NavalAI

NavalAI generates the safest, most energy-efficient, BUILDABLE vessel for a specified mission and budget — and shows its evidence.

— PLM.md §0, which is the home of the vision, the two product families and the gap analysis. This blockquote is a citation, not a second copy.

The gap is not that the industry lacks tools. CAD systems, hydrostatic packages, RANS solvers and open-source naval-architecture calculators all exist, and several of them are excellent. The gap is that there is no broadly accessible, integrated, physics-aware workflow in which a technically capable person can state a MISSION and get back a safe, efficient, structurally realisable vessel with a traceable evidence trail. Today that path is

CAD → hydrostatics → resistance → CFD → manual stability
    → manual structure → manual electrical → manual drawings

— eight tools, every arrow a hand-carried file and a re-keyed number. NavalAI collapses it into one governed pipeline, and the governing is the product: every quantity carries {value, tier, sigma}, the ladder refuses what it cannot compute in-process instead of substituting something cheaper, and a gate that misses its bar stays red, on the record, with an owner and a review date.

Two product families share nearly the whole engine and differ mainly in the mission layer — recreational / DIY (people, range, speed, comfort, cost, coastal conditions) and the autonomous marine drone (payload, endurance, sea state, sensor package). The drone family is DECLARED, not built; PLM.md §2.0 says exactly what is missing and why it is the mission layer rather than the physics.

What the picture is, and is not. These renders are the project's visual vocabulary — the vessel families and hull forms the two product lines aim at, from the solar catamaran and electric day boat through the survey and solar USVs, down to the ten hull-form families along the bottom row. They are REFERENCE IMAGERY, not generator output: in this project a rendered image never defines Cp, LCB, displacement, stability or resistance — geometry comes from the mathematical kernel and every claim above carries its evidence. (Of the ten hull forms pictured, the grammar today expresses the chined families and the round bilge; multi-chine and SWATH are on the roadmap — docs/BUILD-PLAN.md §16 phase PV.)

Licensed under GNU AGPL-3.0 (see LICENSE).

Quickstart

python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt              # pinned, lower and upper bounds
pip install -r requirements-optional.txt     # capytaine/cadquery/mujoco tiers

python3 -m navalai.gates                     # what is proven, and what is not
python3 ui/server.py                         # slider surface -> localhost:8642
python3 benchmarks/wigley.py                 # the Michell Wigley curve
python3 -m pytest tests/ -q                  # the whole ladder (~10 min)

python3 -m navalai.gates is the honest status of the project and the only one — no document in this tree carries a gate verdict. python3 scripts/reconcile_gaps.py does the same for outstanding work.

Where to start reading

You want	Read
what this is for, what a product may be, who owns what	PLM.md
how the system is built, and what order it is being built in	docs/BUILD-PLAN.md
what has been MEASURED, and what was refuted	docs/research/*.md
what this project learned the hard way	docs/LESSONS.md
how to work in this repository	CLAUDE.md
what a production vessel needs beyond a hull	docs/research/PRODUCTION.md

The ladder

Tier	What	Speed	Module
L0	algebraic feasibility (closed-form checks)	~0.05 ms	grammar.py
L1	hydrostatics · Michell+ITTC57 · Holtrop-Mennen · energy/weight budget	~20 ms	hydrostatics.py resistance.py holtrop.py energy.py
L2	Capytaine BEM seakeeping (convergence-swept)	~s–min	seakeeping.py
L3	OpenFOAM interFoam resistance (runs on the Mac node)	hrs	cfd/
R	ISO 12217 / 12215-5 subsets (assessment aid)	~ms	rules/

The L0 count is deliberately not quoted here. The audit measured 9 live constraints against a "49" the plan claimed and a "30+" this file claimed (gap E4) — a count in prose is a count that drifts, so ask grammar.check.

L3 is the one tier that cannot run in-process: evaluate raises TierRequiresOperator and names the operator route rather than quietly handing back an L1 number wearing an L3 badge.

Surrogate spine (surrogate.py): ARD kriging + Kennedy–O'Hagan co-kriging, batched-EI infill, OOD → ladder escalation. Generative (generative.py): GMM family model + performance-conditioned sampling + 2-D latent map (guided tabular diffusion is the planned drop-in upgrade behind the same interface).

