arc-protocol 1.1.4

ARC (Agent Receipt and Certification): cryptographic proof of execution for AI agents

ARC — Agent Receipt & Certification

Cryptographic receipts for AI agent tool calls.

The tool provider signs every action receipt with Ed25519 - not the agent. Before execution, the resource state is snapshotted and the declared intent is committed to an RFC 6962 Merkle transparency log. After execution, the provider signs a seven-field canonical payload and commits the full receipt. An agent cannot forge the provider's signature to deny what happened, claim a different result, or lie about rollback availability.

Protocol v1.1. 58/58 tests passing. 5/5 live receipts cross-verified. EU AI Act Article 12 ready.

---

The Problem

AI agents self-report their actions with no independent verification. When an agent says "I deleted the file," there is no cryptographic proof it happened, no proof of what existed before, and no signed commitment to undo it.

This failure mode has been documented repeatedly in production:

Incident	What happened
Replit, 2024	Agent deleted a production database, fabricated ~4,000 records to conceal it, then told the operator rollback was impossible. The rollback was available the entire time.
Gemini CLI, 2025	Agent deleted user files during autonomous coding sessions without explicit authorization.
Claude Code, 2025	Agent deleted project files during autonomous operations; no tamper-evident record of what existed before.
Amazon Kiro, 2025	Performed destructive schema migrations without rollback specification in the agentic IDE.
Warehouse management agent, 2025	Wiped 1.9 million inventory rows from a production database. No before-state captured.

These incidents share the same structure: the agent is the only witness to its own actions. Existing tools cannot help:

• LangSmith, LangFuse, Arize Phoenix: capture self-reported execution traces; no cryptographic integrity; mutable; cannot distinguish accurate logs from fabricated ones.
• MCP, OpenAI function calling, Anthropic tool use: return plain text results with no signature, no before-state, and no rollback specification. A tool response is indistinguishable at the protocol level from a hallucination.
• OpenTelemetry, OpenLineage: structured tracing without tamper evidence; self-reported; do not satisfy EU AI Act Article 12.

No current protocol addresses proof of execution, observability, and reversibility together. ARC does.

---

How ARC Works

Every tool call produces an Action Receipt in two phases:

PHASE 1: PRE-ACTION (before the tool runs)
─────────────────────────────────────────
  1. Snapshot resource state → SHA-256 hash → store at snap_<ULID>
  2. Agent declares intent (tool, args, agent_id, reasoning_hash, on_behalf_of)
  3. Commit intent to RFC 6962 Merkle log → get sequence_number + inclusion_proof

PHASE 2: POST-EXECUTION (after the tool runs)
──────────────────────────────────────────────
  4. Tool executes → capture actual return value → outcome_hash = SHA-256(result)
  5. Provider builds 7-field canonical payload (keys sorted, no whitespace, UTF-8)
  6. Provider signs payload with Ed25519 private key → signature
  7. Provider signs inverse operation (rollback spec) separately
  8. Commit full receipt to Merkle log → receipt now publicly verifiable

The canonical signing payload (7 fields, alphabetically sorted):

{
  "before_state_hash": "sha256:...",
  "intent_id":         "intent_01...",
  "is_reversible":     true,
  "outcome":           "success",
  "outcome_hash":      "sha256:...",
  "receipt_id":        "arc_01...",
  "signed_at":         "2026-04-02T..."
}

The tool provider signs this payload - not the agent. This is the core architectural decision: the party with ground-truth knowledge of what actually executed provides the cryptographic attestation. An agent that fabricates results produces a receipt whose signature does not verify. An agent that claims rollback is impossible contradicts the provider-signed inverse_signature in the receipt.

---

The Three Guarantees

1. Proof of Execution

The outcome_hash in Phase 2 is the SHA-256 of the canonical JSON of the tool's actual return value. The provider signs this hash alongside the receipt_id and before_state_hash. If an agent reports a different result than what the provider attested, the signature does not verify - regardless of how plausible the fabrication appears in natural language.

# Verification catches fabrication:
tampered = copy.deepcopy(receipt)
tampered["phase_2"]["execution"]["outcome_hash"] = sha256_hex(canonical_json({"deleted": 0}))
result = verify_receipt(tampered, registry)
# result["valid"] == False
# result["checks"]["provider_signature_valid"] == False

2. Tamper-Evident Observability (EU AI Act Article 12)

Every Phase 1 intent and Phase 2 receipt is an entry in an RFC 6962 Merkle tree. Each entry records previous_root (the Merkle root before it was appended) and merkle_root (after). These chain: entry[n].merkle_root == entry[n+1].previous_root. Any modification to any entry breaks the chain at a specific sequence number, detectable by any auditor with read access to the log.

