Skip to content

Paired disposable v3 Fleet launch plan — 2026-09-29 ​

The shortest intended path to actual Norn v3 HA evidence is the existing paired disposable/management/nyc3 and disposable/fleet/nyc3 topology. Do not create a new M3 Terraform profile and do not reuse the external-Mac topology. The paired infrastructure, retirement and normal five-node bootstrap exist. PR #196 corrected paired Linux runner admission. PR #197's cleanup proof-reader fix passed all 1,430 contract tests and merged as 172c2c3. That exact protected head passed 286 qualification tests; its source archive is built and verified.

This document is preparation, not spend approval. A new run needs a new bounded window and explicit approval of the larger paired topology after its resource envelope, current quote and initial backend plan are reviewable.

Exact topology ​

The management root creates one project and VPC, two s-2vcpu-4gb Droplets (authority and protected runner), and a two-node db-s-2vcpu-4gb PostgreSQL 16 cluster. The Fleet root creates a separate project and VPC, three s-2vcpu-4gb control Droplets, two s-1vcpu-2gb ingress Droplets, one regional load balancer, a two-node db-s-2vcpu-4gb PostgreSQL 16 runtime cluster and a two-node db-s-2vcpu-4gb MySQL 8 cluster. Firewalls, databases, tags, project attachments and run-scoped state accompany those billable resources.

The normal bootstrap installs Consul, Nomad, Norn and etcd on the three control nodes. Norn uses NORN_CONTROL_BACKEND=etcd; it does not use the managed runtime PostgreSQL cluster as its control store. The two ingress nodes are Nomad clients. This is the real v3 topology that the external-Mac pilot did not provide.

Current read-only price basis ​

The DigitalOcean API was read on 2026-09-29 with the retained service-scoped token. NYC3 reported s-2vcpu-4gb at $24/month or $0.03571/hour and s-1vcpu-2gb at $12/month or $0.01786/hour. DigitalOcean's current managed database price table reports db-s-2vcpu-4gb at $60.90/month or $0.09063/hour per node. The regional load-balancer price is $12/month; the four-hour projection below uses $0.01786/hour.

RootMonthly configuration quoteFour-hour usage projection
Management$169.80$1.01
Fleet$351.60$2.09
Paired total$521.40$3.10

The bounded request should reserve $10.00 total for four hours from first billable create through verified deletion. The compute/database/load-balancer projection is $3.10; the remaining $6.90 covers a two-hour teardown overrun, network overage and a possible $5 Spaces account subscription minimum. The current API credentials do not expose whether the account has already incurred that Spaces base charge, so it must remain an explicit uncertainty rather than being called included or free. This is not a promise of billing granularity. The Droplet catalog was refreshed at 2026-09-29T23:59:50Z with unchanged prices and NYC3 availability. Bind fresh per-root quote values into the run tfvars and the protected plans before create. The previous external-Mac $128.15 monthly quote does not authorize this topology.

Sources: the authenticated DigitalOcean /v2/sizes and /v2/databases/options catalogs, plus the public managed database pricing and load balancer pricing pages. Provider plan output remains the final resource-count check.

Version and authority prerequisites ​

  1. pilot260929g partial-precreate final-zero and policy restoration are complete; retain both evidence chains.
  2. Use qualified protected 172c2c3 and its verified checkout archive. PR #196 separates the paired Linux/X64 norn-pilot-<run-id>-runner from external-Mac macOS/ARM64 ...-mac.
  3. Pin Norn to the signed v3 integration candidate 8ac21f8e24bc3add8b1f4d3a1ae34fd5d2cd9b21, or a newer separately signed and qualified v3 candidate. The pilot260929g runtime pin ce047020 predates the v3 integration and is not eligible for this run.
  4. Validate the exact Linux Norn archive, manifest, signature and digest. Also validate the protected-root pins and reported versions for etcd, etcdctl, etcdutl, Consul, Nomad and Traefik. The three etcd binaries must report the same version.
  5. Choose a fresh run ID, hostname, two distinct VPC CIDRs, separate state keys, create expiry, Fleet retire-by and later management retire-by. Create fresh run-scoped credentials, PKI, TLS, owner inputs, keys and receipts. No signed pilot260929g artifact is reusable.
  6. Prove empty provider/run inventory and configure fleet-first retirement plus the management-last finalizer before create. The paired lane cannot produce both exact plans before every external mutation: both plans need their new remote Spaces backends, and the protected Fleet plan also needs the run-bound self-hosted runner created by the management root.

Mandatory staged planning sequence ​

Local preparation can bind a candidate commit, owner, fresh run ID, two non-overlapping /24 VPC ranges, exact bootstrap IP, SSH fingerprints, public hostname, authority Tailscale hostname, signed Norn v3 artifact, tool pins, per-root live quotes, expiration/retirement times, provider scopes, tfvars and the cleanup/finalizer inputs. Those values are mandatory; do not copy approval IDs, timestamps, state credentials or signed evidence from another epoch.

The workflow admission fix is merged and qualified. The first externally mutating step is the paired backend bootstrap. It requires a short-lived signed backend approval bound to the exact clean reviewed commit, cost-approval digest, account/team/Spaces owner, run ID, retire-by and two canonical declarations. It creates exactly two versioned buckets and two bucket-scoped read/write keys, then revokes its temporary broad key. Until that step succeeds, neither exact OpenTofu plan can initialize its remote backend.

