Kinesis: dynamics on three tiers
16,629 words24 subsections25 tables2 listingsPlan version 1.1
On this page24
Crates: frob-dyn. The step's kernels live in frob-kernels, which Spark owns (§21, §34). Kinesis calls
frob-contact (§18), frob-phylum (§17), frob-actuate (§15) and, through the airlock frob-fsim, FrankenSim's
fs-mbd, fs-contact, fs-time, fs-query, fs-eproc, fs-uq and fs-rand.
16.1 What Kinesis owns
Kinesis owns the advance of state in time and the rules that say what a computed trajectory may be cited as:
- the three tiers and their step profiles (§16.3);
- the step itself: a transaction with located events on the referee and on the strict profile (§16.5), and the ten-stage lockstep step on the throughput tiers (§16.6);
- the integrators (§16.7);
- the trust ladder: the fidelity ledger, the promotion rule and its tier correction (§16.8), rank agreement between tiers and model-form rank checks (§16.9), the audit for energy created by the simulator (§16.10), and the background referee with its coverage figures (§16.11).
Kinesis does not own contact geometry or contact laws (§18), the packing of different robots into one batch (§17), device back ends (§21), clocks and packet queues (§20), or the storage of telemetry and of the energy ledger (§29). It calls them and states what it needs from each.
Two units recur and are never mixed. A world-step is one physics step of one world. An env-step is one control interval of one world, that is, all the world-steps between two decisions. A rate is stated in one of the two with the decimation beside it (§36).
16.2 Decisions
| ADR | Decision | Reason |
|---|---|---|
| ADR-KIN-01 | One model, three tiers (referee, batch-cpu, gpu-*), and a ledger of their disagreement. |
A fast answer is useful only if its distance from a trusted answer is a measured number. The CPU reference is permanent, so a GPU world can always be re-run where it can be inspected. |
| ADR-KIN-02 | Two step disciplines, chosen by tier. The referee and the strict profile of batch-cpu run a transaction per accepted interval with located events, a commit fence and retries that restore all state. The throughput tiers keep a fixed step and a lockstep convex solve. |
Where a result adjudicates another, a controller must never observe a half-step and a tunnelled contact must be a detected refusal. Where throughput is the purpose, adaptive splitting would destroy the static launch graph (§21) and the fixed shapes that make batches fast. |
| ADR-KIN-03 | A world's trajectory is a function of its address and identities. The batch it ran in is not an input. | Replay of one world outside its batch, mixed cohorts (§17), slot-level dispatch and paired statistics all need it. It is a tested invariant (INV-KIN-04), not a hope. |
| ADR-KIN-04 | Promotion returns an estimate of the higher-tier value with a correction and an interval, as well as a label. | A pass or fail label discards the measured tier bias. Audit draws are addressed, so audit probabilities are known exactly and the correction is unbiased by construction (§16.8). |
| ADR-KIN-05 | Physical parameters and numerical parameters are declared separately, and only numerical ones may change in a retry, a refinement study or a model-form variant of type "step" or "tolerance". | A smaller time step must not silently stiffen a fitted foot spring. A stiffness that scales with h is a numerical parameter disguised as a physical one, and it makes step-halving tests meaningless. |
| ADR-KIN-06 | The referee's law is specified in this plan (CL-H, §18.5) and FrankenSim is the implementation consumed through the airlock. Where FrankenSim's limits bind, the system refuses by name and routes the adjudication to the strict profile, recording which adjudicator answered. |
Sibling limits are facts at the pinned commit, each with an interim posture until its upstream work item lands (§16.4). A specification that lives here lets two implementations be compared and lets an upstream change be admitted by a re-run instead of an argument. |
| ADR-KIN-07 | The mass matrix is small and dense per world for 18 to about 60 degrees of freedom. A tree-operator route serves larger mechanisms F. | At these sizes a dense Cholesky is about 2,000 to 14,000 flops (n³/3 at 18 and 35) and, as a HYPOTHESIS that the ledger of §16.23 measures, under a tenth of the step; sparse bookkeeping costs more than it saves. Above about 60 the cubic term and padding in mixed cohorts (§17.4) reverse that. |
16.3 One model, three tiers
| Tier | Purpose | Arithmetic | Foundation | Worlds |
|---|---|---|---|---|
Referee (referee) |
The most trusted answer we can compute. Adjudicates the others. | f64, deterministic elementary functions, exact Coulomb friction, certified collision queries |
fs-mbd, fs-contact, fs-time, fs-query through the airlock, driven by the transaction of §16.5 |
One to a few dozen |
Batch CPU (batch-cpu) |
The permanent reference for the batched algorithm. Runs anywhere, including where there is no GPU. Debugging target for a teleported world. | f64 or f32; the same derived kernels and the same stage order as the GPU |
frob-kernels in Rust on fs-exec tiles |
Hundreds to thousands |
GPU (gpu-cuda, gpu-metal) |
Throughput for training and for design sweeps. | f32 |
The same kernels in the shared dialect (§21) | Thousands to tens of thousands per device |
The referee is not called ground truth. Its outputs carry the evidence class FrankenSim's contracts give them, and Reality is a separate claim (§8).
