counts: list<item: int64>
                child 0, item: int64
              forced_choice_mean_expected_utility_on_ties: double
              context_resolution_records: string
              replacement_ambiguity_counts: list<item: int64>
                child 0, item: int64
              ambiguity_records: string
              continuous_refinement_final_accuracy: int64
              threshold: double
              mean_first_replacement_ambiguity_competitor_count: int64
              to
              {'additive_ambiguity_counts': List(Value('int64')), 'additive_transition_counts': List(Value('null')), 'ambiguity_records': Value('string'), 'ambiguous_surface_count': Value('int64'), 'ambiguous_surface_multi_basin_rate': Value('int64'), 'boundary': Value('string'), 'composition_records': Value('string'), 'composition_success_rate': Value('int64'), 'context_resolution_accuracy': Value('int64'), 'context_resolution_records': Value('string'), 'continuous_refinement_final_accuracy': Value('int64'), 'continuous_refinement_records': Value('string'), 'forced_choice_mean_expected_utility_on_ties': Value('float64'), 'mean_first_additive_ambiguity_competitor_count': Value('int64'), 'mean_first_additive_transition_competitor_count': Value('null'), 'mean_first_replacement_ambiguity_competitor_count': Value('int64'), 'mean_first_replacement_transition_competitor_count': Value('int64'), 'perturbation_records': Value('string'), 'replacement_ambiguity_counts': List(Value('int64')), 'replacement_transition_counts': List(Value('int64')), 'schema_version': Value('string'), 'supports_projection_hypothesis': Value('bool'), 'threshold': Value('float64'), 'tie_policy_lift_vs_forced_choice_on_ties': Value('float64'), 'tie_policy_mean_expected_utility_on_ties': Value('float64')}
              because column names don't match

Need 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.

Evidence Collection — Episteme Recurrence-Sensitive Memory Substrate

This directory collects all evidence referenced by the episteme paper. Every claim in final_synthesis.md points to specific files here. Provenance is preserved: each file is labeled by which project produced it.

Provenance Map 

episteme/      — this project (the paper's subject)
  code/          all experiment implementations (reproducible)
  results/       raw JSON outputs (primary evidence)
  reports/       human-readable analysis derived from results
  ldgr_history/  LDGR project observations/artifacts used as memory content
  phase0/        the phase-0 toy semantic retrieval benchmark
topology/      — parent project (identity-free recurrent topology memory)
  reports/       findings and meta-analysis
  source/        canonical experiment implementations (the ports episteme inherits)
ecphory-2/     — parent project (trajectory vs semantic memory split)
  docs/          architecture decisions and experimental reports
  results/       raw JSON including the run-46 ceiling experiments

Claim → Evidence Map

Each numbered claim lists: the claim, the supporting files, and the headline metric.

Lineage claims

C0. Identity-free recurrent topology is a valid constraint memory.

  • evidence: topology/reports/meta_analysis.md, topology/reports/final_findings.md
  • code: topology/source/recurrence_topology_experiment.py, topology/source/set_valued_prediction_experiment.py
  • metric: set-valued retrieval N=3, coverage 0.88, inclusion 1.0, compression 23.9x

C1. Coarse-to-fine two-stage filtering is the production pattern.

  • evidence: topology/reports/two_stage_granularity_experiment.md
  • code: topology/source/two_stage_granularity_experiment.py
  • metric: entity→command_refined, utility 0.689, mean final set 2.85 vs single-stage coarse 3.36

C2. The typed-edge question was left open (topology phase 4).

  • evidence: topology/reports/meta_analysis.md (section "Methodological Guardrails": topology deliberately strips relation types)

C3. Trajectory and semantic memory are distinct retrieval problems; the graph is a projection of basins.

  • evidence: ecphory-2/docs/trajectory-vs-semantic-memory-architecture.md

C4. Unlabeled constraint basins recover human labels above shuffled chance.

  • evidence: ecphory-2/results/unlabeled-constraint-basin-pilot.json, ecphory-2/docs/unlabeled-constraint-basin-pilot.md
  • metric: query accuracy 1.0 vs shuffled 0.0; basin purity 1.0 at threshold 0.40

C5. Agreement-gated dual lookup beats naive union.

  • evidence: ecphory-2/results/promoted-dual-lookup-control.json, ecphory-2/docs/promoted-dual-lookup-control-experiment.md
  • metric: agreement-gated 1.0 vs naive union 0.567 (worse than labeled-only 0.700)

C6. Run-46 ceiling: synthetic relaxation could not beat a label-overlap index.

  • evidence: ecphory-2/results/scale-local-dynamics-vs-index-retrieval.json, ecphory-2/results/scale-trajectory-to-semantic-basin-bridge.json
  • metric: any_broad_metric_win_for_relax: False; relaxation's only unique act is reconstructive completion, not retrieval advantage

Episteme core claims

C7. Typed edges carry discriminating information beyond pure topology.

