AIRLOCK-IN — the airlock-in/out constellation (judgment-grade, by hand)

Reflexive note. The airlock does not yet exist; the discipline does. This is the first application of the entry membrane — applied, fittingly, to the act of building the entry membrane. It models the shape the Phase-3 HATCH ADR will mechanize. Operator-attested go/no-go is the Gate; this document is the seam-map and volume declaration that Gate reviews.

Mode: Design (authoring/evaluating intent — Constitution → PRD → ADR → OBPI → REQ → TASK; campaign §3). Authority: Operator-ratified design dialogue 2026-06-30 + Magna Carta Movement III. The campaign governs sequencing. Source-ranking: L1 canon (pool ADRs, work-phases-and-airlock.md) > L2 ledger (this session's agent-insights.jsonl records) > L3 derived views. External refs (graphify / LightRAG / Plumb / BEADS / Symphony) are bones, never authority. Topology-purpose: a sequential keel-up constellation build (one parallel-early node: corpus). Purpose: realize the §8 "work-phase theories lawful" 1.0 gate. Each new harness mechanism (monitor, graph engine, membrane) declares its own function · topology · verifier · termination (the absorbed loop/topology discipline).

The move

Promote/author four feature ADRs (pool→feature) realizing the airlock-in/out system: KEEL (obpi-state-machine) → HULL (one graph substrate superseding artifact-graph-navigation + execution-memory-graph + covers-source-anchors) → HATCH (new airlock membrane) → RECALL (rag-anything, deferred).

Seam-map — footprint (breadth)

PUSH edges — what this move may break (blast radius, ↓ into fact)

Surface	Why it is in the blast radius
3 superseded pool ADRs	Anything referencing artifact-graph-navigation / execution-memory-graph / covers-source-anchors must be repointed when they are superseded by the HULL ADR.
Every governance read/write path	The KEEL's runtime invariant monitor sits on the artifact-graph read/write boundary — it touches every gz command that mutates artifact state. Largest push surface in the constellation.
ADR-0.0.16, ADR-0.0.24	Subsumed by the KEEL (coherence guard → monitor rule 4; receipt-binding → attested transition witness). Their behavior must be preserved through the subsumption.
~30 audits / reconcilers	The KEEL retires the choreography; each retired reconciler is a behavior that must be re-homed as a transition, not dropped.
gz adr promote + taxonomy schema	pool→feature promotion runs while the foundation enum still exists in schema (mechanical abolition is Movement IV). Promote with --kind feature; do not assume the enum is gone.
ADR-0.20.0-spec-triangle-sync	Coupled to Plumb's spec↔test↔code reconciliation — the HATCH's airlock-out reuses Plumb-bones; keep the two coherent.
work-phases-and-airlock.md / four-phases-of-work.md	Promoted from "Draft North Star" to lawful by the HATCH ADR — their final terminology becomes binding.
This campaign file	Edited in the same move (Movement III rewrite + §3a pivot).

PULL edges — what binds this move (constraints, ↑ into intent)

Constraint	Binding
Arch-Boundary §12.3	"Do not build the graph engine without locking state doctrine first." → KEEL (Phase 1) MUST precede HULL (Phase 2). The monitor is the lock.
Arch-Boundary §12.1 / §12.6	No post-1.0 pool promoted into active work (ours are pre-1.0, OK); derived views never source-of-truth (the Tier-B graph cache + L3 recall must never gate).
§5 enforcement-claim rule	The HATCH's "refuses GO on un-accounted seam" claim REQUIRES a live negative control; no NC ⇒ facade ⇒ rejected.
STDLIB-FIRST	tree-sitter + networkx is a departure — must be attested in the HULL feature ADR (named: deterministic multi-surface extraction + topo-sort/cycle-detection stdlib cannot supply).
state-doctrine Rule 5	L3 never fail-closes a gate. The fence that makes graph.json (Tier-B) and the recall tier (L3) legal assets.
OBPI↔ADR 1:1 mandate	Every OBPI traces to a parent feature ADR's Feature Checklist. No headless OBPIs.
pool→feature taxonomy (§3a)	foundation abolished; all four promote/author as feature, earning back to the release line one at a time with executable proof.