What the kernel does — each line checked against the code, 2026-08-13

This section states only what was re-verified by reading the module named beside it. This project's recurring defect is a claim that is not true of the code — four documents once asserted that two shipping subsystems did not exist — so nothing goes here on the strength of having been written before.

• The sectional-area curve and the design waterline are DESIGN CURVES, not outputs. geometry.sectional_area(params, x) and geometry.design_waterline(params, x) are closed form. Cp and lcb are GENES in grammar.PARAMS, and geometry.sac_exponents SOLVES the area-curve exponents to deliver the requested pair, refusing a (Cp, lcb) it cannot reach at that x_mb instead of silently landing somewhere else. A naval architect chooses Cp for the design Froude number; here that is an input.
• The section kernel is not three points. Hull.section_control() returns the exact five-point description — two legs and ONE quadratic Bézier arc — and immersed_section integrates it in closed form. Hull.section() is adaptive: 3 points at roundness == 0 (bit-for-bit the old polyline, fenced at 1e-12) and 257 above it.
• Resistance knows where it stops being valid. resistance.FN_MICHELL_MAX is 0.45; past it the result is badged L1-INVALID, and evaluate.tier_rank puts that at −1, BELOW L0's 0 — so an out-of-envelope number can never win a comparison against a valid one.
• The genome is SIXTEEN parameters (grammar.N_PARAMS), not fifteen.
• The optimiser minimises energy per mile, not drag. optimize.HullProblem returns (wh_per_nm, build panel area, distance from the GM band) — build area is in the objective vector, so panel cost is optimised against, not checked afterwards.

Honesty rules (enforced by tests, not vibes)

1. every quantity carries {value, tier, sigma} — no bare numbers
2. any kept design re-validates up the ladder; surrogates refuse OOD queries
3. the LLM translates missions and explains — it has no code path to geometry
4. retrained surrogates that degrade the frozen benchmark never deploy
5. rules output leads with ASSESSMENT AID — NOT CERTIFICATION and declares every approx-basis threshold
6. never soften a failing gate threshold to make it pass — a failing gate is information. RED gates are recorded in data/gate-ledger.json with a measured watermark, an owner and a review-by date, never reworded

Gate registry

This table is GENERATED. It used to be hand-maintained, and every number in it drifted: the per-gate test counts were stale throughout (Gate 1 read 13 against 22 actual), Gate 2U — a RED gate — had no row at all, and the Gate 2M row carried a figure three later commits had superseded. Regenerate with python -m navalai.gates --readme --write; a test fails if you do not.

Run python -m navalai.gates for live status; this table is the REGISTRY (what is gated by what), regenerated from navalai/gates.py. A RED row's measured watermark, owner and review-by date live in data/gate-ledger.json — never in prose here, which is how five different Gate 2M numbers came to circulate at once (gap J1).

