External Proving Ground Note — 2026-06-24¶

Status: Operator-supplied strategy synthesis. Captured from the operator design dialogue of 2026-06-24. It is not a new campaign, not an ADR, and not an authority layer — a direction to align the apparatus to, deferred for later routing. Natural home when picked up: a design note that may seed a pool-ADR candidate.

Thesis¶

gzkit cannot validate its own theories — the airlock / seam-map discipline, reasoning-mode induction, "governed loop engineering" — by dogfooding on itself. gzkit-governing-gzkit is a maximally self-referential, confounded test surface: it exercises governance-meta seams, never the real application seams (imports, call-sites, product contracts) the theories actually make claims about. Self-dogfooding therefore cannot show whether a theory generalizes, and it can get messy.

An external project is the live negative control for gzkit's strategic bet itself. The North Star claim is "governed loop engineering improves agent-driven work." If the only loop gzkit ever governs is its own, that claim is unfalsifiable — meta-level vibing, the exact failure class the apparatus exists to make inert. A proving ground is what turns "gzkit helps" into a claim a negative control can refute.

The missing forcing function¶

"Wait-and-see" caution — e.g. deferring tool-computed AIRLOCK-IN until the graph engine lands — is only legitimate if there is a mechanism to actually see. Today gzkit has no forcing function, no trial-run surface, no scouting loop, so that caution is currently indistinguishable from plain deferral. The Loop Leverage Test asks "does it enable faster failure before state corrupts?" — you cannot fail fast with no external surface to fail on. A skunk-works problem set is the prerequisite that makes both the caution and the theories falsifiable.

Candidate selection — rhea¶

Three candidate proving grounds exist, all at github.com/tvproductions (exact repo slugs to confirm):

repo	shape	fit
rhea	small repository pattern library	best fit — small, bounded scope = clean trial surface
pueo	more ambitious "afk" (away-from-keyboard) agent framework	too large/ambitious for a controlled first trial
gzfactor	more ambitious "afk" agent framework	too large/ambitious for a controlled first trial

rhea has been used as a test target, but only briefly. Its small, pattern-library scope is the reason it is the right first proving ground: fewer confounds, real-but-bounded application seams.

Mechanism — genesis-forward, lateral branches, gzkit bootstrapping¶

The affordance the operator named: take rhea back to its genesis document — the earliest commit, when it is "just its original design doc" — and from that clean start:

start lateral branches off the genesis document, each a discovery probe;
allow gzkit bootstrapping into each branch (init the apparatus from scratch on a real, non-gzkit project);
run the work forward through gzkit's full apparatus, exercising design → build → fix → refactor on real product seams.

This is a controlled from-clean trial: known starting state, real seams, gzkit as the instrument rather than the subject.

Relationship to current work¶

This is upstream of the AIRLOCK-IN sequencing question (judgment-grade vs tool-computed, and the time-vs-contract-adjacency tradeoff): a proving ground may be the thing that resolves that fork by letting the theories be scouted under real use, rather than a peer to it. It touches the Build-to-1.0 campaign (Magna Carta) because it concerns how gzkit validates anything at all — so any adoption is operator-ratified, not agent-initiated.

Explicitly deferred — not yet routed to any ADR, OBPI, or GHI.

Provenance¶

Captured from the operator design dialogue, 2026-06-24. Operator anchors:

gzkit needs a "problem set" within which it can perform "skunk works" validation and discovery.
applying gzkit outside its own dogfooding — which can get messy — is needed; otherwise we have no places to test our theories.
rhea is the right scope; start lateral branches with just the "genesis" document and allow for gzkit bootstrapping.
rhea is a small repository pattern library; pueo and gzfactor are more ambitious "afk" agent frameworks; rhea is likely the best fit.