arcaeon-ledger

Observability tools show you what your agent did. arcaeon-ledger lets you prove it.

Every record is hash-chained to the one before it. Edit a row, delete one, or reorder history, and every later link breaks — verify names the exact line. You own the record, and you can prove it wasn't altered. Zero dependencies, one JSONL file, two verbs.

pip install arcaeon-ledger      # then:  from arcaeon_ledger import Ledger

from arcaeon_ledger import Ledger

log = Ledger("agent.log.jsonl")
log.append({"tool": "web.search", "query": "weather in LA", "result_ok": True})
log.append({"tool": "payment", "amount": "49.00", "currency": "USD"})

log.verify()          # VerifyResult(ok=True, rows=2, chained=2, ...)

Tampering is caught, not hoped against:

# someone edits row 2's amount in the file by hand...
log.verify()          # VerifyResult(ok=False, first_break="line 2: chain mismatch")

CLI (wire it into CI or a pre-ship gate — a tampered log exits nonzero, and a log that could only be partially vouched for no longer exits like a fully verified one):

python -m arcaeon_ledger.cli append agent.log.jsonl '{"tool":"search","ok":true}'
python -m arcaeon_ledger.cli verify agent.log.jsonl
python -m arcaeon_ledger.cli verify --strict agent.log.jsonl

verify exit codes (0.5.7):

A CI gate should treat only 0 as green:

python -m arcaeon_ledger.cli verify agent.log.jsonl
case $? in
  0) echo "fully verified" ;;
  3) echo "chain intact but prechain rows skipped unverified — inspect, or use --strict" ; exit 1 ;;
  *) echo "ledger broken" ; exit 1 ;;
esac

Prove who acted, not just the order

A hash chain proves sequence integrity — it can't prove who wrote each entry or whether they were allowed to. Attach an authority block to bind the actor and their permission surface into the chained (tamper-evident) row:

from arcaeon_ledger import Ledger, authority

log = Ledger("agent.log.jsonl")
log.append(
    {"tool": "payment", "amount": "49.00"},
    authority=authority(
        "agent://billing-7",
        capability_version="v3",              # what they were allowed to do
        tool_schema={"name": "payment", "args": ["amount"]},  # hashed, not just named
        time_source="ntp",                    # trust surface of the clock
    ),
)

Now the audit question sharpens from "was this edited?" to "was this edited and was the writer authorized?" — editing the principal, capability, or schema hash breaks the chain like any other tamper. This composes tamper-evidence with permission-replay. (Shipped in response to community feedback on launch.)

Related MCP server: AgentAudit MCP

Why this exists

The loudest unmet pain for agent builders in 2026 is the reliability/audit gap: an agent "completes" a task and the result is quietly wrong, and you can't reconstruct — or prove — what actually happened. Observability platforms trace runs; none give you a tamper-evident, portable, ownable record. Regulation is arriving too: the EU AI Act requires high-risk systems to technically allow automatic recording of events over their lifetime (Art. 12(1)) and requires providers and deployers to keep those logs, to the extent under their control, for at least six months (Art. 19(1), Art. 26(6)). The Act mandates recording and retention — tamper-evidence is not its word, it is ours: when someone asks whether a retained log is still the log, that question needs an answer stronger than trust. arcaeon-ledger is the smallest honest version: a cryptographically chained action log you drop in, own, and verify.

How the chain works

chain = sha256(prev_chain + json.dumps(row_without_chain, sort_keys=True, ensure_ascii=False))[:32]

Note the separators: the chain body uses Python's default ", " / ": " spacing, not the compact json-c14n:v1 form the artefact digests use. A cross-language verifier has to reproduce that spacing exactly.

The chain value is truncated_sha256_128 — the first 32 hex chars (128 bits) of SHA-256, not the full digest. Named so nobody cites it as full SHA-256: 128 bits is plenty for edit/accident detection, thinner if you want the chain itself to be expensive to grind after a rewrite (credit: atomic-raven's review).

Each row commits to the entire history before it. The first row chains from a fixed "genesis" seed. Rows without a chain field are tolerated only before the first chained row (so you can adopt it on an existing log); an unchained row appearing after the chain begins is itself flagged. On a mismatch, verify keeps going from the claimed value so it counts later damage honestly instead of cascading one break into noise.

What it proves — and the five things it doesn't

Being precise here is the product, not a disclaimer. A hash chain proves the recorded content of each row was not altered in place after writing: mid-file edit, delete, and reorder all break it and verify names the row.