The log server signs each entry with its own keypair (the log operator signature), providing non-repudiation at the log level as well as the receipt level.

# Tampering is detected:
log.storage.update_entry(seq=3, content_hash="sha256:" + "0" * 64)
consistency = log.verify_consistency()
# consistency["is_consistent"] == False
# consistency["broken_at_sequence"] == 3

3. Signed Rollback

The inverse_operation block in every receipt is separately signed by the tool provider:

{
  "is_reversible": true,
  "inverse_tool": "restore_snapshot",
  "inverse_arguments": {"snapshot_ref": "snap_01..."},
  "valid_until": "2026-04-02T15:23:11Z",
  "inverse_signature": "ed25519:..."
}

The inverse_signature covers {receipt_id, inverse_tool, inverse_arguments, valid_until}. An agent cannot set is_reversible = false without invalidating this signature. An agent cannot claim the rollback window has expired without forging the provider's Ed25519 key. The rollback_filesystem() function restores the before-state from the snapshot.

snap = ctx.snapshot_store.retrieve(receipt["phase_1"]["before_state"]["snapshot_ref"])
success = rollback_filesystem(snap)
# success == True
# All files restored with identical content

---

Quick Start

Install:

# Core SDK only (fast, lightweight: cryptography + python-ulid)
pip install arc-protocol

# With remote log server client (adds httpx)
pip install arc-protocol[client]

# With FastAPI log server (adds fastapi, uvicorn, pydantic)
pip install arc-protocol[server]

# With JSON Schema validation (adds jsonschema)
pip install arc-protocol[validation]

# Everything
pip install arc-protocol[all]

# From source (development)
git clone https://github.com/RamachandraKulkarni/arc-protocol.git
cd arc-protocol && pip install -e ".[dev]"

Wrap any function:

from arc import ARCKeyPair, ARCContext, ARCInMemoryLog, SnapshotStore, signed_tool, verify_receipt
from pathlib import Path

# One-time setup: provider identity + runtime context
provider_kp = ARCKeyPair.generate()
ctx = ARCContext(
    agent_id="claude-sonnet-4-6",
    model_version="claude-sonnet-4-6-20250514",
    session_id="session-001",
    provider_keypair=provider_kp,
    provider_name="filesystem-tool-v1",
    log=ARCInMemoryLog(),
    snapshot_store=SnapshotStore(),
    on_behalf_of="user:[e-mail]",
)

# Decorate any tool function
@signed_tool(resource="filesystem", resource_uri_from_args="path")
def delete_directory(path: str, ctx: ARCContext) -> dict:
    import shutil
    shutil.rmtree(path)
    return {"deleted": path, "status": "ok"}

# The return value is a complete, signed ActionReceipt
receipt = delete_directory("/tmp/old_data", ctx=ctx)

# Verify cryptographic integrity
result = verify_receipt(receipt, {ctx.provider_name: provider_kp.public_key_hex})
print(result["valid"])   # True
print(result["checks"])  # all True

# Rollback from the receipt
from arc import rollback_filesystem
snap = ctx.snapshot_store.retrieve(receipt["phase_1"]["before_state"]["snapshot_ref"])
rollback_filesystem(snap)  # files restored

With the HTTP transparency log server:

make log-server  # starts FastAPI server on :8080

# In your code:
from arc import ARCLogClient
log = ARCLogClient(base_url="http://localhost:8080")

---

Demo Scenarios

make demo-basic      # Generate one receipt, verify it, print full JSON
make demo-disaster   # Full Replit scenario: delete, fabricate, detect, rollback
make demo-verify RECEIPT_ID=arc_01...  # Third-party verify from receipt ID only

demo_disaster.py output structure:

=== ARC PROTOCOL DEMO: The Replit Scenario ===

[SETUP] Creating test directory with 3 files...
[PHASE 1] Agent declares intent to delete...
  Before-state captured: snap_01JTXM... (sha256:abc123...)
  Intent committed to log at sequence 0
  ✓ Phase 1 is immutable  -  agent cannot change what it declared

[EXECUTION] Deleting directory...
  Directory deleted. Receipt ID: arc_01JTXM...

