ABOUT

Reliability is a measurement, not a feeling.

FAULTLINE exists because production dashboards answer the wrong question. They tell you a service was available. They do not tell you whether the agent produced the correct user-visible outcome, at what cost, and how often it will again.

01 / MISSION

Engineered boundaries over probabilistic assumptions.

Autonomous software breaks down when engineering organizations confuse service uptime with agentic success. An agent can maintain 99.99% network availability while silently corrupting user workflows, spending uncapped retry budgets, or falling prey to hostile tool instructions.

FAULTLINE was started to establish deterministic, mathematically grounded verification for autonomous agents. We build runtime boundaries, pre-deployment release gates, and evaluation benches that treat reliability not as an aspirational sentiment, but as a repeatable, empirical invariant.

CORE OPERATING INVARIANTS
  1. [01]
    Outcome ownership — An agentic system is measured solely by the correctness of its completed outcome, not intermediate token fluency.
  2. [02]
    Experimental judgment — Every claim is anchored to paired, frozen-seed trials evaluated against deterministic oracles.
  3. [03]
    Systems thinking — Reliability is an end-to-end discipline spanning tool interfaces, sandboxes, and verification contracts.
  4. [04]
    Operational discipline — Hard resource caps, deterministic retries, and strict failure budgets prevent silent degradation.
  5. [05]
    Research integrity — All benchmarks, harnesses, and derivations are open source, reproducible, and bound to frozen contracts.

02 / PRINCIPLES

Six non-negotiable architecture invariants.
  1. [01]

    Correctness is independent of availability.

  2. [02]

    The evaluator is outside the policy.

  3. [03]

    Every recovery action spends a budget.

  4. [04]

    Comparisons share the experimental unit.

  5. [05]

    A fix must replay.

  6. [06]

    Claims fail closed.

03 / MAINTAINER

Built in public with reproducible benchmarks.
FAULTLINE mark

Samir Sawarkar

Maintainer · reliability engineering

Built FAULTLINE over 30 days of sequential modules and a phase-2 platform. Author of two research publications on pass^k reliability bounds and MCP tool poisoning defense. Maintained 684 deterministic tests across the evaluation suite.

GitHub

04 / TIMELINE · 2026

Sequence of delivered work.
  1. STAGE 01
    Phase 1 — 30 sequential modules, deterministic harness
  2. STAGE 02
    Publication 01 — pass^k reliability
  3. STAGE 03
    Publication 02 — MCP provenance contracts
  4. STAGE 04
    mcp-guard released, MIT
  5. STAGE 05
    faultgate v0.2.0 released, MIT
  6. STAGE 06
    faultgate v0.3.0 released: business rules, INCONCLUSIVE verdict, vendor trace import
  7. STAGE 07
    Reliability Workbench — in the workbench
OPEN INQUIRY

Questions, ideas, or a benchmark to break? Open an issue.

GitHub issues ↗Email