After backend bootstrap, prepare and review the saved management plan on the owner Mac. scripts/disposable_create.py prepare proves an empty state key, fresh provider absence and the exact expected management resource set. A separate short-lived create approval binds that saved plan before apply. Management create and bootstrap are then required to establish the authority and exact norn-pilot-<run-id>-runner Linux/X64 protected runner. Only after that runner is registered and authenticated can .github/workflows/plan.yml produce the protected Fleet plan; the workflow has no management target and routes disposable Fleet planning to that run-bound runner after the admission correction above.

Use three review points because both plans cannot exist before spend:

  1. Approve backend bootstrap only, with zero compute, database or load-balancer creation. Review its canonical two-bucket/two-key plan and bounded cleanup path first. This step may activate the account's uncertain $5 Spaces minimum.
  2. Review the exact saved management plan, then approve management create within the same four-hour retirement window. Its current resource projection is about $1.01 for four hours, plus the possible Spaces minimum already called out above.
  3. Review the exact protected Fleet plan generated on the new runner, then approve Fleet create subject to the original $10 total paired-run ceiling, including all costs already incurred by stages 1 and 2. If Fleet approval is withheld or its plan is wrong, retire management and clean both backends.

These are three evidence-review stages, not a requirement to ask the user three times. A single explicit authorization may cover this exact topology, $10 total ceiling, four-hour lifetime and staged execution/retirement. Within that authorization, validate each generated plan and produce its bounded approval artifact. Ask again only if the plan exceeds the authorized resource set, budget, time window or other material scope.

The local neutral input manifest is /Users/arti/.config/norn-fleet/v3-launch-prep-20260929/paired-plan/manifest.json. Its adjacent paired-plan-proposal.json selects pilot260929h, owner aaronbarton, the verified account, canonical bucket/state names, hostname and non-overlapping VPC CIDRs. It leaves absolute windows and run authority unset until authorization and records the dependency that blocks each exact plan.

Locally verified public binary material ​

Neutral preparation is retained at /Users/arti/.config/norn-fleet/v3-launch-prep-20260929. The signed Linux amd64 Norn 8ac21f8e bundle passed platform-release-artifact verify against the release public key; receipt.json records the exact archive, manifest, signature and SBOM digests. The older ce047020 companion runtime must not supply the Norn executable.

The official etcd v3.5.17 Linux amd64 archive matched its upstream SHA256SUMS; etcd-receipt.json records URLs and individual executable pins. All three executables were run locally in an isolated Linux amd64 container with networking disabled and a read-only filesystem. etcd, etcdctl and etcdutl each reported 3.5.17; etcd-runtime-version-check.json retains that observation. These public inputs carry no run authority or spend approval.

Four-hour execution window ​

Reserve the final hour for retirement. Stop expansion immediately if bootstrap or readiness consumes the first two hours; preserve the retirement reserve.

  1. Create and bootstrap management. Prove its PostgreSQL, authority, protected runner, private/Tailscale access and recovery inputs before Fleet create.
  2. Apply the exact Fleet plan. Reconcile five nodes, load balancer, two managed database clusters, DNS, provider inventory and state before bootstrap.
  3. Run the normal bootstrap and retain exact binary, PKI, host, ACL and service receipts. Require all three etcd members healthy, the exact member list, empty alarm list, accepted bootstrap marker, initial managed credential publication, bootstrap-role removal and runtime-role access.
  4. Prove the authenticated Norn v3 API reads/capabilities and a separately labeled etcd read/write/CAS marker. Do not count direct etcd writes as Norn application-write evidence. Verify the Consul-locked active API identity and absence of PostgreSQL control-store configuration.
  5. Stop one control member. Prove two-member etcd quorum, continuing Norn API reads and no alarms; restart the member and prove healthy rejoin.
  6. Deploy the digest-pinned hello workload to both ingress nodes, migrate the managed MySQL database once, and prove load-balanced HTTPS write/read plus one-ingress continuity.
  7. If time remains, create and verify an etcd snapshot and its digest. Do not perform destructive restore without its separate reviewed recovery plan.
  8. Remove the workload, retire Fleet first, prove Fleet provider/state/DNS/ Tailscale zero, then retire management and prove paired final zero. Retain a billing review after provider absence.

The exact command recipe is retained in the neutral prep directory as paired-plan/fitness-execution-recipe.md. The released ce047020 companion fixture supplies /version, /records/*, /readyz and /healthz checks; it supplies no Norn executable. The v3 Fleet-target API is revision-fenced but has no standalone harmless create/delete cycle. This run therefore reports Norn API continuity and direct etcd CAS separately; it does not fabricate a Norn application CRUD or normal-release acceptance result.

Milestone interpretation ​

Successful bootstrap, one-member loss/rejoin and Norn API continuity provide meaningful M3 multi-host evidence. The workload and ingress check provide a bounded M4 slice. Protected qualification and exact version receipts advance M8. Complete retirement advances operational assurance but is not itself a v3 release sign-off.

The run does not complete M3 rotation, rolling upgrade, destructive restore or soak requirements. Fixed two-node ingress cannot prove M4 2→3→2. It does not upgrade the Mini, move a real application, complete the version matrix, or sign M5–M9. Record only the checks actually observed.