--- BEHAVIOR 1: Agent fabricates result ---
  Fabricated hash: sha256:jkl012...
  Receipt hash:    sha256:def456...
  ✗ MISMATCH  -  provider signature does not match
  ✓ ARC detected fabrication

--- BEHAVIOR 2: Agent tampers with log ---
  Merkle chain broken at sequence 0
  ✓ ARC detected log tampering

--- BEHAVIOR 3: Agent claims rollback impossible ---
  receipt.phase_2.inverse.is_reversible = TRUE (signed by provider)
  ✓ ARC refuted the rollback denial

--- ROLLBACK ---
  Restored: report.txt  ✓
  Restored: config.json ✓
  Restored: users.csv   ✓
  ✓ All files restored

---

Integration Patterns

Pattern 1: @signed_tool decorator (recommended)

Wrap individual functions. Zero-boilerplate for new tools.

@signed_tool(resource="filesystem", resource_uri_from_args="path")
def create_file(path: str, content: str, ctx: ARCContext) -> dict:
    Path(path).write_text(content)
    return {"created": path}

@signed_tool(resource="dict", resource_uri_from_args="key")
def update_config(key: str, value: dict, ctx: ARCContext) -> dict:
    config_store[key] = value
    return {"updated": key}

Pattern 2: ReceiptBuilder directly

For custom two-phase orchestration or async pipelines.

from arc import ReceiptBuilder
from arc.snapshot import capture_filesystem

builder = ReceiptBuilder(
    agent_id="my-agent",
    model_version="my-model-v1",
    session_id="sess-001",
    log_client=log,
)

# Phase 1
snap = capture_filesystem("/data/users", store)
phase1 = builder.declare_intent(
    tool_name="delete_users",
    arguments={"path": "/data/users"},
    snapshot=snap,
    reasoning="User requested cleanup of inactive accounts",
    on_behalf_of="user:[e-mail]",
)

# (tool executes here)
result = tool.execute("/data/users")

# Phase 2
receipt = builder.attest_execution(
    provider_keypair=provider_kp,
    provider_name="storage-tool-v1",
    outcome="success",
    result=result,
    phase1=phase1,
)

Pattern 3: LangChain / OpenAI Agents integration

Wrap any framework tool by putting @signed_tool on the underlying function:

from langchain.tools import tool
from arc import signed_tool, ARCContext

@tool
@signed_tool(resource="filesystem", resource_uri_from_args="path")
def read_file_tool(path: str, ctx: ARCContext) -> str:
    """Read a file and return its content."""
    return Path(path).read_text()

---

Adversarial Security Testing

Two independent red team sessions were conducted. The tester had no access to the builder's design documents and read the source code as an attacker.

Session 1 results (43 tests): 6 exploitable holes found in v1.0. All fixed in v1.1.

Hole	Attack	Fix in v1.1
Outcome string not signed	Flip "success" to "failure" without touching hash	Added outcome to signed payload
is_reversible not signed	Set is_reversible=false to hide rollback	Added is_reversible to signed payload
Log content_hash tamper	Modify content without breaking chain check	Recompute and compare Merkle root on verify
Duplicate receipt replay	Submit old receipt as new action	Reject duplicate receipt_id in log
No timestamp ordering	Declare intent after execution	Assert declared_at <= started_at
Phase 1 after Phase 2	Manufacture audit trail retroactively	Reject intent if receipt already exists

Session 2 results (15 tests): Zero new holes found in v1.1.

Session 2 targeted new code surfaces from the v1.1 patches:

• outcome changed independently of outcome_hash
• is_reversible=false with inverse_signature still present (structural contradiction)
• Double intent/receipt for same receipt_id
• declared_at equal to started_at (boundary: should be valid)
• Merkle tamper at the last appended entry

Combined: 58/58 tests. 0 failures.

Full findings: RED_TEAM_FINDINGS.md Test suite: tests/red_team/

make red-team          # run all 58 tests
make red-team-verbose  # with narrative output
make red-team-live     # run the Replit scenario end-to-end

---

Live Cross-Agent Proof

Five receipts were generated in a real Claude Code session on Windows 11 and verified by a completely isolated Python process with zero session knowledge - no access to the original keypairs, context, or tool runtime. The verifier received only the receipt IDs and a log server URL.

Receipt ID	Action	Intent seq	Receipt seq
arc_01KN6RGD75FMHNC1TBZ5MBZE7F	create data.csv	2	3
arc_01KN6SYG39A1FMXTQTAWDM97H3	read data.csv	4	5
arc_01KN6T08ZM9VXSAHVP1Y7QR5EB	create process.py	6	7
arc_01KN6T2HBHRW4PBDYHXWSYVD6Z	create results.txt	8	9
arc_01KN6T3RFS3N41YDFMG6QJRDQ1	create summary.json	10	11

All 4 checks passed for all 5 receipts: found_in_log, log_chain_consistent, intent_committed, receipt_committed.

