Glossary.

The load-bearing terms this site coins, defined plainly. Each definition names the concept and nothing more; where a term names a capability, its status is noted.

governed decision ✅ live

A decision that was routed across model failure families, criticized by rival models, and checked against executable ground truth before it shipped. It is the unit of work the whole site is scoped by: a status is only claimed for work that met this bar. One agent makes governed decisions today, on the lab's own operations.

failure family

A group of models that share the same underlying way of being wrong, because they share the same attention design. Members of one family tend to fail on the same inputs, so agreeing with each other proves little. Routing a decision across different families is what makes a second opinion an independent one.

cross-family critic ✅ live

A critic drawn from a different failure family than the builder, so it does not inherit the builder's blind spot. It is given the artifact in isolation and asked to refute it. A disagreement is resolved by a concrete refutation or by running the work, never by a vote.

executable ground truth ✅ live

A check that can be run, not a matter of opinion: a test, a computation, a real-world result. An outcome is accepted when it passes such a check. This is the truth that verification is measured against, which is why the site can say verified rather than agreed.

training triple ◷ coming

The record left by a governed decision when it is complete: the action taken, the outcome it produced, and the correction it earned. A triple is trainable data, not a log. Today the lab captures the action and the reasoning reliably but the outcome and correction legs are still being closed, so today's records are traces, not yet complete triples.

correlated failure

A failure that slips past every check at once, because the checks share a weakness. It is the rare case that offset defense is designed to make rare rather than to pretend away. The site names it plainly instead of claiming the defense is total.

offset defense ✅ live

Layering several verification mechanisms whose weak points sit in different places, so a hazard that slips one is caught by the next. The defense holds not because any single check is complete but because the gaps do not line up. Fig. 03 is the diagram of this idea; the residual path it shows is the correlated failure that survives when the gaps are forced to align.

specimen ◷ coming

A single agent in the fleet, presented as an entry in a field guide (marked SP-0X). One specimen is live, running the lab's own operations; the rest are designed and not yet running. The field-guide framing is a convention for reading the site, not a claim about any specimen's status; each one carries its own live or coming tag.

hazard paths (α, β, γ)

The three example hazards drawn approaching the panels in Fig. 03. Each is caught by a different panel because the panels' gaps sit at different heights: α is stopped at the second panel, γ at the fourth, and β at the first. β is the one that can be made to thread all four panels when the reader forces the gaps into alignment, so it stands in for the rare correlated failure.

The status taxonomy: ✅ live means running the lab's own operations today; ◷ coming means designed, not yet running. Terms without a tag name a pure concept, with no live-or-coming state of their own.