What Is an Evidence Bundle? (REA Glossary)
Evidence bundle (noun): the exportable package of every content-addressed evidence record an analysis agent retained during one session. Each record inside it pairs a claim — "31 sections at entry point 0x400320" — with the proofs that make the claim checkable: the analyzed input's sha256, the provider and exact library versions used, and an evidence ID derived from the record itself. An evidence bundle is how "trust me, I analyzed this binary" becomes "here are the receipts — re-run it yourself."
evidence bundle, in plain words
Think of the difference between a screenshot of a lab result and a lab report with samples attached. A plain tool output is the screenshot: true at the moment it appeared, impossible to audit later. An evidence bundle is the report: every finding carries its own chain of custody, bundled so a reviewer can verify each link independently.
In the REA context specifically, records accumulate in-session and are retrieved as a bundle through the MCP layer, then converted to a portable canonical file on disk. We ran that exact loop on 2026-10-12 in our own lab. To see where these tools are developed, take the rea GitHub repository tour.
MCP tool chain — lab@realearn verified: ran 2026-10-12
inspect_binary_layout(hello) → ev_6e479c83… (session-retained) $ get_evidence_bundle(detail:summary) total_retained_records: 1 ✓ $ export_evidence_bundle → hello-bundle.json (39113 bytes, records:1, unknowns:0)
Two details matter for the definition. First, the bundle is session-scoped: it is what the agent retained during one working session, not a global database. Second, the counts are explicit — 39113 bytes, 1 record, 0 unknowns — because an evidence bundle that does not state its own contents is just a file.
what goes inside each record
A bundle is only as good as its records. Each one we inspected carried four proof fields:
- evidence_id — content-addressed, so the ID itself is the integrity check (see: content-addressed ID).
- subject sha256 — exactly which bytes were analyzed.
- provider + pinned versions — in our run, pwntools-elf @ pwntools 4.15.0; pyelftools 0.33; unicorn 2.1.2.
- normalized_result — the finding, in a form your own tools can re-check. Ours matched manual
readelfoutput.
evidence bundle vs. related terms
- vs. evidence contract — the contract is the design rule (every result must carry its proofs, and the system must fail closed rather than guess); the bundle is the artifact that rule produces. Rule vs. output.
- vs. a raw evidence record — a single record is one entry; the bundle is the verified collection. The exporter refuses a bare record fed directly: it demands a session bundle whose integrity it can check end to end. We hit that refusal ourselves (
evidence_integrity_mismatch) — walkthrough in the contract guide. - vs. a log file — logs are append-only text for humans; a bundle is structured, addressed, and machine-verifiable.
when you would actually use one
Three everyday cases. Handing analysis to a teammate: the pinned provider versions end "works on my machine" before it starts. Writing a report: attach the bundle and reviewers spot-check instead of trusting. Reproducing weeks later: re-run the same command against the same sha256 and the content-addressed IDs tell you immediately whether anything drifted.
FAQ
What is an evidence bundle in simple terms?
The exportable package of every evidence record a session produced. Each record pairs a finding with its proofs — the input sha256, the exact tool versions, and a content-addressed ID — so anyone can re-verify the claim instead of taking it on faith.
Is a rea evidence bundle the same as the rea evidence contract?
No. The contract is the design principle — every result carries proofs, and anything unverifiable is refused rather than guessed. The bundle is the artifact that principle produces. Our REA evidence contract guide covers the mechanism; this page covers the vocabulary.
Can I export a single evidence record directly?
In our 2026-10-12 run, feeding one raw record into the exporter failed with evidence_integrity_mismatch. The exporter only accepts a session bundle exported through the MCP layer first — a refusal that is itself the audit trail working.
Is realearn affiliated with morluto or rea.tools?
No. realearn is independent and non-affiliated; we ran every command referenced here in our own lab and mark nothing verified that we did not execute.
Related: content-addressed ID · fail-closed · the REA evidence contract (deep guide) · all our evidence practices