This report makes two kinds of claim. Claims about the published literature are transcribed from primary sources and can be checked directly against those papers; the transcriptions are mirrored here as CSV so a reader can verify the arithmetic without re-reading the originals. Claims about a deployed AGNT instance (§6) rest on three InsightEngine records, reproduced here verbatim.
| File | Backs | Contents |
|---|---|---|
| insights.json | §6, Table 4; §8 | Three AGNT InsightEngine records retrieved from GET /api/insights/:id on 2026-07-25, verbatim except user_id, which is redacted. Records d9a5a00f (tool_preference, conf 0.88), 7e949534 (antipattern unrelated-trace-context-contamination, conf 0.60), 6ce90af2 (memory, conf 0.95, status applied). |
| ctim-rover-tables.csv | Tables 1–2, Figures 1–2 | CTIM-Rover per-repository success rates joined to turn and token statistics. Transcribed from arXiv:2505.23422, Tables 1 and 3. |
| awm-webarena-table1.csv | Table 3 | AWM WebArena results and average step counts. Transcribed from arXiv:2409.07429, Table 1. |
| SHA256SUMS.txt | — | SHA-256 over the three files above. |
On a running AGNT instance belonging to the same user:
curl -H "Authorization: Bearer $AGNT_AUTH_TOKEN" \
http://localhost:3333/api/insights/d9a5a00f-b3fe-4eaa-81ca-51856e790ab5
curl -H "Authorization: Bearer $AGNT_AUTH_TOKEN" \
http://localhost:3333/api/insights/7e949534-96bf-4668-acea-970372ed6c9a
curl -H "Authorization: Bearer $AGNT_AUTH_TOKEN" \
http://localhost:3333/api/insights/6ce90af2-6f37-4073-b2ff-384f246c1a0c
The two records supporting Table 4 were created 27 minutes 31 seconds apart (2026-06-27 04:01:59 and 2026-06-27 04:29:30) from an identical tool sequence, and differ in source_type (agent_chat vs goal) and therefore in the task the recalled material was serving. The evidence field of each record enumerates the tool calls observed, which is what establishes that the sequences match.
github.com/Liqs-v2/ctim-rover, github.com/zorazrw/agent-workflow-memory).insights.json establishes that the two opposing verdicts exist, with the stated timestamps, categories, confidences and evidence strings. It does not establish that the contaminated run performed worse on its task; both verdicts are judgments produced by an LLM-based classifier over traces. §6 states this limitation in the body text and no causal claim is made from it.instrument-first-engineering skill quoted in §8 is a user document on a private machine and is not redistributed here. The quoted passage and the file size (15,502 bytes) are reproduced in the paper; insight 6ce90af2 in insights.json records its creation path and applied status.