Bolt PCHIE releases deterministic, tamper‑evident safety harness for AI agents

What is Bolt PCHIE and why was it created?

Bolt PCHIE is an open‑source safety harness for autonomous AI agents that records each decision the agent makes in a cryptographically linked hash chain. The author, posting on Reddit’s MachineLearning community, described the project as a way to "dispell all the hoopla and vibes" by providing a verifiable receipt for every action an agent takes. By making runs deterministic and auditable, the harness aims to move trust from a black‑box assumption to a reproducible proof that can be inspected after execution.

How does the deterministic hash‑chain mechanism work?

According to the Reddit post, the harness inserts a lightweight hashing step after each inference or policy call. The output, together with a timestamp and any relevant context, is fed into a hash function; the resulting digest is then combined with the previous digest to form a chain similar to a blockchain ledger. Because each link depends on the prior one, any alteration to an earlier decision would break the chain’s integrity, making tampering evident. The author emphasizes that the process is deterministic: given the same inputs and model state, the chain will always produce the same sequence of hashes, allowing exact replay of a run.

What benefits does tamper‑evident logging bring to AI agent safety?

The primary advantage is verifiability. Researchers can replay a recorded run, recompute the hash chain, and confirm that the agent behaved exactly as logged. This mitigates concerns about hidden or post‑hoc modifications to an agent’s behavior, a common criticism in safety‑critical deployments such as autonomous vehicles or financial trading bots. Additionally, the harness stores episodic memory – the agent’s short‑term observations – alongside the hash chain, enabling auditors to cross‑check what the agent claimed to have seen against the immutable log. The author notes that the system provides "receipts" for those who demand proof rather than trust.

How can developers and researchers try Bolt PCHIE?

The code is publicly available on GitHub at https://github.com/wetwaferworks/bolt-pch-ie. To experiment, users can clone the repository and integrate the provided BoltHarness class into any Python‑based agent framework. The harness exposes a simple API:

from bolt_pch_ie import BoltHarness
harness = BoltHarness()
action = agent.decide(state)
harness.record(action, state)

Developers can then call harness.verify() after a run to ensure the chain is intact. The repository includes a minimal example using a reinforcement‑learning loop, which can be a starting point for custom agents. For those who prefer a visual inspection of the decision history, the Grafy Graph app can be used to plot the hash chain alongside agent metrics, turning the raw log into an interactive branching diagram. See the Grafy Graph page for more details (/graph). If you need to modify or extend the harness code, the Grafy Code environment offers an in‑browser IDE with instant execution, making it easy to prototype changes without leaving the platform (/code).

Where does Bolt PCHIE fit within the Grafn ecosystem?

While Bolt PCHIE itself is a standalone library, its design aligns closely with Grafy’s Graph tool, which visualizes prompt flows, model outputs, and edit histories. By importing a run’s hash chain into Graph, creators can trace exactly how a generative model arrived at a particular output, a capability that complements the platform’s existing focus on reproducibility. Researchers can also leverage Reader to ingest the repository’s README and source files, asking natural‑language questions about implementation details without leaving the chat interface (/reader). In short, Bolt PCHIE provides the low‑level audit layer; Grafy supplies the high‑level UI and analysis layer that makes those audits actionable.

Frequently asked questions

Does Bolt PCHIE guarantee that an AI agent cannot be altered during execution?

No. The harness records actions deterministically, but it does not prevent an external attacker from modifying the agent’s code or environment. It only makes any such modification evident after the fact by breaking the hash chain.

Can the hash chain be used to replay an agent’s exact runtime environment?

The chain captures decisions and the associated context, enabling a logical replay of the agent’s behavior. However, it does not store the full system state (e.g., GPU memory), so a perfect binary replay would require additional snapshot tools.

Is the harness compatible with all model providers and frameworks?

Bolt PCHIE is written in Python and designed to be framework‑agnostic, but integration still requires inserting the record call at each decision point. Compatibility with specific libraries (e.g., OpenAI, Vertex AI) depends on the developer’s willingness to add the hook.

Source: Reddit ML (https://www.reddit.com/r/MachineLearning/comments/1xvyvbd/p_bolt_pchie_a_deterministic_tamperevident_safety/)

---

Source: Reddit ML — [P] Bolt PCHIE: a deterministic, tamper-evident safety harness for AI agents [P]