One word in that sentence changed in 0.5.8, and the reason is the kind of thing this section exists for. It used to say "the recorded bytes", which claims more than the chain does. The chain is computed over each row parsed back from the file, and the reader normalises byte sequences it cannot decode — so two different byte strings inside such a region read identically and produce the same verdict. What is protected is the meaning of every row, not the exact bytes of the file. If you need byte-level custody, hash the file itself alongside this.

It does not by itself prove five other things:

1. Truncation. Lop off the most recent rows and what remains verifies clean — no append-only chain catches this alone. Close it by publishing the head somewhere outside your own control, on a cadence:

pin = log.head().as_pin()
# -> "arcaeon-ledger head chain=9f3c… rows=204 as_of=2026-08-13T17:40:00Z"
# post `pin` to a git commit / public comment / notarization anchor.
# a reader compares a fresh head() against the last pin; a truncated or
# re-minted history disagrees. the MAX gap between pins is your security
# parameter, not the average — an attacker picks the gap.

2. Truth. The chain notarizes whatever was written — a tamper-evident record of a hallucination is still a hallucination with a checksum. To make a row speak about the world, hash a re-fetchable artefact (URL+bytes, a snapshot, tool stdout) and store that digest in the row, so a third party can re-get it and compare.

3. Authorship. authority() (above) records who-claimed-what, but it is data in the row, not a signature — a rewriter who re-mints from genesis re-mints it too. External head-anchoring (#1) is the thing a re-minter cannot advance.

4. Fabricated-legacy-prepend. Rows with no chain field are tolerated before the first chained row — that is deliberate, so you can adopt the chain on top of an existing log without rewriting its history. But skipped rows are unverified rows, and the verifier cannot tell real legacy history from a fabricated prepend. So (0.5.7) a non-strict verify that skipped any rows never mints a green: ok is None — "no break found, verified within scope" — falsy, with the scope in-band (verified_scope: "bounded_prechain_skipped") and the count in prechain; the CLI exits 3, not 0. Only a scan that checked every row returns ok=True. If your log is chained from genesis and must have no legitimate legacy rows, pass verify(strict=True) / --strict — it treats any unchained row as a break, hard red. (An unchained row inserted after the chain begins is already flagged in every mode.)

5. Completeness. This is the big one, and it is structural: the agent decides what to call append on. A tamper-evident log of the calls an agent chose to report is still self-report. Nothing inside this library can close that, because anything the agent invokes, the agent can decline to invoke.

Close it by moving the pen out of the agent's reach — record at the seam instead, in a separate OS process the agent does not own, cannot skip, and cannot see:

pip install arcaeon-adapter

python -m arcaeon_adapter --ledger seam.log.jsonl -- <your mcp server command...>

arcaeon-adapter is a stdio proxy that forwards JSON-RPC byte-for-byte between an MCP client and server, writing one hash-chained row per tools/call to its own ledger. Wrapping it around this library's own MCP server produced the number that makes the point: the server's own diary wrote 0 rows while the seam log captured 5. The gap between what a system reports about itself and what the seam observed is the thing worth measuring.

Scoped honestly, the primitive is "this file was not rewritten in place" — small, true, and testable. The layers above (external anchoring via head(), artefact binding, signed authorship, seam recording) are how you extend it toward a full evidence claim.

verify() on missing or empty ledgers

The two look like the same thing ("no data"), and verify() keeps them apart, on purpose:

Ledger("never/written.jsonl").verify()
# VerifyResult(ok=False, rows=0, first_break="unreadable: [Errno 2] No such file...")

open("touched/empty.jsonl", "w").close()
Ledger("touched/empty.jsonl").verify()
# VerifyResult(ok=None, rows=0, chained=0, first_break=None, verified_scope="empty")

A path that was never created can't be vouched for — ok=False, "unreadable," same as any other read failure. A path that exists and is genuinely empty has zero rows to tamper with, but zero rows checked is not a green either (since 0.5.8): ok=None, rows=0, verified_scope="empty", falsy, CLI exit 3. Automation that branches on verify().ok gets a red for the missing file and a not-a-pass for the empty one; read first_break and verified_scope to tell the two apart by name.

When the log was written out of band: declare the break, don't re-forge it

Sooner or later something writes to your JSONL without going through append() — a script, an incident, a person with an editor. The chain breaks there and stays broken, because that is the true record. Your two obvious options are both bad: live with a permanent red that tells a reader nothing, or recompute the chain so the file goes green — which is forging it, and a chain you can silently re-forge is not evidence of anything.

declare_break is the third option. It appends a row naming the break:

from arcaeon_ledger import declare_break, verify_file

declare_break("agent.log.jsonl", 25,
              "Written out of band 2026-08-15 by a session hand-appending JSON "
              "instead of calling append(). Content is true and preserved verbatim; "
              "no chain value was ever computed for it, so none can honestly be supplied.")

r = verify_file("agent.log.jsonl")
r.ok               # None  — bounded, NOT True. Falsy.
r.verified_scope   # "bounded_declared_break"
r.breaks           # 0
r.declared         # ["line 25: declared break (Written out of band 2026-08-15 ...)"]

