No Artifact Left Behind (Day 31, metadata-forensics) - agent-4
Result: Single, Time Trial and Mastery all CORRECT.
| Mode | Result | Flag |
|---|---|---|
| Single Solve | CORRECT | 824a1d0aba99da644c12c8c40dd5bd1e |
| Time Trial | CORRECT (see note on the first attempt) | fb557f6dae4ee8e4c2ef4084a0cf21a7 |
| Mastery | CORRECT (ranked) | b86c1056ddc42450c42993a8a7e1eeed |
The new layer
Only one field was new: graph.execution_chain_artifact_ids for case-001.zip (challenge artifact, not published). Every other id in the Time Trial (36) and Mastery (540) used extractors that already existed. The answer is a compact JSON array of node ids (member basename, extensionless), ordered from the chain start to its end.
case-001.zip (challenge artifact, not published) is a graph of extensionless nodes, each pointing at the next by name:
- PowerShell stage script: non-comment line with
.\<node>(Invoke-Item / & / Start-Process) - PNG:
tEXtCommentnext=<node>; the end node carriesdone <tag>(Single only) - SQLite:
launch_queue.target - nested zip:
run.txt=execute <node>(readme.txtis noise) - EML: attachment
route.txt(open:/forwarded to:/continue with:); a plain-body EML has no route - JSON ticket:
{"task":..,"next":"<node>","owner":..}(Time Trial only)
Trap ("the tempting conclusion the corpus does not support"): most nodes form one closed loop (decoy) that includes the delivery hops. The real chain is the separate, acyclic component: exactly one node that nothing points to, walked forward to its sink.
Single answer: ["birch-bd18","nectar-1767","bramble-e622","orchard-e5bc"]. Time Trial variant (no done PNG; the sink is an EML note): ["hollow-f48b","ember-b6eb","quartz-9497"].
What happened / stall
- Claim, autosolve HELD on the unknown field
graph.execution_chain_artifact_ids(1.7 s). The GM residue carried no words, so the structure came from reading the downloadedcase-001.zip(challenge artifact, not published) directly. - First extractor keyed on the PNG
donemarker (walked backward from it). Single went CORRECT. - The runner then opened the Time Trial on its own and submitted; the batch came back incorrect. The TT
case-001.zip(challenge artifact, not published) had nodonePNG (JSON ticket nodes, an EML sink), so my extractor returned None for that id. I confirmed this offline againsttime_trial_0.bin(challenge artifact, not published) and rewrote the finder to be structural (unique source -> sink, closed loops rejected,doneused only as a tie-break). That spent the first TT attempt; the rerun took the next fresh variant and went CORRECT, then Mastery CORRECT. Caveat: I cannot tell from the GM whether the None was the only wrong id in that first batch; I did not see a per-id verdict. - Total stall on the new field: roughly 25 minutes including the wasted TT attempt.
Lessons + D+1
- Missing: the
graph.*family. Addedgraph.execution_chain_artifact_idsto the per-challenge clone (scripts/metadata_wip.py), registered through FIELDS.setdefault. Port it to the sharedautosolve_metadata.pyif D+1 asks anygraph.*field again; the node parsers (ps1/png/sqlite/zip/eml/json) are the reusable part. - Do not key a graph answer on one marker seen in the Single (
done). The Time Trial/Mastery variants change node types and the terminal. Build the extractor from structure and verify it against every artifact on disk (single_0.bin(challenge artifact, not published) ANDtime_trial_0.bin(challenge artifact, not published)) before the runner is allowed to start a timed mode. The ranked TT attempt was burned because only the Single was checked. - Sibling fields to expect: the same corpus likely supports delivery-chain, loop-members (the decoy cycle) and chain-length questions; the edge table built here answers them with a few lines.
- Offline helpers kept in
scripts/:test_graph_chain.py,offline_tt.py,dump_case.py,dump_eml.py.