Gate	Scope	Verified by
Gate PROP	propulsion integration: motor power and prop space are constraint rows; transom Froude, bilge-keel span and pitch entry are measured reports	tests/test_propulsion.py (27 tests)
Gate MORPH	morphological plausibility: a hull must LOOK like a boat, judged on descriptors measured against 58 published hulls (89-92% of L0-valid generated hulls fail; 0 of 58 real ones do)	tests/test_morphology.py (19 tests)
Gate HULL-KB	the reference corpus is learned and reconstruction is proven: KB citations resolve, and the measured cruiser / warped-V / split-topology reconstructions stand	tests/test_hull_kb.py (6 tests)
Gate BARGE	a houseboat is a rectangle: the widened envelope draws it, the frozen DRAW box protects history, and the pointed bow stays reachable	tests/test_barge_bow.py (6 tests)
Gate PREFLIGHT	before a solve is bought: theory answers what it can, a surface already solved is recognised, the family band speaks, an impulsive start above the measured Fn ceiling is refused — and after it, Kelvin validates and an A/B is only attributable when the meshes match	tests/test_preflight.py (11 tests)
Gate DELIVERED-FORM	form_coefficients measures the surface the ladder floats — notch, hole and knuckle chain included; no feature reports an impossible Cm; the delivered SAC is the commanded SAC for every feature that folds its area into the solve, and the two that do not DECLARE their drift	tests/test_delivered_form.py (27 tests)
Gate SWEEP	seam properties swept over the production flow: descriptors vs ladder, the SAC contract, declared-vs-drawn levers, cache identity, gene reachability, repair-vs-constraint, exception containment, receipts, numbers declared twice, the two design routes, and station convergence	tests/test_gap_sweep.py (5 tests)
Gate REACHABILITY	every post-hoc gene is classified BLIND, REQUESTED or WITHHELD-with-a-measurement; a mission that asks for a protected prop gets a hull that draws the tunnel, one that does not gets none, and the aft prior's share is recomputed rather than remembered	tests/test_feature_reachability.py (11 tests)
Gate CFD-KB	the CFD anchor book: every record honest about settledness and GCI, unsettled records never support a prediction, family bands refuse outside measured Fn support, and the hb19 L1 anchor carries its sigma and basis	tests/test_cfd_kb.py (14 tests)
Gate RHO-X	the bilge radius varies along the length: the warp is the flare law's shape, rho_len = 0 is bit-identical legacy, the control points and the area coefficients are per-station, and a hull may be hard aft and round forward	tests/test_rho_x.py (6 tests)
Gate SPLIT	the split stern: split genes are no-ops at zero, the SAC stays the net contract, the wall is vertical at y_split, the waterplane integrals subtract the hole, and split + tunnel on one station is refused by name	tests/test_split.py (10 tests)
Gate TUNNEL	the W-stern: tunnel genes are no-ops at zero, the SAC stays the net displacement contract to the bit, a crown above the waterline is refused by name, and the wetted surface counts the tunnel walls	tests/test_tunnel.py (8 tests)
Gate DWL	the design waterline: dwl = 0 is bit-identical legacy, the SAC stays the displacement contract at every dwl, the barge passes the general critic, and the requested-vs-achieved gap is a receipt (Hull.dwl_deviation)	tests/test_dwl.py (9 tests)
Gate PARENTS	the parent library is proven and the distortion operators deliver: parents pass their family critic, Lackenby clips into the deliverable band, rescale stays homothetic, and family missions seed the optimizer without touching recorded streams	tests/test_parents.py (8 tests)
Gate HB16	the 16 m liveaboard probe's refusals stay refusals: 3 t is refused naming the family, and the motor-power fix the old sentinel demanded stays landed	tests/test_houseboat_16m.py (9 tests)
Gate G-VISUAL	the canonical views ship with every deliverable and a refused shape is SEEN: fixed cameras, descriptor sheet beside the pictures, written on refused hulls too	tests/test_views.py (4 tests)
Gate CLI	the CLI face DESIGNS for a brief rather than grading the reference hull under its name, says which of the front's designs it is reporting and why, and refuses an impossible brief with the rows that refused it instead of a plausible boat	tests/test_design_report_cli.py (3 tests)
Gate UI-SEED	the studio's default genome BUILDS before it is served, and a refused mesh is PAINTED in the viewport rather than leaving a blank stage with the reason in the response body	tests/test_ui_seed_builds.py (8 tests)
Gate 0G	the ladder cannot be talked into passing	tests/test_gate_integrity.py (39 tests)
Gate 0R	a missed clause is RED BY RECORD, never prose in a scope	tests/test_red_by_record.py (14 tests)
Gate 0K	geometry kernel: SAC/DWL design curves + N-point section	tests/test_geometry_kernel.py (29 tests)
Gate 1E	the stages agree with each other: one geometry, one resistance, one ply, tier+sigma across every handoff	tests/test_end_to_end_flow.py (14 tests)
Gate 0F	the hull-form library: bands ordered, every band carries its basis, no family contradicts its own Froude regime	tests/test_formlib.py (51 tests)
Gate 0X	the experiment suite: controlled sweeps hold their controlled quantities, out-of-envelope points are refused, and the Michell interference phase matches an independent superposition	tests/test_experiments.py (52 tests)