A tier runs under a step profile. The profile is part of the step-plan identity that every result carries
beside its ModelId (§16.6).
| Profile | Tiers | Step discipline | Law | Use |
|---|---|---|---|---|
lockstep |
batch-cpu, gpu-cuda, gpu-metal |
Fixed step h, ten stages, a fixed number of Newton iterations under masks | CL-1 by default; CL-2 and CL-H as Frontier options (§18.5) | Training, sweeps, design populations. On batch-cpu it is the bit-level reference for the GPU. |
strict |
batch-cpu only |
The transaction of §16.5 in f64, one world at a time per worker, events located |
CL-1 with located events; CL-H once GATE-TRAC-09 passes | Adjudication where the referee cannot complete; the "step" and "law" arms of model-form checks; debugging |
referee |
referee only |
The transaction of §16.5 with certified queries | Exact Coulomb through FrankenSim's solve, specified as CL-H | Adjudication |
A request for strict or referee on a GPU tier is refused (FROB-REF-222). DESIGN, reason: a profile is a
promise about what the step did, and a GPU cannot keep this one without giving up fixed shapes.
16.4 The referee as it can exist at M1, and where FrankenSim's limits bind
FACT: fs-mbd provides floating-base forward dynamics and a contact step
(articulated_contacts::step::step_articulated_frictional_contacts) that is first-order kick-contact-drift, takes
the contact set from the caller, and states that "independent endpoint gap checks are not continuous collision
detection". A 29-DoF G1 rollout at 480 Hz with Hunt–Crossley feet and fs-tribo friction already runs
(frankensim/crates/fs-cmaes-viz-wasm/src/g1_walking.rs). The M1 referee is therefore: FrankenSim's articulated
dynamics and contact solve, plus contact-set generation that frob-contact supplies from fs-query convex
separation and signed-distance queries, plus the transaction of §16.5 around them, plus an audit pass with
certified continuous collision detection on sampled windows. It improves as FrankenSim lands its persistent
steppers (upstream work items FD-FSIM-02 and FD-FSIM-04).
The limits below are FACT at FrankenSim commit 483d09c682, read on 2026-11-21, and each is re-checked by an
admission test at INT-FSIM-01 before a gate cites it. Each has an upstream work item in Appendix F, done in
FrankenSim's repository under its rules; the posture in the last column holds until it lands.
| Limit | Where it is stated | What it costs the referee | Posture until the upstream change lands |
|---|---|---|---|
| At most 21 contacts in one solve | fs-contact/src/sustained/solve.rs: pub const MAX_CONTACTS: usize = 21, also the default budget of the frictional multi-contact solve |
A single robot is one island, because every foot couples through the floating body. A sole with eight samples per foot fits when standing (16) and overflows in a fall or a crawl. A tracked robot with patch samples along two belts overflows in normal driving. | A named capacity refusal (FROB-REF-721), never a truncated contact set. The caller's policy then chooses: a documented patch reduction with its qualification fixture (§18.11), or the strict profile as adjudicator. |
| No persistent contact time stepper | The Moreau–Jean and barrier lanes are open beads without code; solve_frictional_contacts is an event solve |
Time stepping, warm starts, friction memory and the commit fence are ours | frob-dyn owns the transaction. Upstream work item FD-FSIM-02. |
| Fixed material witnesses and a fixed contact frame in the sustained step | fs-mbd/src/articulated_contacts/step/sustained.rs: "Body pairs retain their ORIGINAL material witnesses and fixed contact frame"; tangentially separating witnesses are refused |
A wheel rolling over a sharp lip, a sliding sole and a foot rocking heel to toe all move their witnesses | frob-contact migrates witnesses between calls (§18.4), so each call sees an interval short enough that its witnesses are fixed inside FrankenSim's own gate. If the gate still refuses, the refusal is surfaced and the strict profile adjudicates. |
| Contact set supplied by the caller; endpoint gap checks only | articulated_contacts::step contract |
Contact generation and tunnelling detection are ours | frob-contact generates sets; fs-contact::certified_ccd audits the path. Upstream work item FD-FSIM-03. |
| Joint classes limited to fixed, revolute, prismatic and helical | fs-mbd JointKind |
Spherical, universal and planar joints are lowered to products of one-axis joints | Upstream work item FD-FSIM-01. |
Platform sqrt, sin and cos; no cross-ISA claim |
fs-mbd |
The referee claims same-platform determinism only | Upstream work item FD-FSIM-07 (§35). |
Because these limits decide how often the referee can answer at all, the share of scenarios it completes is a reported figure with a gate (§16.11, GATE-FID-03), not an assumption.
