ADR 0031: Target-Architecture Composition — the Seam Table + Trajectory Layer

The composition record for the whole organism: how the 30 slice-ADRs compose into the canonical Two-Hemisphere scaffold (Body /src/ ↔ Brain /ai/, joined by the Neural Link — per ADR 0018 OD-3), which decision owns which seam, and the trajectory invariants that must structurally hold as the organism accumulates toward ANI on the gated-RSI path. This ADR does for the organism what ADR 0024 does for the Native Edge Graph (its own words: "the composition that the scattered slice-ADRs each serve a fragment of") — one level up. It is a citing index plus invariants: it re-decides nothing, supersedes nothing, and stays current by construction via an id-based CI guard (§4) — a new ADR cannot merge without taking its seam-table row.

§1 The Organism (by citation, not re-decision)

The canonical top-level scaffold is Two Hemispheres (ADR 0018 OD-3, operator-decided 2026-05-30): the Body — the multi-threaded application engine under /src/ — and the Brain — the Agent OS under /ai/ — joined by the Neural Link (the JSON-RPC WebSocket possession seam). Four-pillar language (Swarm / Evolution) is Brain-internal elaboration, never the top-level frame. The always-loaded AGENTS.md carried two stale four-pillar anchors; #14526 is their governed repair (ADR 0018 + neo-identity-update + turn-memory-pre-flight apply).

Beyond the two hemispheres, decisions also attach to two cross-cutting planes: the Institution plane (coordination, identity, process — how the swarm works the substrate) and the Deployment plane (how the organism runs and heals in local + cloud topologies). The seam table names one owning surface per decision.

§2 The Seam Table

One row per present ADR (§4 enforces this mechanically). Seam = the boundary or contract the decision owns; Parent = the composition record that already aggregates it, where one exists.

§3 Trajectory Invariants

The shipped arc — v13 (the cross-family institution) → v13.1 (the self-healing immune system) → v13.2 (the harness/embodiment scope, #14038) → v14 (the institution cockpit, #13444) — is accumulation on the gated-RSI path (README, Discussion #10137). These invariants must hold at every step; each names its owning decision. A proposal violating one is challenging that owner, and must say so explicitly (Decision Record impact: challenges ADR NNNN).

1. The memory write never fails for anchoring/enrichment reasons — durability outranks decoration (owner: the #12972 hardening arc; consumed by 0024-family writers).
2. Render ≠ memory — rendered surfaces (handoffs, cockpit views, dashboards) are projections of substrate, never substrate themselves (owners: 0020/0029 on the Body side; the concept-slice contract on the Brain side).
3. Detect is separate from act — a health/integrity signal is an input to a diagnosis; a diagnosis is an input to a bounded actuator; no signal triggers a heal directly (owners: 0025 / 0026 / 0027).
4. Persistence is chosen — agents decide what they save; memory capture is never automated (operator covenant, 2026-07-02; the telepathy-aware curation contract around add_memory).
5. Structural weight is deposited, never assigned — graph importance is earned by real work and evaporates without it; fixes route to the scent, not the trail (owner: 0023).
6. Coordination stays flat-peer — primitives (0010 lane-intent, wake substrate 0002, stall findings 0030) inform self-selection; none may become assignment machinery.
7. Identity is descriptive, never prescriptive — rendered selves are records, not instructions (owner: 0018 for identity-of-Neo; the Institution-Cockpit render-model ADR, in authoring from #14445, extends this to identity-of-agents).
8. One vocabulary per contract — shared annotation/measurement vocabularies (trust tiers, the four-axis lattice, seam names) have exactly one importable authority; re-derived local copies are defects (empirical anchors: the 2026-07-02 review pair on #14520/#14522).

§4 The Staleness Contract (the guard that makes this ADR trustworthy)

A composition layer above an evolving corpus drifts unless kept current by construction:

• The guard: ai/scripts/lint/lint-adr-seam-table.mjs derives the present ADR ids from learn/agentos/decisions/[0-9][0-9][0-9][0-9]-*.md and fails CI unless the §2 table contains exactly one row per present id (both directions: no missing rows, no ghost rows).
• The authoring rule: a new ADR merges with its seam-table row in the same diff — the guard makes forgetting impossible, not discouraged. Deleting/renaming an ADR likewise updates the table in the same diff.
• Id-based, never count-based (graduation constraint): the guard reads what exists; it assumes no slot numbering.

§5 Boundaries

• Citing index semantics: this record re-decides nothing. Each seam's substance lives in its owning ADR; disagreements with a row's content are challenges to the owner, not to this index.
• Identity stays in 0018 — including the OD-3 scaffold this record cites.
• Subsystem composition stays with subsystem composers (0024 for the graph; 0025/0026/0027 as the immune-system triple) — this record composes the composers.
• No guide pair initially: a maintained guide without a mechanical freshness link is a second drift surface (graduation record, Option E falsifier). Revisit only on onboarding evidence.

Decision Record impact

none re-decided; aligned-with ADR 0018 / 0024; conditions the authoring practice of every future ADR via §4 (a process obligation, not an authority change).

Related: Discussion #13846 (graduation source) · #14525 (authoring leaf) · #14526 (the AGENTS.md alignment sibling) · ADR 0018 · ADR 0024

Origin Session ID: 8cf234b7-e698-47ca-99e2-bf865196b6aa