Gate 1M	the vessel: topology/manning/regime, the parallel-axis I_T, separation in the PRODUCTION wave term, and no multihull safety verdict from a monohull GM floor	tests/test_multihull.py (34 tests)
Gate PV-B	vessel-conditional proportion bands, the sourced size box, and the multihull stability refusal	tests/test_vessel_bands.py (20 tests)
Gate 0B	buildability metrics are PROXIES that refuse rather than default, are grid-converged by a measured residual, and price manufacturing in ABSOLUTE m^2 — never in a ratio an optimiser can inflate	tests/test_buildability.py (11 tests)
Gate 0	grammar/geometry/DB	tests/test_phase0.py (16 tests)
Gate 0E5	the geometry kernel against REAL PUBLISHED hulls: offsets from three independent source families round-trip through the genome, and the SHAPE residual is reported, not just the six scalars	tests/test_e5_real_hulls.py (276 tests)
Gate 0E5C	the hard-chine EVIDENCE: published hard-chine geometry is acquired, is PUBLISHED_PARAMETRIC rather than traced off a drawing, carries a real chine, and every refusal is recorded with its reason	tests/test_e5_hard_chine.py (24 tests)
Gate 0E5C-CAP	the hard-chine CAPABILITY: published hard-chine series are EXPRESSIBLE in the grammar -- deadrise distribution within 1 deg at the transom, 50% and 75% LWL, and the hull round-tripping at its published prismatic coefficient. The count and the remaining blockers are the ledger's (data/gate-ledger.json; docs/gates/E5-CHINE.md), not this scope's	RED — data/gate-ledger.json
Gate 1	L1 physics + Wigley anchor + <50ms	tests/test_phase1.py (36 tests)
Gate 1H	Holtrop-Mennen vs the 1982 worked example	tests/test_holtrop.py (57 tests)
Gate 1C	the constraint vector: complete, ordered, finite, and no undefined state reported as ideal	tests/test_constraints_honest.py (25 tests)
Gate 1P	the L1 physics core says what it actually computed	tests/test_gapfix_physics.py (27 tests)
Gate 1b	NSGA-II Pareto front	tests/test_optimize.py (16 tests)
Gate 2	Capytaine BEM (Hulme anchor)	tests/test_phase2.py (18 tests)
Gate 2R	CFD reference parity + GCI honesty	tests/test_cfd_reference_parity.py (100 tests)
Gate 2P	slamming pressure is measurable: a bow patch that partitions the hull, a function object that refuses to point at nothing, and a Wagner C_p guarded at both limits	tests/test_slamming.py (14 tests)
Gate 3	surrogate spine: GP + co-kriging rho, OOD refusal, batched-EI infill + the held-out across-seed error bar	tests/test_phase3.py (25 tests)
Gate 4	generative + slider p95<100ms (raw feasibility: Gate 4F)	tests/test_phase4.py (27 tests)
Gate 4U	the builder surface's honesty contract: every payload declares its source, no NaN on the wire, one home per absence, the kit/mould route reads the refold TREND, KG from the keel plane, Fn > 0.45 refused by name, and no gate verdict without a suite run	tests/test_ui_surface.py (24 tests)
Gate 5	mission translation + LLM seam	tests/test_phase5.py (19 tests)
Gate 6	rules-as-code mechanics	tests/test_phase6.py (6 tests)
Gate 7	flywheel: frozen suite != training draw, monotone regression mark, wall clock, committed baseline	tests/test_phase7.py (15 tests)
Gate B	grammar AST + bend radius + 8-D genome	tests/test_stageB.py (13 tests)
Gate C	agentic PLM network + engineer + STEP/IGES	tests/test_stageC.py (14 tests)
Gate D	waves/RAO response + dynamics + CFD post	tests/test_stageD.py (19 tests)
Gate E	latent-space evolution + latent GP	tests/test_stageE.py (3 tests)
Gate F	panel unroll/DXF + Pareto dash + handoff receipt	tests/test_stageF.py (14 tests)
Gate G	APSE: similitude/ITTC-78/cost/planner/evidence	tests/test_stageG.py (58 tests)
Gate MULTI-CHINE	the topside is a knuckle list: a second chine is an exact vertex, and k=0/1 are unchanged	tests/test_multichine.py (7 tests)
Gate CFD-CMP	CFD comparisons are in newtons, above the scatter, on matched meshes -- or refused	tests/test_cfd_compare.py (6 tests)
Gate V2.1	arrangement grammar: envelope, spaces, deck zones, and an L0-A that names the space it refuses	tests/test_arrangement.py (42 tests)
Gate V3.0	governance compiles to a parameter box and to constraint rows, ratchets only tighter, and the ladder never imports it	tests/test_policy.py (55 tests)
Gate 6P	the product surface: scope guards refuse what does not govern, and the mission contract binds	tests/test_gapfix_product.py (28 tests)
Gate L	one limit, one home; scantling derived from the rule	tests/test_limits_single_source.py (22 tests)
Gate L2	numeric literals are categorised, not banned: an AST fence that catches a physical constant in any expression position	tests/test_constpolicy.py (43 tests)
Gate 6M	manufacturing back end: nesting, BOM, developability controls, export receipt (refold onto the hull: Gate 6D)	tests/test_manufacturing.py (46 tests)
Gate R3	the ladder is climbable: L2 escalation, monotone tier promotion, honest refusal of L3	tests/test_ladder.py (8 tests)
Gate S	the MDO spine: one terminal state per genome, append-only archive, legal-transition graph, unmeasured metric refused	tests/test_pipeline.py (48 tests)
Gate SG	the gap queue: findings are work items, not prose	tests/test_gaps.py (23 tests)
Gate R4	the ladder is WIRED: every claimed tier reachable from evaluate(), L3 read from recorded evidence and never solved, tier R alone can refuse a design	tests/test_ladder_wiring.py (17 tests)
