Pascaal builds AI components for safety and adoption.
Hallmark makes agents accountable. Plots puts them to work. Waterguard keeps them inside the lines. One thesis: agents should be identifiable, employable, and defensible.
Three components. One thesis.
Every product answers the same question from a different angle: can a human stay responsible while an agent does the work? Hallmark is shipping. Plots and Waterguard are in design, built on the same identity layer.
Hallmark
Spec v1Agent identity and accountability. A short-lived passport resolves every agent to one responsible human, with grants that name the action and budgets that cannot be shared. Read the protocol ↗
Plots
In designAutonomous work management. Agents that plan, execute and report real work — with the plan, the spend and every hand-off legible to the human who owns the outcome. See the direction ↗
Waterguard
ResearchAgent security models. Threat models, policy and containment for agents acting on real systems — what they may touch, what they may spend, and when they must stop. See the direction ↗
Intelligence is no longer the constraint. Attribution is.
Agents hold emails, burn API spend, write to workspaces and move money — and platforms still cannot answer the only question that matters when something goes wrong: which paying human is responsible for this action?
Blocking agents loses usage. Accepting them blind loses attribution and creates liability. The missing layer is not another detector or a better bot fingerprint — it is identity that resolves to a human, authorization that names the exact action, and provenance that survives the session.
Hallmark is that layer: Ed25519 with an ES256 fallback, RFC 8785 canonical JSON, short-lived passes, and budgets split — never shared — at spawn. The specification and its executable conformance suite are open.
Six components. One loop.
Identity that resolves. Authorization that names the action. Provenance you can show a user. Each component is useful alone; together they close the loop from first contact to a receipt.
Imprint — identity that resolves
A passport is a signed lineage from agent → parent → … → root attestation, bound to a human’s passkey. Verification is stateless and measured in microseconds.
Grants — authorization that names the action
Capability, budget slice, action, lifetime and jti in one token. Route auth succeeds only when a capability matches; budgets are disjoint at spawn, charged on observed usage.
Barricade — enforcement at the platform
Provider-side plugin that requires an imprint and a grant per surface. Enroll users at their first responsibility, not their first contact.
Freeroam — observation with zero friction
Cooperative agents self-identify; you get attribution before you require anything. Observe-only, no accounts, no interception — the wedge to enforcement.
Trackline — provenance users can read
Every action becomes a receipt: what the agent did, what it cost, which grant authorized it, and how much budget remains.
Seal v2 — ownership of original work
Ownership seals on original works (hallmark+work), enabling licensed use and payment instead of silent copying.
Our promise is identity that resolves: the kind you can prove in a receipt, not just assert in a policy.
Every number below is produced by code in the repository — npm test and npm run bench. If a number weakens, the suite fails.
| Measurement | Measured | Budget |
|---|---|---|
| Passport verification | 116µs | 5ms |
| Edge registry walk | 11.8ns/hop | — |
| Seal | SHA-256 | 1 hash |
| Algorithms | EdDSA / ES256 | allowlist |
Means from bench.mjs. p99 must be measured on your traffic before any SLO is promised.
“Where prose and code disagree, prose wins and this file is a bug.”
Hallmark runs where your traffic runs.
Verification is stateless and the passport carries no PII — only a lineage, a grant reference and public key material. Put it in front of the thing that already sees agent traffic.
In your gateway
One header check before routing. Reject, challenge, or allow with attribution — your call per route.
At the edge
Stateless verification plus registry lookups; no per-request call back to a control plane.
In your VPC
Self-host the verifier and the registry anchor. The passport format does not care where it is checked.
Designed to fail closed.
The passport is the only thing crossing the trust boundary, so it is small, canonical, and boring on purpose. Everything that could be abused is either forbidden by the format or pinned by the registry.
Algorithm allowlist. [EdDSA, ES256] — nothing else is accepted, ever.
Canonical JSON only. RFC 8785 JCS; duplicate keys and non-canonical encodings are rejected.
No key discovery by URL. jku, x5u and x5c are forbidden; keys come from the registry.
Replay suppressed. A jti cache plus a sequence window; reuse is a hard error, not a warning.
Revocation is per capability. Stale revocation data fails closed for that capability only.
Short lifetimes by construction. TTL ≤ 120 s, hard cap 300 s — containment is a constant, not a promise.
Autonomous work, legibly owned.
Plots is our work-management component. Agents plan, execute and report real work, while the human who owns the outcome can see the plan, the spend and every hand-off. It is not a task board with an agent bolted on; it is a work loop where accountability is the default.
Plan. Decompose an objective into steps a human can read, approve and re-scope before anything runs.
Execute. Run each step under a Hallmark grant, so scope and spend are bounded by construction.
Report. Close the loop with a receipt the owner can audit — what ran, what it cost, what changed.
Agent security, as a model.
Waterguard is our security-model component. Not a scanner and not a prompt filter: a way to name the assets, the actions and the failure modes of an agent that touches real systems, then enforce the answer where the action actually happens.
Model. Name the assets, the actions and the failure modes before an agent is deployed.
Enforce. Policy that fails closed at the gateway — not in a prompt, and not in a hope.
Contain. A bounded window between detection and stop, backed by short-lived credentials.
Design notes, in the open.
Start with Hallmark. The rest plugs into the same identity.
If your platform already sees agent traffic, start with Freeroam: observe-only, zero accounts, no interception. Add Barricade when you are ready to enforce — and Plots and Waterguard inherit the same lineage.