Dataset Viewer
Auto-converted to Parquet Duplicate
text
stringlengths
0
166
f979705293849de48e4b40aa8988c046fa5871da
8fdc369223509baee3422914efb3053adbd27b6391073644599a3a9ca1519cb3 experiments/LDE2G_PREREG.md
b02983dbd22674c5b234aaada26e5003133f6a70464859e831267089c24d0137 experiments/lde2g_generator.py
496e31ab7740c1f6264b94811f2d49b3a1f359a22be7df4ada78507620518638 experiments/lde2g_information_priority_confirmatory.py
0d37b5c676c7fa1f6248db6fc62b82bf64a98d60c7e0025bd387a0f63c834367 lde/swarm/protocol.py
d1e3c827a1811094ba93cf03ca9d26a1155a7f7f9883cf33608fe88062f43c2a lde/swarm/priority.py
5f2ba5f1c60718dce95bdefc86b3e77a1799934843f535388a89b58321f3524f lde/swarm/selection.py
cb004d9136656f1843dcf22b9b9a4fc5bd612962de04160781ad9c4c1e3d12f8 lde/swarm/discrimination.py
fa72322e0152f269abc67929ee17abe713206ebc0bfd07588d9d5912f32c2c38 tests/test_lde2g_confirmatory.py
=== LDE-2G S25 PRE-EXPERIMENT ENVIRONMENT ===
Captured UTC:
2026-10-07T16:05:38Z
=== REPOSITORY ===
commit: f979705293849de48e4b40aa8988c046fa5871da
working_tree_status:
=== PYTHON ===
Python 3.14.6
machine: aarch64
platform: Android-16-aarch64-64bit-ELF
implementation: CPython
python: 3.14.6 (main, Jul 5 2026, 10:35:55) [Clang 21.0.0 (https://android.googlesource.com/toolchain/llvm-project 5e96669f0
=== BUILD ===
compiler: aarch64-linux-android-clang
SOABI: cpython-314-aarch64-linux-android
=== HOST ===
Linux localhost 6.6.98-android15-8-pd6ff1cd-abogkiS938USQSCCZF9-4k #1 SMP PREEMPT Tue Aug 11 06:32:03 UTC 2026 aarch64 Android
arch: aarch64
=== ANDROID ===
model: SM-S938U
abi: arm64-v8a
android: 16
build: S938USQSCCZF9
=== STORAGE ===
Filesystem Size Used Avail Use% Mounted on
/dev/block/dm-72 459G 197G 261G 44% /data/user/0
=== LDE-2G SOURCE HASHES ===
8fdc369223509baee3422914efb3053adbd27b6391073644599a3a9ca1519cb3 experiments/LDE2G_PREREG.md
b02983dbd22674c5b234aaada26e5003133f6a70464859e831267089c24d0137 experiments/lde2g_generator.py
496e31ab7740c1f6264b94811f2d49b3a1f359a22be7df4ada78507620518638 experiments/lde2g_information_priority_confirmatory.py
0d37b5c676c7fa1f6248db6fc62b82bf64a98d60c7e0025bd387a0f63c834367 lde/swarm/protocol.py
d1e3c827a1811094ba93cf03ca9d26a1155a7f7f9883cf33608fe88062f43c2a lde/swarm/priority.py
5f2ba5f1c60718dce95bdefc86b3e77a1799934843f535388a89b58321f3524f lde/swarm/selection.py
cb004d9136656f1843dcf22b9b9a4fc5bd612962de04160781ad9c4c1e3d12f8 lde/swarm/discrimination.py
fa72322e0152f269abc67929ee17abe713206ebc0bfd07588d9d5912f32c2c38 tests/test_lde2g_confirmatory.py
e5a9d5e372e69061e2db8d9a820121637a67d61fd800bc5f8403586cab983c39 evidence/05_exploratory/generator_independence/azure/README.md
0569445b9d77d7cc21cb9d4d67499f0d5782382f2cf8a1d6b0294c54fbf03989 evidence/05_exploratory/generator_independence/azure/command.txt
95374d9ce31783a1756f5e196fe9fe90745627bba4725aeb157dbf6a0fb793a3 evidence/05_exploratory/generator_independence/azure/environment.txt
4a8ba5ad855c89dcd2746eb2e4e4cd2ce243fb2e88fcc85dedb4cec2dd4bb33f evidence/05_exploratory/generator_independence/azure/summary.json
e42987d1cf52ac827ef5039c902fb9f87a2e51371c09c2de5bbcd6b019f338d9 evidence/05_exploratory/generator_independence/azure/stdout.txt
e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855 evidence/05_exploratory/generator_independence/azure/stderr.txt
9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa evidence/05_exploratory/generator_independence/azure/exit_code.txt
PYTHONPATH="$PWD" python3 - <<'PY'
from experiments.lde2g_generator import generate_case
from experiments.exploratory.lde2g_generator_independence import generate_case_independent
from lde.swarm.blackboard import ResearchBlackboard
from lde.swarm.discrimination import candidate_discrimination
from lde.swarm.priority import priority_score
from lde.swarm.selection import select_next_experiment
def evaluate(case):
board = ResearchBlackboard()
for hypothesis in case.hypotheses:
board.add_hypothesis(hypothesis)
for request in case.requests:
board.add_experiment_request(request)
selected = select_next_experiment(board)
if selected is None:
raise AssertionError("case produced no selectable request")
hypothesis_records = tuple(
{
"hypothesis_id": h.hypothesis_id,
"author": h.author,
"statement": h.statement,
"predictions": h.predictions,
}
for h in case.hypotheses
)
scores = {
request.request_id: candidate_discrimination(
hypothesis_records,
request.input_value,
)
for request in case.requests
}
End of preview. Expand in Data Studio