Gate SR	gap state is derived from the code, not from prose	tests/test_reconcile_gaps.py (38 tests)
Gate 2S	one settled_drag: one cell-count rule, one settledness rule, and a component that oscillates cannot hide under a stable total	tests/test_settled_drag.py (52 tests)
Gate 4H	the surrogate refuses what it has not seen, and the feasibility bar measures the model rather than the sampler	tests/test_surrogate_honesty.py (18 tests)
Gate 2T	tank resonance is diagnosed, and a period is never claimed from too few cycles	tests/test_tank_resonance.py (25 tests)
Gate V2.0	refdata spine: every constant carries source + basis	tests/test_refdata.py (27 tests)
Gate 2A	CFD-admissibility screen: a grammar-valid hull is not a CFD-meshable one	tests/test_admissibility.py (29 tests)
Gate 2D	the admissible design space refuses pathological hulls before OpenFOAM, by the named gate, at milliseconds per genome	tests/test_admissible_space.py (11 tests)
Gate 2L	the prism-layer cap is a measured value, and a clean mesh with no boundary layer is not a pass	tests/test_layer_cap.py (7 tests)
Gate 2F	STL forensics: watertight is not valid, and the surface handed to snappy is measured rather than assumed	tests/test_stl_forensics.py (16 tests)
Gate 2B	Blender-native hull generation, measured and REFUSED on the hull path: a 0.05 m voxel remesh destroys the chine	tests/test_blender_hull.py (12 tests)
Gate 2C	the campaign classifier names the mechanism that actually failed, and refuses one it cannot measure	tests/test_campaign_classifier.py (17 tests)
Gate 2K	the smoke verdict is the runner's own bar, an absent or truncated log is never promoted, and a smoke refusal can never read as a solve failure	tests/test_smoke_verdict.py (9 tests)
Gate RT	the resistance tier: its envelope, and bit-exact values under any optimisation of the Michell integral	tests/test_resistance.py (17 tests)
Gate XP	cross-machine parity: platform-class differences are distinguished from code-class ones, and a verdict is never excused by a tolerance	tests/test_parity.py (5 tests)
Gate EP	evidence promotion: a RECEIPT ONLY or EMPIRICAL basis may never be reworded into a mechanism, the crossover is never cited without its standing, and the refutation record cannot thin out	tests/test_evidence_promotion.py (22 tests)
Gate HC	the hull evaluation contract: four separate verdicts, a regime, a derived mesh+solver prescription and its cost, in one receipt that never reads OK without evidence	tests/test_contract.py (28 tests)
Gate 2Y	population identity: (arity, seed) not seed alone, an undeclared seed is UNKNOWN and never silently development, and NO committed artifact may draw from the validation or held-out seed	tests/test_population_split.py (20 tests)
Gate 2N	physics sanity: a force must have the right sign, a finite magnitude, and a defensible ratio to the cheap model — and an unreadable history is never a result	tests/test_physics_sanity.py (11 tests)
Gate 2W	the case writer consults the physics floors it was given: cells-per-wavelength and the Reynolds regime decide before a mesh is sized	tests/test_case_wiring.py (19 tests)
Gate 2E	the tank contains the ship's OWN wave: the domain length scales with 2piU^2/g, not with hull length alone, and the design point is left bit-identical	tests/test_domain_wavelength.py (4 tests)
Gate 2X	h011/h012: the geometry is provably clean, and no measured descriptor separates the failures from the passers — a criterion that does not separate is not a criterion	tests/test_h011_h012_invariant.py (15 tests)
Gate 2H	surface repair on the import boundary, and generated geometry refused rather than healed	tests/test_mesh_repair.py (9 tests)
Gate 2G	KCS benchmark geometry: present and accepted (scripts/fetch_benchmark_geom.py)	tests/test_benchmark_geom.py (4 tests)
Gate 6R	review-record mechanics, basis routing, AND edition parity: both implemented standards name the dated edition held	tests/test_phase6r.py (12 tests)
Gate PF	physical form regression: six deterministic hulls stay boat-like (descriptors, SAC shape, sourced bands, ratchet)	tests/test_physical_form.py (46 tests)
Gate VM	vessel matrix: five vessel classes end-to-end (ladder + manifest), refusals explicit, roles judged apart	tests/test_vessel_matrix.py (9 tests)
Gate DC	design certification: cheap classify-and-rank with receipts, banded speed curves, loading matrix, regime refusals	tests/test_design_certification.py (13 tests)
Gate FG	fidelity governor: five gates route (not badge) a design to a tier, with named receipts and no silent passes	tests/test_select_fidelity.py (42 tests)
Gate 4F	raw generative feasibility: UNFILTERED model draws vs the >=99% bar (BuildPlan Phase 4)	RED — data/gate-ledger.json
Gate 2P-Stack	Gate 2 physics stack: hydrostatics as a MATHEMATICAL invariant (Wigley 4LBT/9, second-order convergence), and KCS pinned as solver verification only	tests/test_gate2_physics_stack.py (5 tests)
Gate 2M	SOLVER VERIFICATION ONLY (not small-craft validation): KCS/JBC OpenFOAM numerics w/ per-case GCI	RED — data/gate-ledger.json
Gate 2U	unattended meshing (plan: >=95% of a 200-hull batch)	RED — data/gate-ledger.json
Gate 6D	developable-panel refold: EXPORTED panels back onto the hull vs the 5 mm bar (BuildPlan 12.3)	RED — data/gate-ledger.json