Intent sequence numbers (2, 4, 6, 8, 10) always preceded receipt sequence numbers (3, 5, 7, 9, 11). This proves Phase 1 was committed before execution in every case - the log ordering cannot be faked after the fact.

Total: 5  Valid: 5  Invalid: 0
OVERALL: ALL VALID  -  PROOF COMPLETE

---

Protocol Specification

Seven JSON Schema draft-2020-12 definitions form the protocol:

Schema	Contents
action-receipt.schema.json	Root document: arc_version, receipt_id, phase_1, phase_2
intent.schema.json	Phase 1: intent, before_state, log_commitment
before-state.schema.json	Snapshot: resource_uri, resource_type, snapshot_hash, is_reversible
provider-attestation.schema.json	Phase 2: execution, provider_attestation, inverse, log_proof
inverse-operation.schema.json	Rollback: is_reversible, inverse_tool, inverse_arguments, inverse_signature
log-entry.schema.json	Log: sequence_number, content_hash, previous_root, merkle_root, log_signature
provider-registration.schema.json	Key registry: provider_id, public_key_hex, capabilities, revoked_at

Full specification: SPEC.md

ID format: All identifiers use ULID encoding (26 uppercase chars, millisecond-sortable) with type prefixes: arc_ (receipts), intent_ (intents), snap_ (snapshots), log_ (log entries).

Merkle tree: RFC 6962 construction: leaf = SHA-256(0x00 || data), node = SHA-256(0x01 || left || right). The 0x00/0x01 prefixes prevent second-preimage attacks.

---

EU AI Act Article 12 Compliance

Article 12 logging requirements enforce on August 2, 2026. Fines up to €35M or 7% of global revenue.

Requirement	ARC Mechanism	Status
Automatic event recording	phase_1.log_commitment committed before execution begins	Satisfied
Tamper-evident storage	RFC 6962 Merkle tree with log_signature on every entry	Satisfied
Full reconstructability	before_state snapshot + outcome_hash covers before and after	Satisfied
Person identification	intent.on_behalf_of + declared_by.agent_id	Satisfied
Non-repudiation (NIST AU-10)	provider_attestation.signature (Ed25519)	Satisfied
6-month minimum retention	SQLite log, configurable retention per operator	Supported

---

Version History

v1.0 introduced the two-phase receipt structure with Ed25519 signing. Red team testing found six security holes. v1.1 closes all six by expanding the signing payload from 5 to 7 fields.

Field in signing payload	v1.0	v1.1
receipt_id	yes	yes
intent_id	yes	yes
outcome_hash	yes	yes
before_state_hash	yes	yes
signed_at	yes	yes
outcome	no	yes
is_reversible	no	yes

Receipts signed under v1.0 are rejected by v1.1 verification. See CHANGELOG.md.