16.5 The transactional step S
The referee and the strict profile advance each world by one transaction per accepted physics interval
(one base interval, defined below). A transaction either commits a complete, checked interval or leaves no
trace.
What is in the restore set. A checkpoint taken at the start of a transaction, and restored in full on every retry, contains: the physical state (q, v); actuator, thermal and battery states; contact memory (witness pairs, contact frames, tangential memory, the warm-start impulses and the active set); per-contact film thickness (§18.8); clock phases and packet queues (§20); controller and estimator states; the ordinals of every sample-stamped random draw; terrain edits; the accumulators of the energy ledger; event counters; and the command cursor. Omitting any of these makes a retried interval differ from a first attempt, which INV-KIN-02 forbids.
The fence. Committed state carries a monotone commit_epoch. Controllers, sensor samplers, telemetry and
snapshots read a named epoch and nothing else. A candidate interval is built beside the committed state and
becomes visible only when the epoch advances, in one step. On the lockstep tiers the same rule holds at the
kernel-batch boundary: a host reader sees a completed batch or the previous one.
Time. Time is rational (§11). The physics grid and the clock boundaries of §20 together cut time into base
intervals [t0, t1) of length ℓ₀ ≤ h. That partition is fixed by the schedule and never by the state. Inside a
base interval, sub-intervals are dyadic: a split may fall only at t0 + ℓ₀·k/2ᴸ for L ≤ L_max. DESIGN:
L_max = 16, which at ℓ₀ = 2 ms resolves 31 ns, so an approach at 3 m/s is located to within 0.1 µm. An event that
cannot be bracketed without exceeding L_max, or more than E_max = 64 events inside one base interval, ends the
transaction with FROB-REF-220 instead of an approximation (the guard against accumulation points of impacts).
Both constants are OPEN-KIN-5.
DESIGN transaction(X_committed, [t0, t1)) # t1 = min(t0 + h, next clock boundary of §20)
T1 S ← checkpoint(X_committed) # the whole restore set above
T2 for attempt in 0 ..= N_retry: # N_retry = 3; attempt selects a numerical plan only
T3 X ← restore(S); work ← [(t0, t1)]
T4 while work is not empty:
(a, b) ← work.pop_front()
T4.1 read pending actions, external forces and the noise increment for (a, b) # refinement law below
T4.2 broad phase on swept bounds over (a, b); narrow phase at a returns, per candidate pair:
features, normal, witnesses, material pair, signed gap g and a bound ε_g on its error (§18.4)
T4.3 t* ← earliest time in (a, b] at which a guard may change sign:
touchdown, lift-off, stick↔slip, joint limit, backlash closure, actuator protection, clock boundary.
Conservative advancement on gap ÷ bounded approach speed; bisection on the dyadic grid;
on the referee, fs-time event location and fs-contact certified CCD.
T4.4 if t* < b: (a, b) ← (a, t*); work.push_front((t*, b_old))
if the split would pass L_max or the event count would pass E_max: refuse FROB-REF-220
T4.5 advance smooth or sustained-contact dynamics from a to b with the active set held fixed
T4.6 at b, if an event fired: evaluate every event row on the one common pre-event state; group the
events that fall in the same finest dyadic cell; partition bodies and rows into independent islands;
solve the coupled impulse once per island; apply reaction impulses to dynamic props; credit
boundary work from prescribed surfaces to the energy ledger
T4.7 invalidate affected witnesses, redetect, migrate tangential memory (§18.4)
T5 checks: sub-intervals sum to (t0, t1) exactly; solver residuals; joint limits and closures inside
tolerance; group-membership error; energy-ledger residual; path admission (no missed contact on the
swept path, inside ε_g); loads and currents inside admitted domains; event count
T6 if every check passes: COMMIT (advance commit_epoch; publish state, impulses, sensor inputs,
diagnostics and the attempt index together) and return
T7 after the last attempt: restore(S); mark the world Faulted with cause numerical-fault (FROB-REF-221);
publish nothing from the failed interval
Retries change numerical parameters only (ADR-KIN-05). DESIGN, the ladder: attempt 0 is the declared plan; attempt 1 halves the largest admitted sub-interval and forces bisection for event location; attempt 2 tightens the solver tolerance by ten, doubles the iteration cap and clears the warm start; attempt 3 switches the island solver (semismooth Newton to the fixed-point cone iteration of §18.5). The attempt that committed is part of the result. The ladder is a deterministic function of state, so a replay repeats the same attempts.
