atmem/code-pruner-click-evidence-20261020

Click: native, Cognee, AtMem, and GPU source pruning

Six recorded coding attempts on one historical Click issue, b67832c. This dataset makes the successful and failed attempts inspectable: measurements, tool calls, model CLI streams, proposed source changes, test output, protocols, GPU qualification and original encrypted archives. It is evaluation evidence, not memory-model training data or a SWE-bench leaderboard submission.

Results at a glance

Optional reader: four configurations, one attempt each

Configuration	Task passed	Selected tests passed	Input tokens including cache	Agent seconds
Native	Yes	1,868	76,994	72.275
Cognee	Yes	1,868	57,724	49.544
AtMem curated	Yes	1,868	56,795	49.964
AtMem, pruning enabled	Yes	1,868	53,864	69.841

No arm called the graph. The pruning-enabled agent did not invoke the reader. The ranking therefore does not demonstrate graph or neural-pruning savings.

Controlled reader: two separate new attempts

Configuration	Task passed	Selected tests passed	Input tokens including cache	Agent seconds
AtMem full source	No	1,867/1,868	75,178	57.365
AtMem + GPU SWE-Pruner	Yes	1,868/1,868	88,552	76.306

Both agents had to read source before workspace commands were enabled. Pruning actually ran twice: 3,456 source bytes became 1,174 (66.0% reduction), in 1.155 GPU seconds combined. The agent subsequently reread 3,002 unpruned bytes through the shell. Total input increased 17.8% and agent time 33.0%. Source compression did not become end-to-end token savings. The full-source patch missed a custom help-option collision case. One pair cannot establish a causal quality benefit.

Browse and use

The default dataset viewer shows all six trial summaries. Select events for decoded, privacy-redacted event payloads. Nested payloads use JSON strings to keep heterogeneous events readable without forcing a misleading common schema.

from datasets import load_dataset

trials = load_dataset("atmem/code-pruner-click-evidence-20261020", "trials", split="test")
events = load_dataset("atmem/code-pruner-click-evidence-20261020", "events", split="test")
print(trials.to_pandas())

Use an immutable dataset revision when reproducing an analysis; the publication receipt in the AtMem repository records the uploaded revision and checksums.

Directory/file	Contents
data/trials.jsonl	Six summary rows; tokens, tests, latency and actual tool usage
data/events.jsonl	Readable event payloads, prompts and complete attempt/grade records
trials/	Per-attempt source changes, decoded CLI stream, tool audit and grader output
protocols/	Separate fixed four-way and controlled-reader protocols
gpu/	Seven recovered runtime/qualification/audit/source artifacts
harness/	Privacy-redacted source snapshots retained by the original archive
reports/	Detailed interpretation, machine-readable results and rental-cost record
archives/	Byte-identical signed, recipient-encrypted originals; keys remain with the owner
manifest.json	SHA256 and byte size of every other publication file
verification.json	Decryption verification, original event digests and redaction counts

Readable exports replace credentials, local home paths and IP addresses. They are derivatives, not the original signed bytes. Encrypted originals provide owner-recoverable preservation; public readers can analyze the readable files without private keys. Environment files, SSH keys, archive secret keys, runtime databases and owner configuration are not uploaded. No solver was rerun for publication. No new labels or conversation examples were generated.

Experimental scope

Claude Opus 5.5 medium, one fixed-order attempt per arm, one previously inspected public task. Cached tokens are included in input; GPU skimmer tokens are not Claude tokens. The two protocols must not be pooled. Frozen grading also skipped 25 tests, deselected 31,001 and xfailed one. No experience memory was used. These are workflow diagnostics, not an untouched confirmation set or evidence of general superiority. Full method/runtime details are in reports/.

RTX A4000 16 GB; torch 2.8.0+cu128; transformers 4.57.1; flash-attn 2.8.3.post2; swe-pruner 0.1.1; unmodified upstream wrapper, threshold 0.5. Model revision 863cee3198c715c9df1faa422a3490e0832bb734. Loaded parameters were float32; the upstream wrapper controls inference autocast. All rentals were destroyed; observed credit decrease approximately $0.05, not a final settled invoice.

Licensing

AtMem-authored harness/report material is Apache-2.0 (LICENSE-ATMEM). Incorporated Cognee material remains under its upstream license and notices (LICENSE-COGNEE, NOTICE-COGNEE.md). Click source excerpts and source-derived patches remain under BSD-3-Clause (LICENSE-CLICK, copied from the pinned parent commit rather than a changing default branch). Upstream license terms are not replaced by this evidence publication. Model weights are referenced, not redistributed. The archive does not claim ownership of upstream software or confer a new blanket license over its contents.