Capability truth — IMPLEMENTED / PARTIAL / REFUSED / UNKNOWN

Re-verified against the modules named, 2026-08-14 (the 2026-08-13 list below it claimed four absences that code and gates had since disproved — prose claiming a capability is absent when the tree proves it exists is the same defect as overselling, mirrored). A README that oversells is worse than one that undersells, and honesty rule 6 applies in both directions:

• multihull hydrostatics + righting arm: PARTIALLY IMPLEMENTED, verdict REFUSED. The parallel-axis sum I_T = Σ_j [ I_T,j + A_wp,j · d_j² ] is in hydrostatics.solve (Gate 1M measures it); the heeled-waterplane solve exists (gz_curve — the catamaran curve saturates at ~2.46 m where GM·sin claims 10.3 m, peaks and declines, computed not narrated); NZ Part 40A cl. 1.4 clauses (a)/(b) are MEASURED (multihull_gz_assessment). The CRITERION still cannot pass: clause (c) needs windage the genome cannot declare, clause (d)'s text is unread — so every multihull carries an explicit stability refusal, never a monohull GM verdict. Refusal-first until the vessel can declare its windage.
• catamaran resistance: IMPLEMENTED in the production path. total_resistance(..., separation=) carries the demihull spacing into the interference term (Gate 1M: ratios 1.107/1.002/0.996 at s/Lwl 0.40/0.60/1.00 measured through evaluate). Experimental validation: UNKNOWN — Insel & Molland (1992) are cited, not transcribed; self-consistency is not validation, and no tank anchor exists in tree.
• multi-chine hulls: NOT IMPLEMENTED. The section law delivers hard chine and the roundness bilge fillet (round bilge IS expressible — formlib re-audited 2026-08-14); a second chine is not. Pinning roundness = 0 for sheet-built typologies is a stopgap correct for the unroller we have, not a principle.
• CFD: READY, not VALIDATED. cfd.manifest.CFDManifest is the one vessel description a case renders from (mass/centres/attitude from the L1 weight model at the SOLVED equilibrium — never rho·V_STL); Gate 2M's watermark is NONE and stays red until a converged campaign exists on a CFD node. CFD-READY is a bookkeeping claim, CFD-VALIDATED an experimental one; this repo currently earns only the first.

