Nishi Frontier Program — the critical-path project plan

A PMP-grade plan for taking the from-god graph ecosystem to frontier on every node: a work-breakdown structure with a computed critical path, float, a Gantt, a dependency network, milestones, resource loading, and the live domain board — all wired to our own instruments and modeled as data our dependency engine consumes.

v3 PMP · published 2026-07-17 · unit = one work-session (a focused team+Claude block); external lead times noted in calendar. CPM computed by forward/backward pass over the dependency network below (48 tasks). Backlog is real data at knowledge/registry/frontier_queue.tsv — consumed by nx_frontier (READY/BLOCKED) today; CPM+visuals emitted by nx_frontier_board (F-202, staged — see §How this is wired).

1 · The critical path (computed, zero float)

Project span (all lanes, max parallelism) = 14 sessions of critical work + external board lead time for the silicon-boot milestone. The seven lanes run concurrently on disjoint file footprints (confirmed by the resource model in §7); the plumbing that makes the plan self-updating (F-201…208) runs alongside them.

2 · The three findings a PMP read surfaces

3 · Milestones (target session-offset from the computed schedule)

4 · Work-breakdown structure + CPM table (48 tasks, computed)

Every task: duration (sessions), predecessors, earliest start/finish (ES/EF), total float, and critical flag. Red rows = zero float = on the critical path. Float = how many sessions this task can slip before it delays M7. Source data: frontier_queue.tsv (id·dur·owner·deps), CPM computed by forward/backward pass.

5 · Gantt — the whole program on one timeline

Bars positioned by earliest-start × duration (one grid = one session). Red = critical path (NishiOS) · amber = GPU + silicon · blue = autonomy loop · green = early-win milestones · ◆ = milestone. Silicon boot is dashed (calendar-gated by board arrival, not effort).

6 · Critical-path network (the chain that sets the finish)

The critical path is a single serial chain — that is exactly why it governs. Each node is strictly gated by the prior (a native core needs the contract; UEFI needs the core; os-compare needs a booting OS). No amount of extra people compresses it; only early staffing and zero blocking help.

7 · Resource loading — who is the bottleneck

Each lane has one owning role; tasks within a lane are serial. So the schedule is feasible with exactly the seven lane-owners + the plumbing/support roles — no owner is ever double-booked, which is why "7 concurrent workstreams" is the correct parallelism, not an arbitrary number.

8 · The domain board — every /compare node, coverage vs EARNED maturity

Wired from the live nishi_compare_index (46 domains) × the maturity ruler. Coverage = capability-presence (inflatable, honest-labeled). Maturity = evidence-capped L0–L5 (≈ coverage × 0.2 at floor; climb by earning LIVE→MCP→multi-node→hardened). The gap is the anti-navel-gazing signal: it is the real backlog of hardening work, and it is exactly what the autonomy feeder (F-208) turns into new F-rows.

The aggregate: coverage across the plane is ~5× the earned maturity. That gap is not failure — it is the measured, prioritized hardening backlog, and closing it (LIVE→MCP→multi-node per domain) is what moves the ecosystem's 328‰ toward S-class. The feeder (F-208) reads this table and momentum radars to file the next rocks automatically.

9 · Experiential census (top-down, 19 axes) — GAP 11 / PARITY 8 / AHEAD 0

From the self-emitting /experiential. Competitor attestation = sovereign-fetched READMEs (playwright/puppeteer/cypress/selenium/axe-core/lighthouse/backstop/browser-use). PARITY caps until a measured win; the AHEAD conversion is lane F-106.

10 · The ladder from god (condensed — state + named next)

11 · RACI — the house lived in (PM Accountable everywhere)

Roster / what each teammate watches: pm ranks+owns the door · researcher 2025/26 corpora · census disk-vs-evidence · referee grades check-ins · adversary/critic refute every green · architect daily family-tree board · genealogist lineage of every organ (roots-to-god, provenance to seed) · librarian files /evidence · janitor debt+orphans · doctor diagnosis loop · conductor queue+cadence · supervisor fleet+leash · hr staffing+wage ledger (⚠nx_hr* deletions = F-003 gate) · warden policy+money-safety · fin revenue ledger · modelwright model lane · racing crew head-to-heads · engineer lanes ×7 in §4.

12 · Autonomy — 70% → 100%

• Zero-Claude today: cadence loop (60s queue-tend + leash-walk, survived the app crash) · arch board cron 06:20 · autograde pulse (re-grades + republishes itself) · guard respawns · leash reclaim · admission gating.
• The loop (F-204/208): MEASURE → RANK → ASSIGN → EXECUTE → VERIFY → PUBLISH → FEED (files new F-rows from maturity-gap × momentum). Gauge = autograde's zero-Claude share, 70%→100%. The forge closes the last mile once the model lane clears its RED.

13 · Value + revenue — the mission fuel (funnels honestly unmeasured)

F-302 makes income visible on nx_fin_serve; F-303 publishes the living-wage staffing curve here. Revenue → reinvest → staffed roles is reported, not asserted.

14 · Progress ledger (newest first)

15 · How this is wired (the honest architecture)

• The backlog is DATA, not prose: 48 tasks live at knowledge/registry/frontier_queue.tsv in the exact 9-column format the existing nx_frontier engine reads — so READY-vs-BLOCKED (each blocked task naming its first unmet predecessor) is derivable today by our own dependency engine, not drawn by hand.
• The CPM + Gantt + network on this page were computed this session by forward/backward pass over that data. The durable, self-emitting version is nx_frontier_board.nx (rung F-202) — it EXTENDS nx_frontier with the critical-path/float pass + the SVG emit + a compose of the live instruments, and emits this page from the run so it can never drift (the same sync-by-construction pattern as /experiential). Status: staged — one build (nx_sov_build_run nx_frontier_board / /api/build) from live. Until it lands, this page is hand-published via the zero-ssh recipe. That is the honest line: the model and data are graph-native now; the emitter is the next rung.
• Live instruments composed: nx_ecosystem_maturity_rollup (328‰) · nishi_compare_index (46 domains) · nx_workflow WMS board (1,614 registry rows) · nx_status (15/15 fleet) · nx_https_get (edge verify). The family-tree SVGs render today at /compare/atlas from the 16,222-node eco_graph; wiring them straight into /plan needs the eco_graph tools' cap (F-205/F-210).
• Not sprawl by construction: the plan composes the existing graph/queue/compare/rollup organs; F-203 puts every F-row into the WMS registry SSOT so this page and nx_ws_board render the same truth.

Republish (zero-ssh): nx_fs_write chunks → nx_site_publish → nx_https_get verify · durable scoped cap banked · self-emitting page = F-202. Backlog SSOT migrates to the WMS registry at F-203.