  • evidence: episteme/results/basin_results.json, episteme/reports/findings.md
  • code: episteme/code/signature.py (typed_canonical_signature vs label_free), episteme/code/bench_relaxation.py
  • metric: polysemy disambiguation typed 1.0 vs labelfree 0.25; corrupt stability typed 0.69 vs 0.29
  • this answers C2 (topology's open phase-4 question)

C8. Set-valued basin retrieval generalizes to typed content.

  • evidence: episteme/results/basin_results.json, episteme/results/bench_twostage (two-stage report in episteme/reports/findings.md)
  • code: episteme/code/relaxation.py, episteme/code/bench_twostage.py
  • metric: held-out inclusion 1.0, bundle reduction 36x (synthetic)

C9. Noise collapse was a matcher failure, not a representation failure.

  • evidence (diagnosis correction): episteme/code/smoke_identity.py, episteme/results/ (no separate JSON; see episteme/reports/findings.md "Later Correction" section)
  • evidence (ablation): episteme/results/matcher_ablation_results.json
  • code: episteme/code/identity_regimes.py, episteme/code/matcher_relaxation.py, episteme/code/bench_matcher_ablation.py
  • metric: identity smoke showed canonical stable to insertion 9/9, role_payload(local) worse (0/9, 4/9); greedy noise@1 0.60 → DP/LCS 0.75

C10. DP/LCS skip-capable alignment repairs neutral-noise collapse without touching identity.

  • evidence: episteme/results/dp_lcs_surface_results.json, episteme/reports/dp_lcs_surface_report.md
  • code: episteme/code/matcher_relaxation.py (align_dp), episteme/code/bench_dp_lcs_surface.py
  • metric: noise transition greedy mag 2 → DP/LCS mag 3; polysemy 1.0 unchanged; reduction 34x unchanged

C11. Deletion brittleness is a narrow identity boundary (first-occurrence renumbering).

  • evidence: episteme/results/deletion_controls_results.json, episteme/reports/deletion_controls_report.md
  • code: episteme/code/bench_deletion_controls.py
  • metric: canonical first-recurring deletion 0.5, repeat deletion 1.0; anchored_recurring REJECTED (polysemy 1.0→0.43); edge_sequence/bag robust but lose polysemy (0.86/0.57)

C12. Behavioral relevance: the deletion boundary is often tolerable on real content.

  • evidence: episteme/results/behavioral_relevance_results.json, episteme/reports/behavioral_relevance_report.md
  • code: episteme/code/bench_behavioral_relevance.py
  • memory content: episteme/ldgr_history/observations.jsonl, episteme/ldgr_history/artifacts.jsonl
  • metric: clean 1.0/32x; plausible survival 0.8884 (0.8644 with first-rec deletion); control/adversarial 0.4479

C13. Payload-graph refinement beats flat overlap where flat overlap should fail.

  • evidence: episteme/results/payload_graph_refinement_results.json, episteme/reports/payload_graph_refinement_report.md
  • code: episteme/code/bench_payload_graph_refinement.py
  • metric: same-node rewired decoys — node_bag top1 0.0 (decoys tie), content_graph top1 0.7969 aggregate / 1.0 on core+partial+noisy
  • this continues C6 (ecphory-2's run-46 escape generalizes)

Reproduction

All episteme experiments run from experiments/motif-topology-retrieval/:

# smokes (fast, verify mechanisms)
python3 smoke.py
python3 smoke_generator.py
python3 smoke_relaxation.py
python3 smoke_identity.py        # C9 diagnosis correction
python3 smoke_matcher.py         # C9 matcher mechanism

# benchmarks (produce the result JSONs)
python3 bench_relaxation.py              # C7, C8
python3 bench_twostage.py                # C8 two-stage
python3 bench_noise_taxonomy.py          # C9 precursor
python3 bench_matcher_ablation.py        # C9 ablation
python3 bench_dp_lcs_surface.py          # C10
python3 bench_deletion_controls.py       # C11
python3 bench_behavioral_relevance.py    # C12
python3 bench_payload_graph_refinement.py # C13

Seed for all experiments: 20260706. Corpus: build_generated_graph_with_polysemy(n_disjoint_families=20, n_polysemy_bases=10, instances_per=3) for synthetic; LDGR project history (17 observations, 40 artifacts) for behavioral/payload-graph.

File Inventory 

episteme/
  code/        19 Python files (signature, relaxation, matchers, 8 benchmarks, smokes, generator)
  results/     11 JSON result files (primary evidence)
  reports/     9 markdown reports + program doc
  ldgr_history/  observations.jsonl, artifacts.jsonl, manifest.json
  phase0/      10 files (toy benchmark)

topology/
  reports/     12 markdown reports + design principles
  source/      7 Python files (the ports episteme inherits)

ecphory-2/
  docs/        8 architecture/experiment docs
  results/     7 JSON reports + trajectory-memory-lookup multi-stage results

Frozen State 

GitHub:   https://github.com/hydra-dynamix/episteme
Hugging Face: https://huggingface.co/datasets/Bakobiibizo/episteme-evidence
Downloads last month
105
Total file size:
4.21 MB