The break stays a break, forever, in declared. What changes is that a known, explained break stops masquerading as an unexplained one — and the orphan's exact bytes are pinned by sha256, so editing that line afterwards turns the file red again. It never returns ok=True. Only a scan that checked every row does that, and an excused row was not checked. verify(strict=True) ignores declarations entirely.

What this does not do, said plainly: it is a record device, not a cryptographic one. Anyone who can write the file can write a declaration, so it raises no bar at all against an attacker who already has write access. It defends against forgetting, not against tampering. It cannot tell an honest out-of-band append from a malicious one — why is an unverified human sentence. And it can only declare breaks verify() already found; it does nothing about breaks nobody noticed. Use it to keep an honest incident legible, never as a way to make a ledger green.

Bind what the agent actually read (artefact-binding)

The chain proves a row wasn't edited. It does not prove the row was ever true — it will notarize a hallucination as faithfully as a fact. bind_artefact closes that gap for the cases where you can point at a re-fetchable source: hash the actual bytes the agent read and store that digest in the row, so a third party can re-get the source and compare.

from arcaeon_ledger import Ledger, bind_artefact

log = Ledger("agent.log.jsonl")
art = bind_artefact("https://example.com/pricing")   # or bytes, a file path, or a dict
log.append({"tool": "web.read", "url": "https://example.com/pricing", "artefact": art})
# art -> {"subject": {"name": "...", "digest": {"sha256": "..."}},
#         "recipe": "sha256:raw-bytes:v1",
#         "digest": "sha256:raw-bytes:v1:<hex>", "bound_at": "...", "source_meta": {...}}

Digests are self-describing — never a bare hex hash. Each one is sha256:<recipe>:<version>:<hex>, carrying its own recipe so a stranger reproduces it from the string alone: raw-bytes:v1 (opaque bytes as-read) or json-c14n:v1 (a pinned, documented JSON canonicalization — sorted keys, compact, UTF-8). Recipes are frozen and versioned append-only, so old rows keep their recipe forever and a changed rule never makes history look tampered.

Verify honestly:

from arcaeon_ledger import verify_artefact

verify_artefact(art)                    # recipe reproducible + string self-consistent
verify_artefact(art, refetch=True)      # for a URL: re-fetch and compare
# -> {"verdict": "live_match",          # <- THE answer; read this field
#     "digest_ok": True, "reason": None,
#     "refetch": "match" | "mismatch" | "unavailable" | "skipped", "notes": [...]}

Read verdict, not just digest_ok (0.5.7). digest_ok names only the offline leg — recipe reproducible, string self-consistent — and it stays True even when a live re-fetch disagrees. The top-level verdict tag mints the whole answer in one field: "digest_consistent" (offline leg passed, no live comparison made), "live_match", "live_mismatch" (live content no longer matches — changed or tampered, indeterminate), "live_unavailable" (the requested live check could not run), or the typed failure reason itself when the offline leg fails. if out["digest_ok"] after refetch=True used to read green through a live mismatch; out["verdict"] == "live_match" cannot.

A label this build cannot reproduce is a typed failure, never a pass. If the digest names an algorithm, recipe, or recipe version outside the supported registry, verify_artefact returns digest_ok=False with a machine-readable reason — one of unknown_algorithm, unknown_recipe, unknown_recipe_version, malformed_digest, subject_digest_mismatch — and never reaches the re-fetch stage, so an unverifiable recipe can't come back as "match". A digest we cannot recompute is a digest we did not check, and "did not check" must not be reported as "verified." Old versions stay verifiable by staying listed in SUPPORTED_RECIPE_VERSIONS when a new one is minted, so the append-only recipe promise holds without the verifier waving through labels it has never shipped.

The honest boundary, stated loudly because it is the point: a re-fetch mismatch means the content changed or was tampered — indeterminate. It is never reported as proof of tampering. The web mutates, 404s, paywalls, and personalizes; binding proves "this is the digest of the bytes the agent said it read at time T," nothing stronger. For a neutral capture rather than your own fetch, route the source through a notarizing snapshot; for existed-before-T, anchor the digest externally. Each is a layer you add — stated, not implied.

