Design

Turn the inventory, the clusters, and the confirmed preferences into aws-design.json. Two product goals have to hold at once, and they pull against each other the moment a workload-level decision disagrees with a per-resource one:

1. Holistic — the output describes workloads, not independent rows.
2. Pre-determined where there is no ambiguity — an unambiguous service is not routed through a rubric that could reason its way to a different answer.

The precedence order below is what reconciles them.

Precedence order

1. Skip Mappings — not a target at all.
2. Specialist gates → Deferred. “We do not know” must never be overridden by anything below it.
3. Eliminators — hard technical blockers. Lambda’s 15-minute ceiling is physics, not a preference.
4. Direct Mappings → deterministic — the pre-determined recommendations.
5. Pattern constraint — the holistic layer, applied to whatever is left.
6. Six-criteria rubric — chooses within the pattern’s candidate set: Eliminators, Operational Model, User Preference, Feature Parity, Cluster Context, Simplicity, applied in order, first match wins.
7. Preferred-target substitution.
8. Post-selection — right-sizing from measured utilization, then CPU architecture.

INVARIANT: a pattern may never change a deterministic mapping’s target. It may only choose among rubric candidates.

If a pattern could override a fast-path row, the deterministic tier would stop meaning anything and its user-facing label (“Standard pairing”) would be false. So a genuine pattern/fast-path conflict is evidence the row does not belong in Direct Mappings — the fix is to demote the row, never to let the pattern win.

Right-sizing is post-selection, not a seventh criterion. The six criteria select a service and never touch capacity; adding a seventh would break “apply in order, first match wins”. Each rubric file gets an additive ## Right-Sizing section, structurally parallel to ## CPU Architecture.

Status — build step 5 (both passes, partial)

Pass 1 and pass 2 both run. knowledge/design/fast-path-services.json carries the disposition table; eight rubrics are on disk — compute.md, database.md, networking.md, messaging.md, analytics.md, storage.md, identity.md and fast-path.md. A resource routed by index.md to a category file that is not on disk still halts rather than being improvised; ai.md is on disk (the AI-workload rubric); only licensing.md, gpu-hpc.md and patterns.md are still to come.

Check the filesystem, not this paragraph. A status list in prose goes stale the moment a file lands, and the halt decision reads from disk. If this list and ls disagree, ls is right and this list is the bug.

A missing SIZING table is treated more softly than a missing RUBRIC file, on purpose. Without appservice-eb-sizing.json the design still names Elastic Beanstalk and states a dev-tier instance size as a default; without compute.md it would have to invent the service choice. The first degrades a number’s precision, the second fabricates the answer.

The _postconditions encode the FINISHED contract, so some still fail by construction — a halted design does not account for every resource. That is intended: the gate reports the skill’s gap loudly rather than letting a plausible mapping pass for a real one.

Step: Run the phase

1. Order clusters by tier: network/identity/secrets → data → compute → edge.
2. Run each fragment whose _trigger holds.
3. Run design-assemble.md.
4. Evaluate _postconditions. On all-pass emit HANDOFF_OK; on any failure emit GATE_FAIL and stop.