LDE-2G cross-substrate research record

Raw records of one preregistered experiment, LDE-2G, run on the Local Discovery Engine at source commit f979705293849de48e4b40aa8988c046fa5871da (codeberg.org/badatchess/local-discovery-engine). This card was added on 2026-11-07. It describes the files as they are; no data, protocol or historical output was changed.

Result first: the primary hypothesis was not supported

The preregistered hypothesis (01_preregistration/LDE2G_PREREG.md) was that the engine's existing priority selector picks the candidate experiment with maximal measured discrimination in every trial. It did not.

From the root result_summary.json (Azure confirmatory run):

Field Value
cases 60 (seeds 0–59)
selected_maximal 49
non_maximal 11
selection_rate 0.816667
mean_selected_discrimination 3.8
mean_maximum_discrimination 4.0
primary_hypothesis_supported false

The preregistration's falsification rule is that any trial selecting a strictly lower-discrimination candidate refutes the hypothesis; 11 of 60 did.

What "discrimination" means here. For a candidate experiment, it is the number of distinct outputs predicted by the competing hypotheses (a deterministic count). The preregistration says explicitly that it is not Shannon information gain, and that agreement or disagreement with it says nothing about AI intelligence, reasoning, optimal experiment design or generalization.

What did reproduce: the same output on two machines

releases/2026-10-11/lde-cross-substrate-evidence/04_confirmatory/ records the same confirmatory run on:

  • Azure: x86_64, Ubuntu Server 24.04.4 LTS, CPython 3.12.3;
  • a Samsung Galaxy S25 Ultra: aarch64, Android 16, CPython 3.14.6.

Both gave 49/60 (0.816667). cross_substrate_comparison.json records byte-identical stdout (sha256 966108ec…934b), identical empty stderr and exit code 0, and 0 per-seed differences: EXACT_MATCH_FOR_RECORDED_CONFIRMATORY_OUTPUT.

This shows that the recorded behaviour reproduced on these two machines. It does not show portability to other platforms, independent replication by another party, or that the hypothesis is true. The same failed result reproduced.

Preserved separately: an invalid attempt

The root stderr.txt and exit_code.txt (1) are from a first, invalid attempt that failed before any experiment ran (ModuleNotFoundError: No module named 'lde'), as stated in failed_attempt_status.txt. The confirmatory run's own files are confirmatory_stdout.txt, confirmatory_stderr.txt (empty) and confirmatory_exit_code.txt (0).

Exploratory, not preregistered

05_exploratory/generator_independence/ changes how the case generator draws random numbers (one stream vs split streams). With split streams, 46/60 were maximal (0.766667), against 49/60 for the original. Its own summary says this changed the observed behaviour in one 60-seed run and does not establish causality or statistical independence. It is not part of the confirmatory result.

Layout

Path Contents
01_preregistration/ the original LDE-2G preregistration
02_source/ source commit and source-file hashes
03_substrates/s25/ the S25's environment fingerprint
05_exploratory/ the generator-independence perturbation (exploratory)
06_historical_lineage/2026-10-07/ earlier LDE baselines and experiments (LDE0–LDE2F)
07_azure_archive/2026-10-07/ the Azure archive: preregistration, Amendments 1 and 2, source, environment
99_provenance/ the pre-confirmatory manifest (recorded with the outcome not yet known), timeline and hashes
releases/2026-10-11/ the cross-substrate release bundle
education/ a research curriculum document (not evidence)
root files the Azure confirmatory outputs, plus the invalid attempt described above

Amendment 1 reclassifies the first implementation as a pilot / nondiagnostic and keeps it. Amendment 2 freezes the confirmatory procedure (seeds 0–59, four hypotheses and six candidates per case). The original preregistration is kept unchanged.

Limits

  • Synthetic, generator-dependent workload; results are conditional on that generator.
  • Two computational substrates, both run by the same owner. No independent replication.
  • No licence has been declared for this dataset. Ask the owner before reuse.

Provenance

Dataset owner: Chad Holland (Hugging Face: holland202). The files listed above do not record which AI systems, if any, helped design or run the experiment, so this card does not say. This card was drafted by Claude (Anthropic, Opus 5.5) from the files listed above, on 2026-11-06 → human validation: not performed → human responsibility: Chad Holland.

Downloads last month
126