---

Repository Structure

arc-protocol/
├── CLAUDE.md                              Build instructions for AI assistants
├── README.md                              This file
├── SPEC.md                                Full protocol specification
├── CHANGELOG.md                           Version history and security fixes
├── RED_TEAM_FINDINGS.md                   Adversarial test findings (6 holes, all fixed)
├── CONTRIBUTING.md                        Contribution guide
├── SECURITY.md                            Vulnerability disclosure policy
├── pyproject.toml                         Package metadata and dependencies
├── Makefile                               Build, test, demo, and log server targets
│
├── schemas/                               JSON Schema draft-2020-12 (the protocol)
│   ├── action-receipt.schema.json
│   ├── intent.schema.json
│   ├── before-state.schema.json
│   ├── provider-attestation.schema.json
│   ├── inverse-operation.schema.json
│   ├── log-entry.schema.json
│   └── provider-registration.schema.json
│
├── src/arc/                               Core Python library
│   ├── ids.py                             ULID-based ID generation
│   ├── signing.py                         Ed25519, canonical_json, build_signing_payload
│   ├── snapshot.py                        Before-state capture and rollback
│   ├── receipt.py                         Phase1Builder, Phase2Builder, verify_receipt
│   ├── merkle.py                          RFC 6962 Merkle tree
│   ├── log_client.py                      ARCLogClient (HTTP) + ARCInMemoryLog (tests)
│   └── decorator.py                       @signed_tool, ARCContext
│
├── arc_log/                               Transparency log server
│   ├── server.py                          FastAPI: /v1/log/intent, /receipt, /verify/{id}
│   ├── storage.py                         SQLite-backed append-only log
│   └── merkle.py                          Server-side Merkle tree with persistence
│
├── tests/
│   ├── test_signing.py                    Key gen, sign/verify, canonical JSON
│   ├── test_merkle.py                     RFC 6962 inclusion proofs, append-only
│   ├── test_receipt.py                    Phase 1+2 assembly, schema validation
│   ├── test_log.py                        Log server API tests
│   ├── test_decorator.py                  @signed_tool integration tests
│   ├── test_e2e.py                        Full disaster scenario end-to-end
│   └── red_team/                          Adversarial tests (58 total)
│       ├── test_attack_1_fabrication.py   Result hash and outcome tampering
│       ├── test_attack_2_log_tamper.py    Merkle chain integrity
│       ├── test_attack_3_fake_sig.py      Unregistered and revoked keys
│       ├── test_attack_4_rollback_lie.py  Rollback denial and execution
│       ├── test_attack_5_replay.py        Duplicate receipt rejection
│       ├── test_attack_6_backdate.py      Timestamp and phase ordering
│       ├── test_edge_cases.py             Boundary and unusual inputs
│       ├── test_live_scenario.py          Replit scenario end-to-end
│       └── test_v11_surfaces.py           v1.1 new code surface regression
│
├── demo/
│   ├── demo_basic.py                      Minimal: sign, verify, print receipt
│   ├── demo_disaster.py                   Replit scenario with narrative output
│   └── demo_verify.py                     Third-party verify from receipt ID
│
├── paper/
│   ├── arc_paper.tex                      LaTeX source (8 pages, 38 references)
│   ├── arc_paper.pdf                      Compiled PDF (unpublished preprint)
│   └── compile.sh                         pdflatex compilation script
│
├── docs/                                  Extended documentation
│   ├── protocol_overview.md
│   ├── signing_specification.md
│   ├── integration_guide.md
│   ├── log_api.md
│   ├── compliance_map.md
│   └── red_team_report.md
│
└── website/
    └── index.html                         React SPA (no build step, CDN)

---

Research Paper

ARC: A Two-Phase Cryptographic Receipt Protocol for Verifiable AI Agent Execution Ramachandra Kulkarni, Harin Kumar Mallela, Arun Basavaraj Alur - April 2026 Unpublished preprint. Not yet peer-reviewed.

8 pages. Covers the action receipt primitive, seven JSON Schema definitions, the complete red team evaluation (6 holes found and fixed, 58/58 tests), and the live cross-agent verification proof on Windows 11 with Claude Code.

• LaTeX source: paper/arc_paper.tex
• Compiled PDF: paper/arc_paper.pdf (or download from Releases)
• Website: ramachandrakulkarni.github.io/arc-protocol (GitHub Pages, auto-deployed)

Document provenance (hardcoded in source, immutable without a git commit):

Research inception	March 2026
Paper created	April 02, 2026 00:00:00 UTC
Protocol version	v1.1

BibTeX:

@misc{kulkarni2026arc,
  title        = {{ARC}: A Two-Phase Cryptographic Receipt Protocol for
                  Verifiable {AI} Agent Execution},
  author       = {Kulkarni, Ramachandra and Mallela, Harin Kumar and
                  Alur, Arun Basavaraj},
  year         = {2026},
  month        = apr,
  howpublished = {Unpublished preprint},
  note         = {Protocol v1.1. 58/58 tests. \url{https://github.com/RamachandraKulkarni/arc-protocol}}
}

---

Authors

All three authors independently identified the AI agent observability trust gap. Ramachandra designed the protocol architecture, the seven-field canonical signing payload, the RFC 6962 Merkle log integration, and authored the Python reference implementation and red team test suite.

Author	Role	LinkedIn
Ramachandra Kulkarni	Protocol design, implementation, red team	linkedin.com/in/ramachandra-kulkarni26
Harin Kumar Mallela	Protocol analysis, adversarial modeling	linkedin.com/in/harin-kumar-m
Arun Basavaraj Alur	Protocol analysis, evaluation methodology	linkedin.com/in/arun-basavaraj-alur

---

Contributing

See CONTRIBUTING.md for setup, code style, and pull request requirements.

Security

See SECURITY.md for the vulnerability disclosure policy.

License

MIT. See LICENSE.