Agent Loss Record (ALR)

For insurers, AI-safety teams, and researchers who need a shared, structured record of AI-agent failures instead of scattered news coverage.

The Agent Loss Record is an open schema for writing down an AI-agent failure the way insurers write down any other loss: exposure, controls, the event, severity, and dependencies. It ships with a benchmark of 18 real, publicly reported incidents forced into that shape, plus five worked in full YAML. The important caveat: this corpus is built from public reporting, so severity and frequency are modeled and estimated, not measured against real paid claims.

Status: early, solo-maintained, v0.1 draft. The schema and the 18-incident benchmark exist and are internally consistent. No real claims data has been pooled yet, so every severity and frequency figure in this repo is modeled or estimated, never measured from a paid claim.

---

Quickstart

There is no CLI here: ALR is a schema and a corpus, not a tool. The fastest way in is to clone it and look at what is actually there.

$ git clone https://github.com/kylemaps/agent-loss-record.git
$ cd agent-loss-record
$ wc -l schema/agent-loss-record-schema.md benchmark/ai-agent-loss-benchmark.md benchmark/seed-corpus.md
  251 schema/agent-loss-record-schema.md
   79 benchmark/ai-agent-loss-benchmark.md
  176 benchmark/seed-corpus.md
  506 total
$ grep -c "^| [0-9]" benchmark/ai-agent-loss-benchmark.md
18
$ grep -c "^## Row" benchmark/seed-corpus.md
5

schema/agent-loss-record-schema.md defines the format. benchmark/ai-agent-loss-benchmark.md structures 18 real incidents into it. benchmark/seed-corpus.md works 5 of those incidents in full YAML, field by field.

---

How it works

Every incident, real or synthetic, gets forced into the same shape:

LOSS COST = FREQUENCY x SEVERITY, over an EXPOSURE base, adjusted by CONTROLS

A record is five linked objects: an exposure unit (the deployment), a control profile (what is on, off, or partial), a loss event (what happened and how it was triggered), a severity band (S0-S5, by realized or potential dollar loss), and a dependency map (shared model, framework, or vendor implies correlated loss across otherwise-unrelated companies). Every event is also tagged to OWASP-ASI, MITRE ATLAS, and AIUC-1, so a record slots into taxonomies that already exist instead of inventing a new one.

---

Architecture

public incident / red-team finding / real claim
        |
        v
  EXPOSURE UNIT ---- has a ----> CONTROL PROFILE
        |                              |
        `----------- experiences ------'
                       |
                       v
                 LOSS EVENT
                       |
                       v
          SEVERITY & BAND  (S0-S5, the unit of account)
                       |
                       v
       DEPENDENCY MAP  (shared model/tool/vendor -> correlated loss)

  wrapped throughout by: PROVENANCE + INTEGRITY ENVELOPE
  (source_type, content_hash, ed25519 signature, timestamp proof)

Two feeders populate ALR-shaped records today: Rating Bridge derives exposure metrics and incident stubs from robot/AV telemetry (MCAP, sealable with veriseal); Trace Bridge does the same from software-agent execution traces (OpenTelemetry GenAI / MCP tool-call logs). Different capture worlds feed one loss-record standard.

---

What it does NOT prove

• No verified dollar loss for any incident: public reporting supports modeled or estimated severity, never measured, except where a court or regulator already published a figure (4 of 18 rows).
• No exposure denominator: decisions or calls at risk are unknown for all 18 benchmark incidents, because that number only comes from a real deployment review.
• No real frequency: 18 anecdotes describe failure modes, not how often they occur per exposure.
• Not a pricing tool: it is a proposed common shape for loss data, built so real claims data can be poured into it later.

---

Status / Roadmap

v0.1, open draft, solo-maintained. Built: the schema (five linked objects, an autonomy scale, a control catalog, severity bands), an 18-incident public benchmark, and 5 incidents worked in full YAML. Not built yet: any pooled real claims data, a calibrated grade scale (AL1-AL9), or automated schema validation.

The specific open questions the schema leaves to humans are listed in schema/agent-loss-record-schema.md, section 10.

Agent Loss Record is one of a small family of standalone tools (veriseal, rating-bridge, trace-bridge); each stands on its own.

---

Contributing

Contributions of public incidents structured into the ALR schema are welcome: open an issue or a pull request. See CONTRIBUTING.md for the process and sourcing requirements.

---

License

Apache-2.0 (schema and code). Benchmark data may also be reused under CC-BY-4.0.

Maintained by Kyle Mapue.