Everything else:

• the autonomous-drone mission layer: PARTIAL. MissionSpec.payload (PayloadSpec: mass/volume/continuous+peak power/endurance/mission distance/sea-state, positioned mass + hotel-load wiring, Gate VM) ships; sea_state is a DECLARED requirement nothing can assess yet, and optimize.HullProblem still has no wake, acoustic, survey-km/kWh or time-on-station objective (PLM.md §2.0).

• guided tabular diffusion (GMM baseline stands in; same interface)

• LoRA-fine-tuned translator (rule floor + sanitising LLM seam stand in)

• Tokyo-2015 calibration. OpenFOAM now RUNS (Mac simulation node), so this is no longer metal-gated: Gate 2M and Gate 2U are RED, measured. Both are in data/gate-ledger.json with a watermark, an owner and a review-by date, and no figure is repeated here — that habit is what produced five circulating Gate 2M numbers. The register (docs/GAP-REGISTER.md §F) carries what is still wrong with the number.

• a SECOND benchmark anchor. KCS is a 230 m containership with no chine, no immersed transom and no spray; the SKUs are 4–14 m chined craft. A green Gate 2M is not small-craft validation (docs/BUILD-PLAN.md §11.4).

• licensed-text ISO parity — Gate 6R is RED, because the review record cannot name the dated editions it checked. (ES-TRIN itself is no longer absent: navalai/rules/estrin.py ships and Gate 6P proves its scope guard.)

• BuildPlan 2 beyond V2.1: V2.3, V2.5 and V2.6. V2.0 (refdata spine) and V2.1 (arrangement grammar) both ship with gates; V2.2 tier E and V2.4 tier F are partial. V2.3/V2.5/V2.6 have no gate row at all, so their bars — ≥95% of layouts passing Tier E, the USCG worked examples including the −0.81 plywood case — exist only in prose. docs/BUILD-PLAN.md §15.3.

• BuildPlan 3 beyond V3.0: governance ships (navalai/policy/, Gate V3.0). Mission intelligence, component models, BOM closure, procurement, the order package, the twin and fleet learning do not.

• RUN, entirely. No telemetry, no sensor ingestion, no in-service data — and the gap register does not know it, because the audit that produced the register was scoped by four documents that contain no RUN phase. docs/BUILD-PLAN.md §15.2.

  This list previously read "arrangement grammar, tiers E and F do not [exist]" while the GENERATED gate table above carried Gate V2.1 and Gate V3.0 — the file contradicted itself, because the table is regenerated and the prose is not. Corrected 2026-08-11; the durable fix is to stop keeping a second copy (docs/BUILD-PLAN.md §0).

---

Alignment campaign (stages B–F) — historical

The original agentic-PLM plan was audited against the build; all 11 gaps closed behind gates:

Stage	Closed	Gate
B	grammar AST + typology type-checker · plywood bend-radius · 8-D pPCA genome	Gate B
C	async agent network (Orchestrator/Builder/Validator/Engineer, audit trail) · engineer metrics · STEP/IGES export (CadQuery/OCP)	Gate C
D	JONSWAP + heave-RAO response spectra · inertia/mooring/lifting + MuJoCo cross-check · CFD runner + forces parser + Roache GCI	Gate D
E	NSGA-II over the 8-D genome · latent-GP (measured: 8-D costs 2–3× accuracy)	Gate E
F	developable-panel unrolling → DXF · Pareto dashboard · handoff-latency receipt (<1% of physics)	Gate F

Status and test counts are in the generated table above, not here — the same number written twice is this codebase's recurring defect, and this table carried a second copy of five of them.