Accountability Interface
FEATURES

Eleven features, three readers.

The interface serves three distinct readers — the operator deciding, the regulator auditing, and the engineer wiring a new agent system in. Each group's needs are surfaced explicitly rather than hidden in a single feature list.

FOR THE OPERATOR

For the operator

01

Single inbox across all agent systems

TX-1, SS-1, Swarm Lite, and future systems submit through one schema. The operator sees one queue, sorted by urgency, with decision-type accent strips telegraphing classification before they read a word.

Inbox showing 7 proposals from three agent systems
02

The gate prevents rubber-stamping

Approve is disabled on every consequential card until the Strategic Context block has been opened. A subtle interaction cost, applied only where stakes are high.

Card with Approve disabled until context reviewed
03

Modify, don't reject

Override Panel surfaces alongside the card with the original action visible. The operator changes specific values, captures rationale, acknowledges conflicts — both the original proposal and the executed modification are preserved.

Override panel docked beside the proposal card
04

Three actions, equal weight

Approve, Override, and Dismiss are visually identical. Per ADR-006, no action is larger, more prominent, or more coloured than the others. Approval is not the default.

FOR THE REGULATOR

For the regulator

01

Immutable brief snapshot

The exact proposal shown to the approver is captured verbatim. Six years later, an auditor opens the record and reads the same page. Not a reconstruction. The bytes.

Forensic decision record rendering the brief as presented
02

SHA-256 sealed records

Every decision record is hashed over its sealed fields (reasoning + action + governance) at commit time. Any modification invalidates the record. The hash is rendered on the document — an auditor can verify independently.

03

EU AI Act articles satisfied at decision time

Article 12 (logging), Article 13 (transparency), Article 14 (human oversight) — captured-brief + captured-rationale + gate-mechanic + sealed-record satisfies all three, provably, at the moment of decision.

04

Manifest-versioned decisions

Every record carries the manifest version it was made against. When rules change, decisions made under the old rules stay attached to the old rules. No retroactive reinterpretation.

FOR THE ENGINEER

For the engineer

01

One schema for every agent system

Proposal and DecisionRecord interfaces are stable across TX-1, SS-1, Swarm Lite, and any future system. Adding a new system reduces to: serialize → POST /api/proposals.

Architectural diagram showing TX-1 native card, JSON bridge, and Accountability Interface card
02

FastAPI sidecar with SQLite

Tiny portable backend. Boots in 250ms. POST proposals in, GET them out, commit decisions back to source systems via callback URLs. SQLite for Phase 3 — swappable for Postgres in a five-line change.

03

IBM Carbon throughout

Pure CSS using Carbon design tokens. No Tailwind. No rounded corners on data surfaces. IBM Plex Sans + Plex Mono. Engineering can drop in real Carbon React components — they just work.

See it live.

Eleven features sound abstract. Five minutes in the live demo makes them concrete. The Chicago Bottleneck scenario is pre-loaded.

Open the demo →