Volume declaration

• Footprint (breadth): wide — the KEEL alone touches the whole governance surface. Accounted above; the supersession + monitor edges are the heavy ones.
• Reach (depth / tracer): the FIRST end-to-end slice is not the full 8-property state machine. It is the airlock-critical tracer: State/Transition Pydantic models → withdraw/supersede transition → runtime monitor → CLI verb → ledger event, piercing schema → model → monitor → CLI → ledger end-to-end. One working keel→hull-corpus→hatch-resense path before breadth expansion (tracer-bullet discipline; the deferred-in-keel properties wait).

Pre-registered falsifier(s)

1. Landing falsifier (keystone). If the KEEL monitor does not refuse a silent status: frontmatter drift (the GHI #348 class) when run in production config, the keystone is unbuilt → re-sense is theater → NO-GO on Phase 2 until fixed.
2. Preservation falsifier (blast radius). If superseding the three pool ADRs breaks any reference (gz validate --documents / --cli-alignment non-green) → unaccounted push edge → block and repair before promotion.
3. Sequence falsifier (§12.3). If HULL (graph) is attempted before KEEL locks state doctrine → boundary violated → the graph re-senses drift-prone nodes → reject the sequence jump.

Go / No-Go

Recommendation: GO, keel-up, with two gates:

• Phase 0 (this record) gates Phase 1 — operator attestation required before any ADR is authored or promoted.
• Phase 1's landing falsifier (monitor refuses silent drift, live) gates Phase 2.

Phase 4 (RECALL) is severable and explicitly deferred past the first working airlock. The §5 live NC is the binding teeth on the HATCH's GO-refusal claim.

Deferred / future option — refactored graspologic

graspologic (Microsoft/NeuroData graph-statistics: spectral embedding, community detection, graph matching) was stripped from the hull floor — unused for seam-queries and not installable on Python 3.13+ (graphify #290). Kept as a possible future, not discarded: a 3.13-compatible refactor (or a thin extraction of its spectral/community algorithms over the seam graph) could power advanced graph analytics the deterministic floor cannot — "god-node" detection (over-coupled artifacts), community structure (latent module boundaries), surprising-connection surfacing. A post-airlock enrichment only; L3-advisory, never gating (state-doctrine Rule 5), peer to the RECALL tier. Parked here as a named option so the door stays open.

---

Status: Phase 0 EXECUTED — GO attested; Phase 0 → Phase 1 gate now OPEN (2026-07-02). Operator authorization (verbatim): "take on Movement III Phase 0" / "this: Movement III Phase 0 — airlock-in, operator-gated." The go/no-go recommendation below (GO, keel-up) is ratified.

The seam-map was re-sensed against current repo truth before recording GO (judgment-grade re-verification, 2026-07-02) — no blast-radius shift since the 2026-06-30 authoring: all five constellation pool ADRs present (obpi-state-machine, artifact-graph-navigation, execution-memory-graph, covers-source-anchors, rag-anything-governance-retrieval); the foundation enum is still live in src/gzkit/schemas/adr.json:36 (confirming the seam-map directive to promote --kind feature, not assume abolition); the subsumption targets ADR-0.0.16 / ADR-0.0.24 and the Plumb-coupled ADR-0.20.0 all exist; work-phases-and-airlock.md and four-phases-of-work.md present.

Boundary held — nothing authored or promoted in this Phase 0 action. Opening this gate authorizes Phase 1 (KEEL obpi-state-machine promotion) to begin; it does not itself begin it. Per the operator's standing distinction (2026-06-30: "if go means start work, I am not ready to") and Behavior Rule Always #17, Phase 1 ADR authoring/promotion awaits an explicit go-to-work directive. The Phase 1 → Phase 2 gate remains the landing falsifier (monitor refuses silent status: drift, live).

This record stands as the pre-flight; the seam-map, volume, and falsifiers are laid out for audit and reuse as the Phase-3 HATCH checklist. Provenance: design dialogue 2026-06-30; ratifications + dependency weighing in .gzkit/insights/agent-insights.jsonl (2026-06-30 records); Phase 0 GO record (2026-07-02) in the same insights channel.