A numerical failure is not a physical failure. A world that exhausts its retries ends with
numerical-fault, which the failure taxonomy (§29) keeps apart from fall, slip or collision. A design is
never scored as having fallen because a solver gave up, and summary statistics state how many worlds were
excluded for this cause (§35).
Refinement law for random inputs. A random input drawn over an interval must mean the same thing when the
interval is split. A sensor draw scheduled at a sample instant keeps its original stamp and is not affected. A
driver that acts over the interval (a Brownian disturbance force, process noise in an actuator) is defined on the
dyadic tree of the base interval: the root increment over length ℓ₀ is S = √ℓ₀·Z₀; a node with increment S over
length ℓ gives its left child S/2 + √(ℓ/4)·Z and its right child S minus that, with Z a standard normal keyed by
(WorldAddr, base_interval, stream_id, level, index) in the counter-based streams of §35. Each child then has
variance ℓ/2 and the two are uncorrelated, which is the law of Brownian increments. Because every node is a pure
function of its address, no noise state needs storing: a retry, a restart from a checkpoint and a migration to
another device all read the same path, and a retry can refine the path but can never redraw it.
Counterexample, hand-checkable: let the input be a fair sign, and let a step be rejected whenever the sign is negative and then redrawn once. The accepted sign is positive with probability ½ + ½·½ = ¾. Rejection that redraws biases the law; this fixture is in the negative-control registry.
A driver with no refinement law registered is refused under a transactional profile (FROB-REF-223). A
higher-order stochastic integrator would need a joint law for its iterated integrals and is not offered.
Coupled impulses per island. Simultaneous events are solved together, never in an imposed order, because the order can change the outgoing velocity and the basin of a gait. Trot, bound and pronk have simultaneous touchdowns by symmetry and are ordinary inputs, not refusals. With M the mass matrix, J the event rows and A any independent bilateral rows that hold through the event, the contact-space inverse mass is W = J·P·Jᵀ with P = M⁻¹ − M⁻¹Aᵀ(A·M⁻¹·Aᵀ)⁻¹A·M⁻¹, and W = J·M⁻¹·Jᵀ when there are none. For a moving bilateral constraint A·v⁺ = b the freely predicted velocity v_f is first corrected by M⁻¹Aᵀ(A·M⁻¹·Aᵀ)⁻¹(b − A·v_f), and the tangent response P·Jᵀp is added to the corrected velocity.
Counterexample, hand-checkable: two 1 kg sliders joined by a rigid tie reach a wall together at 1 m/s. Stopping both needs a wall impulse of 2 N·s. Solving the wall contact for the first slider alone gives 1 N·s; imposing the tie afterwards leaves 1 kg·m/s shared by 2 kg, so both move into the wall at 0.5 m/s.
Redundant contacts can admit several impulse distributions with the same motion. The transaction compares and
reports observable post-event motion, total impulse and work, and never reports a pressure distribution inferred
from one resultant wrench. A configuration with no solution or several distinct motions (the Painlevé rod of
§18.16) is reported as indeterminate (FROB-REF-224).
Units. Impulses are in N·s and forces in N. A tolerance is typed by the quantity it bounds; one threshold applied to both is a dimension error that the kernel type checker rejects (C-14).
16.6 The batched step S, with per-world variants
Tiers batch-cpu and gpu-* under the lockstep profile execute the same ten stages in the same order. Every
stage is a named kernel or a short chain of kernels (§21), batched over worlds, with no data-dependent control
flow across worlds.
