Skip to the notebook
OpenExecutionPROJECTAEGIS

AEGIS: from records to evidence

An interactive protocol study

AEGIS packages recorded actions into a signed bundle that a recipient can verify with a separately obtained public key.

What changes when a record is altered?

We’ll build a chain of three synthetic events, sign its certificate, then change one event. Run the verifier to compare the original evidence with the altered version.

Protocol construction

H₀ binds the chain ID, agent ID and creation time. Each event includes its predecessor hash.

SHA-256 hashes the framed records. Ed25519 signs JCS(certificate).

The certificate binds the identity, origin, length and ending of the recorded sequence.

Inspect the exact byte contract

H is SHA-256. JCS is the canonical JSON encoding of this v1 profile. frame uses a domain tag, named fields and length prefixes; it is not plain string concatenation.

H₀ = H(frame("genesis", chain_id, agent_id, created_at))
Hᵢ = H(frame("event", seq_number, event_type, timestamp, payload_hash, Hᵢ₋₁))
signature = Ed25519.Sign(private_key, JCS(certificate))
Read the full v1 specification

Try it

The sample records a proposal, approval and deployment. Edit an event type, then verify; the original hashes and certificate stay fixed.

#event_typepayload_hashprev_hashevent_hash
Preparing the evidence…

Synthetic records · real SHA-256 and Ed25519 checks · hash prefixes shown.

verifier> verifyBundle(bundle, externalPublicKey)

READY — run verification to inspect the current evidence

More experiments, exact evidence and downloads

The demonstration public key is generated here and held separately from the editable bundle. In an audit, the recipient obtains its verification key through an independently authenticated channel.

Reference implementation

Explore the code that constructs the same objects and applies the same acceptance rule.

Reproduce the experiment locally

Download the reference archive and export the bundle and public key from the experiment. In the extracted reference directory, run:

node src/verify.js aegis-bundle.json --public-key aegis-public-key.pem

Code and replication archive · Python verifier

Reference implementation 0.2.3 · AEGIS v1. The browser adapter follows its framing and V1–V4 rule, with stricter key import: one canonical Ed25519 SPKI PEM block. Browser exports and mutation examples agree with this pinned public reference.

Where this experiment sits in AEGIS

The architecture connects capture (ARC1), portable commitment (ARC2), outside verification (ARC3), independent custody (ARC4), auditor-owned procedures and reports (ARC5), and forum judgment (ARC6). This notebook exercises the released ARC2–ARC3 profile with synthetic input.

Capture completeness needs mediation and independent reconciliation; a received bundle alone cannot reveal an event missing from the captured record. A valid receipt reports technical acceptance of the received bytes, while audit conclusions also use scope, identity, custody and procedure evidence.

AEGIS is one project within the wider research programme. Programme overview

AEGIS implementations

From research interfaces to working code

AEGIS reference code and downstream research implementations. Each entry links to the code and the evidence of its connection to the protocol.

DOWNSTREAM RESEARCH IMPLEMENTATION

DFAH-Bench: replay-log provenance

IBM Client Engineering · Python · SHA-256 / Ed25519

DFAH-Bench’s provenance pipeline groups financial-agent benchmark replay logs into hash chains, signs a certificate for each chain, then verifies and exports the resulting bundles. Its chain and certificate modules explicitly name AEGIS as an influence and link the reference implementation.

  1. 01Replay logs
  2. 02Hash chain
  3. 03Signed certificate
  4. 04Verified bundle

This entry documents explicitly attributed engineering reuse in an independently maintained research implementation. Its serialization and verification profile are defined by that project; byte-level AEGIS v1 conformance is a separate question.

Attribution and source evidence

Both module headers contain the following attribution:

Inspired by the AEGIS protocol (Li, 2026)

Pinned commit · · 066be0f48a37

Source checked on 6 September 2026. The links above preserve the inspected revision.

AEGIS reference experiment

The Protocol tab explains the AEGIS v1 construction and runs a browser-local evidence experiment. Use it to understand the commitment and verification interface before exploring downstream variations.

Posts & discussions

Writing and conversations around AEGIS

Selected public writing and conversations on AI execution, evidence and accountability. Read a short introduction here, then follow the link to the original context.

PUBLIC DISCUSSION · LINKEDIN COMMENT

AEGIS and DFAH-Bench: research meeting implementation

Raffi Khatchadourian · Verified 6 September 2026

In a public comment mentioning Alex Chengyu Li, Raffi acknowledges AEGIS’s influence on DFAH-Bench’s replay-log provenance modules. He also notes that the revised AEGIS paper cites DFAH-Bench as an example of downstream engineering reuse.

his AEGIS research helped inform DFAH-Bench’s replay-log provenance modules
AUTHOR ARTICLE · LINKEDIN

Who Audits the AI Agent?

Alex Chengyu Li ·

An early practitioner-facing explanation of the AI-agent accountability question. The article discusses tamper-evident records, third-party verification and execution provenance, and introduces AEGIS.

The article records the project’s earlier framing. The current protocol definitions and reproducible experiment are maintained in the Protocol tab.

Original posts and comments remain on their authors’ platforms. LinkedIn may require sign-in to display the full discussion.

Ethical reflections

What changes when execution becomes accountable?

Evidence infrastructure changes incentives as well as technical capabilities. This strand examines who benefits from recordability, how work is allocated, and which institutions can turn records into accountable judgment.

AI and EthicsAccepted

Accountability Arbitrage: Ethical Tensions in AI Agent Accountability Infrastructure

Alex Chengyu Li · AI ethics · institutional accountability

A human and an AI agent can perform the same task equally well while producing records of different institutional value. The paper studies accountability arbitrage: an expected advantage in reviewable evidence can reverse the work choice that performance and ordinary cost alone would favour.

The same record can support a compliance claim and make error or misconduct easier to establish. That dual effect can also encourage accountability avoidance. The analysis connects these incentives to evidentiary parity, performance standards, review authority and proportionate data collection.

Accepted for publication in AI and Ethics on 1 September 2026. The reading link above is the public manuscript record; a publisher Version of Record has not yet been added to this catalogue.

Questions raised by this project

Who gains an evidentiary advantage?

Compare equivalent human and agent workflows when their records are valued differently.

Does a better record mean better conduct?

Examine when recordability changes task allocation while substantive performance stays fixed.

Who has the authority to judge?

Connect preserved evidence to review procedures, empowered forums and meaningful challenge.