PipeRoll - Agent Incident Registry

Verified public records of AI-agent failures: what the agent controlled, what went wrong, what it cost, and what the evidence is. Named for the Pipe Rolls - the English Exchequer's great rolls, 676 unbroken years of audited records kept tamper-evident by a parallel copy in different hands.

Live site: https://piperoll.org

Constitution

The registry's binding rules are published as CONSTITUTION.md (rendered at piperoll.org/constitution). Records cite rules by number; rules are amended by PR and never renumbered.

Principles

• Every record is verified individually against primary sources before publication.
• Corrections are published in the record, never slipped.
• Rejected candidates retire their reserved ids permanently; ids are never reused (CVE convention).
• Ids are permanent opaque names. The initial import (PIR-2026-0001..0045) was numbered by occurrence date as a one-time property; later ids are assigned at registration.
• Near-misses are records too - full exposure with zero realized loss is actuarially precious.
• Sources are archived to the Internet Archive's Wayback Machine at registration (tools/archive.py, run by CI on merge; backfill by dispatch). The Archive's own timestamps are the independent witness of what a source said when it was cited.

Completeness disclosure

This registry records what is publicly reported and independently verifiable - a fraction of what occurs. Most agent failures are never disclosed. Coverage is biased toward the visible: on-chain losses, court and regulator records, published research, viral postmortems, and English-language sources; crypto-agent incidents are overrepresented relative to silent enterprise failures for exactly this reason. The initial 45 records were gathered by structured research sweeps (Aug 2026), not systematic sampling. Consequences: registry counts are a floor, never a frequency estimate; absence of a system from the registry is not evidence of its safety; and no failure rate can be computed from this data alone, because the exposure base - how many agents run, with what authority - is unknown. Correcting that denominator problem is part of the institution's roadmap, not a property of this dataset.

Layout

• incidents/ - the records (one markdown file per PIR id) + INDEX.md
• incident-schema-v0.md - the schema, including its amendment history (v0.1, v0.2)
• tools/build.py - static site generator: python3 tools/build.py regenerates docs/
• tools/linkcheck.py - fetch-grade liveness check for every source URL
• tools/archive.py - submit every source URL to the Wayback Machine (Save Page Now)
• the site is a build product, not a committed copy: python3 tools/build.py renders incidents/*.md into docs/ locally (gitignored); CI builds and deploys it to Pages on every merge, exports included

Maintainer

Srinivas Gumdelli - founding editor. Registration authority currently rests with the editor (see CONTRIBUTING.md); conflicts of interest are disclosed inside the affected records, per the registry's constitution.

Licensing

Records and registry data: CC BY 4.0 (cite PipeRoll and the PIR id). Tooling: MIT. The incident schema is open and will remain open.

Contributing

Open data, open submissions, verified registration: anyone may submit a record via PR, CI enforces the schema and checks every source link, verification is editorial, and merge is the act of registration. See CONTRIBUTING.md and incidents/TEMPLATE.md. Merge authority stays with the editors - that split (Wikipedia intake, CVE authority) is constitutional.