| # | Stage | Kernels | Parallel shape | Per-world data it reads |
|---|---|---|---|---|
| 1 | Kinematics: link poses and twists down the tree | K-fk-level, K-twist-level |
Worlds × links at one depth | Model slot: parents, level ranges, joint placements, screw bases |
| 2 | Collision geometry: world poses of primitives and sample points | K-geom-pose |
Worlds × primitives | Model slot: sample lists |
| 3 | Broad phase for dynamic pairs | K-bp-sort, K-bp-pairs |
Per world, deterministic sort | Model slot: pair filter |
| 4 | Narrow phase: primitive pairs; samples against the scene distance field; thin primitives | K-np-prim, K-np-sdf, K-np-thin |
Worlds × candidates | Scene slot |
| 5 | Contact compaction: count, exclusive scan, write | K-compact |
Fixed capacity per world, overflow flagged | — |
| 6 | Dynamics terms: mass matrix, bias forces, actuators, passive forces | K-crba-level, K-rnea-level, K-act-l1, K-passive |
Worlds × links | Model slot: inertias, actuator-card slots; parameter row |
| 7 | Constraint rows: contacts, limits, closures, joint stiction and backlash, cones, film friction | K-jac-rows, K-joint-rows, K-cone-setup, K-film |
Worlds × rows | Interface cards; parameter row |
| 8 | Solve: diagonal scaling, then Newton on the convex objective with Cholesky and exact line search | K-scale-diag, K-hess, K-chol, K-linesearch, K-cone-eval |
Worlds in lockstep under masks | — |
| 9 | Integrate on the group; accumulate the energy ledger | K-cay-step, K-normalize, K-ledger-step |
Worlds × joints | — |
| 10 | Sensors, observations, rewards, resets under masks (semantics in §23) | K-sense-*, K-obs, K-reset |
Worlds | Task slot |
Design rules for the step:
- Fixed shapes. Every buffer has a capacity fixed at environment creation. A world that overflows its contact
or row capacity is marked
Degradedwith a reason; it is never undefined behaviour and never a silent truncation. The caller chooses refuse, reset or continue-flagged. - Lockstep solve. All worlds run the same number of Newton iterations under a per-world
donemask. This keeps the launch graph static (§21) and makes the iteration count part of the identity every result carries. - No floating-point atomics. Compaction and reductions use count-scan-write and fixed-order trees (§35).
- The mass matrix is small and dense per world. For the robots in scope (18 to about 60 degrees of freedom) H = A + JᵀGJ is factored by a batched dense Cholesky after diagonal scaling (§18.12).
- A tree-operator route for larger mechanisms F. For a contact between one link and the environment, the Jacobian is nonzero only on that link's ancestors, so JᵀGJ has the same branch-induced sparsity as the mass matrix and an LᵀL factorization ordered leaves-first has no fill. A self-contact between two limbs breaks the pattern and routes that world's step to the dense path under a flag. Pilot before funding: count factor flops on the reference robots and on the 64-DoF bucket of §17.4 with and without the tree order. Ceiling: HYPOTHESIS, the factor is under a tenth of our step at 18 to 35 degrees of freedom. An incumbent trace puts constraint assembly at 15% and the whole solve at 47% of its step without separating the factor, so the share is measured, not assumed: the ledger of the ancestor-sparse route (§16.23) decides whether the route applies at 18 to 35 degrees of freedom or only to padded buckets and mechanisms above 60. ✖ KILL-TREEOP: whole-step gain below 1.15 at 64 degrees of freedom on the M2 profile; fallback, the dense path everywhere and a cap of 64.
Identities. A result carries its world's own ModelId and the StepPlanId of the plan that ran it. The
StepPlanId is the content hash of: the shape class (the capacities below), kernel versions, contact law, step
profile, h, the number of world-steps per control interval, Newton iterations and tolerance. A physics change
changes one of the two identities; nothing else does.
Per-world parameter variants S. Each world carries a model_slot, a scene_slot, a task_slot and an
optional parameter row, exactly as it already carries its address. Worlds whose slots point at models with the
same tree and different parameter values (link lengths, inertias, gear ratios, actuator-card parameters, foot
radii, friction) run in one batch with no masks and no padding. The parameter row carries per-world
randomization drawn from card uncertainty (§23). FACT: a leading incumbent already batches model fields per world
(MuJoCo Warp "Batched Model Fields", Appendix H), so this is parity with incumbent practice and not a
differentiator.
Per-world model variants F. Worlds whose slots point at models with different trees run in one batch inside padded capacities, the way a world already indexes one of K scene variants. The tables, the size buckets and the conditions under which this pays are §17. Kinesis guarantees one thing to §17: INV-KIN-04, that padding and co-tenants change no world's digest.
Why padding can be exact. Padded links are identity transforms with zero inertia, appended after the real ones.
Padded degrees of freedom get identity rows and a zero right-hand side, and a Cholesky factor's leading block does not
depend on later rows, so the real block of the padded factor equals the unpadded factor bit for bit (PO-KIN-02).
