Evidence ledger

The evidence ledger is the record of every tool call a run made. Each entry has an id such as ev_0094. Agents cite these ids in their claims, the judge cites them in the verdict, and the report cites them in every section. A reader can open any citation and see exactly what the tool was asked and what it answered.

How an entry is madeCopy Anchor Link

The tool loop wraps every tool. When an agent calls one, the loop times the call, records the arguments and the answer, and hands the answer back to the model with the entry's id on the front:

[ev_0094]
{"0x1360bc0b344": "int FUN_1360bc0b344(undefined8 *param_1) { ... }", ...}

The model sees the id with the answer, so it can cite it. The entry holds the answer whole, exactly as the model received it.

Ids count up across the whole run. The triage pack's calls come first, under the agent pipeline. The judge's own calls and one agent's question to another go through the same recorder, so a verdict that leans on them can cite them too.

What an entry holdsCopy Anchor Link

This is ev_0094 from the benchmark run, as GET /api/v1/jobs/{job_id}/evidence returns it, shortened:

{
  "entry_id": "ev_0094",
  "stage": "reversing",
  "agent": "all_tools_reverser_ghidra",
  "tool": "batch_decompile",
  "source": "provider:ghidra",
  "model": "anthropic/claude-haiku-5-5",
  "ok": true,
  "duration_ms": 474,
  "args": {"program": null, "functions": "0x1360bc0378c,0x1360bc0cf14"},
  "output": "{\"0x1360bc0378c\":\"...bool FUN_1360bc0378c(...) ...\"}",
  "truncated": false,
  "repeated_of": null,
  "facts": {"v": 1, "kind": "decompile_batch", "effects": {"network": false, "executes": "no", "third_party": "no"}, "arguments": {"program": "program"}},
  "digest": "94745b5d…",
  "chain": "75a595bd…"
}

How claims use itCopy Anchor Link

An analyst's claim names the entries it rests on. This is one of the triage agent's claims from the same run, as the stored findings hold it:

{
  "claim": "The decoded strings include system-discovery command lines: ipconfig /all, systeminfo, ver, whoami /groups, net config workstation, net view /all /domain, nltest /domain_trusts (and /all_trusts), net group \"Domain Admins\" /domain, and the WMI query for AntiVirusProduct. Code refers to these strings through the command routines. This is evidence of intent to run them, not of their execution.",
  "evidence_ref": "[ev_0019] [ev_0021]",
  "confidence": 0.75,
  "technique_id": "T1033"
}

ev_0019 is the pack's FLOSS answer and ev_0021 its decoded string blobs. The command lines the claim names are decoded strings in them.

Code checks every citation. A claim that cites no entry, cites one that does not exist, or cites one whose answer does not hold the value the claim states is sent back to the analyst with a validation code such as isr.ungrounded_technique or report.citation_wrong_entry. The analyst gets a turn to fix it.

The report turns citations into numbered notes. Each note lists its entries, and the evidence bundle beside the report prints every entry the notes name.

Sealed when storedCopy Anchor Link

When the run is stored, the worker seals the ledger. Each entry gets digest, a SHA-256 of everything stored for it, and chain, an HMAC that links it to the entry before. The chain head is written to the audit chain beside the report's own digest. A changed, removed or reordered entry, or a changed report, fails verification.

curl -s http://localhost:8000/api/v1/jobs/$JOB/evidence/verify \
  -H "Authorization: Bearer $TOKEN"

An enrichment that changes a sealed report later appends an amendment. Sealed rows are never rewritten. See Sealed evidence ledger.

Reading the ledgerCopy Anchor Link

The full entry format is in Architecture.