The outside check: an external witness

The chain can't catch truncation alone — lop off the most recent rows and what remains verifies clean (stated in "what it doesn't prove", above). The fix is a witness: a record-keeper outside your own control that holds your head (rows, chain) on a cadence. Once a witness has a pin from time T, a truncated log has fewer rows than the witness saw, and a rewritten one has a different chain at the witnessed row. Neither can hide.

from arcaeon_ledger import Ledger, WitnessStore, publish_head, verify_against_witness

log = Ledger("agent.log.jsonl")
witness = WitnessStore("witness_pins.jsonl")   # ideally on a host you don't control

publish_head(witness, "billing-agent", log)    # record the current head — do this on a cadence

# later — did the log survive intact?
v = verify_against_witness(witness, "billing-agent", log)
v.verdict     # "consistent" | "truncated" | "rewritten" | "no_record" | "witness_broken" | "local_broken"
bool(v)       # truthy ONLY on "consistent" — a missing pin is no_record, never a false ok

# READ THE VERDICT WITH ITS QUALIFIERS, never the bare string alone:
v.witness_self_integrity   # "verified" | "unestablished" | "broken"

A bare "consistent" is not the whole answer. The verdict also carries witness_self_integrity: whether the witness store could prove its own pin chain intact. A hosted client that only exposes latest() cannot self-verify, so its verdicts read unestablished — the comparison ran honestly, but a forged pin served by that store would compare clean. verified means the store's own chain was recomputed; broken means it failed. A consumer that branches on v.verdict == "consistent" without reading witness_self_integrity is trusting the store's honesty exactly as much as it would trust the log's — which is the arrangement a witness exists to replace. (Found in the 2026-08-23 pre-invite audit, C14; the field exists so "not checked" can never render as "checked and fine.")

WitnessStore is the reference witness: one append-only JSONL file of pins. A hosted witness is a thin HTTP wrapper over exactly this object; run it locally and you have a complete, offline, zero-cost witness you fully control (with the obvious caveat that a witness you control is only as independent as its host).

What this proves, exactly. A witness proves your log wasn't truncated or rewritten only relative to what the witness saw, and only as recently as the last pin. Rows appended after the last pin are unprotected until the next one — so the MAX gap between pins is your real security parameter, not the average, because an attacker picks the gap. And it says nothing about whether the logged content was true — that's artefact-binding's job (above); the witness only guards the history's shape.

What the witness holds. Only fingerprints — (namespace, rows, chain, time) — never your log content. Password-nowhere by design: if the witness is breached, there is nothing sensitive to steal, only hashes useless without the original log.

Drop it into any MCP agent

arcaeon-ledger ships a zero-dependency MCP server, so any MCP client (Claude Code, etc.) can give its agent tamper-evident logging with no code. Wire it in:

{
  "mcpServers": {
    "ledger": {
      "command": "python",
      "args": ["-m", "arcaeon_ledger.mcp_server", "--log", "agent.log.jsonl"]
    }
  }
}

The agent then has five tools. Two are operator tools over one file: ledger_append(record) to log an action (returns its chain hash) and ledger_verify(strict?) to prove the log is intact (or get the exact tampered line back). The verify verdict is three-valued, same as the library: ok: true = every row verified, ok: null = chain intact but unchained prechain rows were skipped unverified (verified_scope: "bounded_prechain_skipped" — not a green), ok: false = broken. Pass strict: true to make any unchained row a hard failure.

Three are agent tools (0.7.0), for when the output is going to somebody — a principal who wants proof, or a peer deciding whether to trust you:

prove_my_conduct re-verifies after appending, so an agent whose ledger has been tampered with gets chain_verified: false rather than a head hash with a green attached. verify_peer_ledger returns first_break as an integer line number (or null) so a calling agent can point at the exact bad row — it never touches your ledger (the export is verified from a throwaway temp file, and the call itself lands one row in <log>.calls.jsonl like every other tool call), and an export with no parseable rows returns ok: null (verified_scope: "empty"), because a green for sending nothing is the cheapest possible forgery. declare_break refuses when nothing is broken, and never restores a green.

Agent ledgers live one file per namespace under --ns-dir (default ledgers/ beside --log). A namespace is a name, not a path: [A-Za-z0-9._-], traversal refused rather than sanitized.

MCP is JSON-RPC over stdio and this server speaks it directly — no SDK, no extra install.

Status

Core library, CLI, and a drop-in MCP server, all tested: the library against edit / delete / reorder tampering (test_ledger.py), the MCP server through tools/list → append → verify at the request handler including tamper detection, and the agent tools against namespace traversal, peer-export tampering by exact line, and declared breaks (test_agent_tools.py). Extracted from a hash-chained action ledger running in production. External anchoring ships via head() (publish the pin yourself) and the reference witness (WitnessStore, above); a hosted witness tier (retention, automatic pin cadence, compliance export) is the next layer.

MIT.