Reductions over a padded axis use pairwise trees aligned to powers of two with the real entries in a prefix: any
subtree of trailing −0.0 padding sums to −0.0, and x + (−0.0) = x for every x, so the tree over capacity P reduces to
the tree over the smallest power of two that holds the real entries, whatever P is (PO-KIN-03). The canonical zero is
−0.0, as §35.5 requires for sums: +0 is not neutral, because (−0.0) + (+0.0) = +0.0 changes the bits of a sum that is
−0.0. Counterexample that motivates the rule, in f32: with a = 10⁸, b = −10⁸, c = 1, (a + b) + c = 1 and a + (b + c)
= 0, because the spacing near 10⁸ is 8. A reduction whose grouping depended on capacity would make a world's bits
depend on its neighbours.
Capacities per reference workload. DESIGN defaults, TARGETs until the M2 battery records overflow rates.
C is the active-contact capacity; R is the row capacity, four rows per contact (normal, two tangential, spin;
the spin row is masked where the effective radius is zero) plus joint rows (one limit row and one stiction row
per actuated joint), rounded up to a multiple of 8. Bytes are per world in f32 with a dense Jacobian, from
4·(61·n_b + 17·n_v + 3·n_v² + 24·C + R·n_v + 12·R + 10·n_v), before a few kilobytes of state.
| Workload (§36) | n_v | Links n_b | Samples tested | C | R | Bytes per world | Fits the referee's 21-contact cap |
|---|---|---|---|---|---|---|---|
| Quadruped, 12 joints, 16 foot samples and 8 body samples | 18 | 13 | 24 | 32 | 160 | 31 kB | Standing and gaits: yes (at most 16). Falls: no. |
| Hexapod, 18 joints | 24 | 19 | 32 | 32 | 168 | 41 kB | Tripod gait: yes. Six feet down with four samples each: no (24). |
| Humanoid, 29 joints, 8 samples per sole, hands and body | 35 | 30 | 40 | 64 | 320 | 92 kB | Standing and walking: yes (at most 16). Hands down, crawl, fall: no. |
| Wheeled base, four wheels with rim samples | 10 | 7 | 32 | 16 | 72 | 12 kB | Yes on flat floors; thresholds and ramps approach the cap. |
| Wheel-leg quadruped | 22 | 17 | 36 | 32 | 160 | 37 kB | Rolling: yes. Stair transitions: marginal. |
| Tracked skid-steer with two flippers | 12 | 8 | 80 | 64 | 264 | 36 kB | No, in normal driving. |
The quadruped and humanoid rows are the reference workloads Q and H and take the capacities of §36.3, where Q's 152 rows are raised to 160; with the state, warm starts and solver vectors of §36.10 they are about 35 kB and 97 kB.
FACT (a derived estimate, Appendix H): per-world state is about 20 to 53 kB for a quadruped and 56 to 180 kB
for a humanoid over row capacities of 60 to 280 and 125 to 700; the defaults above sit inside those ranges. The
dense Jacobian is 24 to 49% of the bytes in every row of the table, so row capacity, not world count, decides
memory. The path table of §14 (each contact row touches only the degrees of freedom on its support path) replaces
R·n_v by about R·(6 + depth) and is the default once GATE-KIN-05 shows identical digests with it.
TARGET: under the default capacities, Degraded world-steps are at most 10⁻⁴ of all world-steps per robot class
on the standard battery (§38), and every one is flagged (GATE-KIN-07). OPEN-SPARK-3 owns the capacities of Q and H
(§36.3); OPEN-KIN-3 owns the other rows.
16.7 Integrators
- Default, S: semi-implicit Euler. Stage 8 returns v₊; stage 9 sets q₊ = q·τ(h·v₊). Joint damping and reflected inertia are implicit in A. Default physics step h = 2 ms with control every tenth step (OPEN-KIN-1).
- Structure-preserving, F: a linear-time variational integrator for articulated bodies in reduced
coordinates on
refereeandbatch-cpu, for long-horizon energy studies and passive-dynamics work. - World-steps per control interval are fixed at admission by the clock contract of §20 (rational periods
with phase accumulators) and enter the
StepPlanId. No schedule is required to be an integer ratio of the physics tick: on the lockstep tiers a period that is not a multiple of h is served by its phase accumulator under the boundary ordering rule of §20, with the quantization recorded; under a transactional profile every clock boundary is an exact sub-interval boundary (step T4.3).
The free-body variational step S, the concrete content of the structure-preserving lane. For one rigid body with configuration g, body velocity ξ, inertia 𝕄, retraction τ = cay and discrete Lagrangian L_d = (h/2)·ξₖᵀ𝕄ξₖ − (h/2)·(V(gₖ) + V(gₖ₊₁)), the step is a symmetric kick-drift-kick:
- Half kick: μ⁺ = μ⁻ + (h/2)·f_pot(gₖ) + (h/2)·uₖ.
