Boundary tests, ties, losses, and supporting diagnostics.

The flagship evidence tests the core enforcement surfaces. This archive preserves the harder questions: where authority is absent, where simpler controls tie, where an answer cannot be attributed to its source, and where an internal stress test should remain an internal stress test.

FieldHash is architecturally independent of the systems it governs. Its current public diagnostics are self-administered; third-party validation is a separate milestone.

Boundary research

Keep authority intact as agents work together

An authored, self-administered synthetic multi-agent study tests whether organizational authority remains binding while useful work completes. Full mediation safely completed 84/84 main workflows versus 75/84 under strong per-action authority; all seven known comparator violations exceeded the shared budget. A separate larger repeat-protection follow-up retained 157 matched mechanism witnesses across Kimi and Terra; neither prespecified study-wide criterion was met. Its 720 scheduled workflows and retained failures remain separate from the main study.

Read study

The agent found another channel. The authority held.

A live execution-surface study retains the failure created by a planted unmediated path, then tests one authority boundary across eight enumerated synthetic enforcement points.

Read study

The agents changed course. The authority held.

A live closed-loop flagship tests whether configured effect authority remains binding after Kimi and Terra observe denial, change routes, and continue.

Read study

A changed route does not create authority

A sealed matched-counterfactual study holds model choices constant across prompt-only, exact-action, and FieldHash effect-authority control. It is mechanism evidence, not a live adaptive-agent trial.

Read study

Your agent can choose a version you revoked

A PyPI-derived version-selection diagnostic: a prompt missing the relevant withdrawal chose the withdrawn version in 540 of 540 trap evaluations, against zero under Governed Tool Use. With the complete, current authority log, prompting also reached zero. No packages were installed.

Read study

Where simpler controls are sufficient

Pre-registered selection arms, retrieval-scale pressure, and the boundary between selection quality and governed accounting.

Read study

Corpus authority series

Public-record falsification, recovery, and hidden-authority tests, including where metadata-aware filters tie.

Read study

Every answer receives a disposition

Support, contradiction, and unattributable accounting across live answers.

Read study

Supporting diagnostics

These pages remain public for technical scrutiny. Internal synthetic results are labeled as such and should not be read as customer validation or production safety rates.