- Solve the discrete Legendre relation for ξₖ: the momentum that the left-trivialized derivative of τ at −h·ξₖ (equally, the right-trivialized derivative at +h·ξₖ; the matrix of C-36) assigns to 𝕄·ξₖ equals μ⁺. Newton with the closed-form 6×6 Jacobian; TARGET two to three iterations to 10⁻¹³ relative.
- Drift and transport: gₖ₊₁ = gₖ·τ(h·ξₖ), and μ⁺ is carried to the new body frame by the wrench pull-back μ° = (Ad_c)ᵀ·μ⁺ of §11.3 (the group coadjoint of c⁻¹, not of c), where c = τ(h·ξₖ), with the sign convention of §11 (the one for which μ̇ = ad*_ξ μ holds for a free body).
- Second half kick with f_pot(gₖ₊₁) and uₖ₊₁.
What is claimed: second order in h, time-reversible to round-off, exact conservation of the discrete momentum map
in the absence of potential and input. What is not claimed: any symplectic or energy-drift property for an
articulated tree with a configuration-dependent mass matrix, where an explicit splitting is not symplectic, or
for any trajectory with dissipative contact impulses, where the property has no meaning. The articulated lane
stays F until GATE-KIN-01's long-horizon rows pass on a tree. Pilot before funding: the asymmetric free body
and the heavy top under this step against the Cayley semi-implicit step at h = 2 ms and 0.25 ms; falsifier, an
energy error that does not fall by a factor between 3.6 and 4.4 per halving of h. Ceiling: none on throughput;
the lane exists for trust in cost of transport over minutes of simulated time. ✖ KILL-VARIATIONAL: the implicit
solve of step 2 costs more than twice a semi-implicit step on batch-cpu, or the articulated form fails its
momentum row after one repair cycle; fallback, the free-body step is kept as a known-answer oracle and every
rollout uses the default.
16.8 The trust ladder, the fidelity ledger and the promotion rule
The fidelity ledger is a table, keyed by (model class, scene class, quantity, tier pair, profile pair, contact-law pair, kernel versions), of the distribution of disagreement between a faster configuration and a slower one on the standard battery (§38). Quantities include base displacement, yaw, stance impulse, slip distance, peak joint torque, energy per metre, time of first failure, the count of contacts the certified detector found and the fast tier missed, the energy-ledger residual, and task decisions: whether the two configurations agree on success, on the terminal class and on the verdict of a comparison (the decision-level equivalence of §35). A state error inside its envelope with a different decision is a disagreement.
Promotion rule. A design score or a policy return computed on a fast tier is labeled provisional. It
becomes established only after re-evaluation of a sample of episodes on a higher tier, with the difference
inside the promotion envelope recorded in the ledger for that quantity. The sample size is chosen by the
e-process of C-15, so clear cases promote after a handful of referee episodes. A result that fails promotion is
labeled provisional: tier-sensitive and reported with both numbers (NS-09).
This is FrankenSim's "speculate low, verify high" pattern and its Goodhart guard (addendum proposals 9 and D, FACT), applied to robots.
The rule has two parts, both S. The first decides the label; the second produces the number that is reported.
The label: a promote and reject pair on exceedances. Each audited episode yields one bit: did the fast and the
higher tier disagree beyond the ledger's frozen envelope for the quantity. Two e-processes run on those bits with
rates p₀ < p₁ fixed in advance: E⁺ tests "the exceedance rate is at least p₁" and promotes when it crosses 1/α⁺; E⁻
tests "the rate is at most p₀" and labels the result tier-sensitive when it crosses 1/α⁻. For n audited pairs with k
exceedances, ln E⁺ = (n − k)·ln((1 − p₀)/(1 − p₁)) + k·ln(p₀/p₁). Worked values, illustrative: with p₀ = 0.05, p₁ =
0.30 and α⁺ = 0.01, a clean episode multiplies the evidence by 1.36 and an exceedance divides it by 6, so promotion
takes 16 clean pairs, or 22 with one exceedance; at α⁻ = 0.05, two exceedances in the first three pairs reject. Both
compile to an integer table (n, k) → promote, reject or continue, regenerated and compared in CI. The envelope is
hashed into the ledger before sampling, and episodes are chosen by address before outcomes exist; under those two
conditions the error rates hold at any stopping time by Ville's inequality. Independence of episodes is not needed if
the null is read conditionally: given the earlier audited pairs, each new pair exceeds with probability at least p₁
(for E⁺) or at most p₀ (for E⁻). Budgeted mode (DESIGN): 32 pairs, after which the result stays provisional.
Deterministic fallback when a ledger stratum is missing or the null self-test fails: a fixed 30 pairs, promoting on at
most two exceedances.
The number: a doubly robust tier correction. For N addresses on the fast tier, let Y⁰ᵢ be the fast outcome at address i (for a comparison, the paired difference B − A at that address).
- Before any higher-tier outcome exists, fix a prediction mᵢ = ĝ(Y⁰ᵢ) of the higher-tier outcome. ĝ is the affine map c + β·Y⁰ᵢ that the ledger holds for that quantity and stratum, and the identity (β = 1, c = 0) in a new stratum (§29.11); it is frozen and hashed first.
- Fix an audit probability pᵢ ≥ p_min for each address and log it. Draw Aᵢ ~ Bernoulli(pᵢ) from the stream
keyed by
(WorldAddr, promotion_id). The assignment is persisted before dispatch and survives retries. - Where Aᵢ = 1, run the same address on the higher tier (same scene draw, disturbances, randomization and policy) to get Y¹ᵢ.
- Form Zᵢ = mᵢ + Aᵢ·(Y¹ᵢ − mᵢ)/pᵢ and report the mean of the Zᵢ with a confidence sequence valid at any
stopping time (
fs-eprocprovides mixture confidence sequences, FACT). The weights 1/pᵢ are bounded by 1/p_min = 20 without clipping: the floor p_min = 1/20 bounds every weight, so nothing is clipped and there is no clipping bias (§29.11, which owns the ledger).
Because pᵢ is set by the system and not estimated, the expectation of Zᵢ given everything earlier equals the expected higher-tier outcome whatever ĝ is: a poor prediction widens the interval and cannot bias it (PO-KIN-05). With mᵢ = 0 the estimate is the plain inverse-probability mean; with pᵢ = 1 it is the plain paired mean, which is the regression test. The higher-tier episodes needed for a given interval width scale by about (1 − ρ²) against a higher-tier-only estimate, ρ being the correlation of paired fast and higher-tier outcomes: 1.3 times fewer at ρ = 0.5, 2.0 at 0.7, 2.8 at 0.8, 5.3 at 0.9, 10.3 at 0.95. HYPOTHESIS: ρ for whole contact-rich episodes is high enough to matter. Contact chaos decorrelates tiers, ρ is unmeasured, and it is reported separately for continuous metrics and for binary outcomes (H-22b). Test: ρ on 200 addresses of the standard battery for energy per metre, displacement over 2 s and success, from paired data that GATE-FID-01 already pays for.
Variants, each labeled in the receipt: a three-level telescoping estimate E[Y₂] = E[Y₀] + E[Y₁ − Y₀] + E[Y₂ − Y₁]
over gpu-*, batch-cpu in f64 and the referee, with samples allocated N_ℓ ∝ √(V_ℓ/C_ℓ) from pilot variances
(FACT: fs-uq/src/mlmc.rs and product_mlmc_allocation.rs implement the allocation for a per-call closure; a
batch entry point is an upstream work item, FD-FSIM-17); and a segment mode that re-simulates one-second
segments from fast-tier snapshots when whole episodes decorrelate. Segment mode estimates per-segment tier bias,
which is a different quantity from the episode outcome, and its reports say so.
Budgeted mode (DESIGN): 256 higher-tier episodes per promotion; on exhaustion the result stays provisional. Fallback
trigger: effective sample size below 30% of the audited count, or ρ below 0.7 for the quantity; then promotion
uses the plain paired e-process. ✖ KILL-DR records the second case as a decision (§16.17).
Labels a result can carry. provisional; established, with the adjudicator named (referee or
batch-cpu:strict); provisional: tier-sensitive; model-conditional (§16.9); energy-created (§16.10);
capacity-biased (§18.11). Citing any label but established as established is refused (FROB-REF-820). The
weakest label among a result's inputs is the label of the result, by the claim product of §8.
16.9 Rank agreement between tiers, and model-form rank checks
A fidelity envelope bounds how far a number moves between tiers. A design studio also needs to know whether the order of designs moves. Both checks below write to the fidelity ledger and both have consequences.
Rank agreement between tiers F. For a set of at least 40 designs in one robot class, drawn by a keyed stream before any score exists, and one fixed controller, each design is scored on the fast tier and on the higher tier at the same addresses. The record holds Kendall's τ, a top-weighted τ, top-k recall for k = 10 and bootstrap intervals, per stratum of design space, because a cheap rung is biased toward designs that are easy to control and the bias is not uniform. Consequences, TARGET thresholds on the lower interval bound:
| Top-weighted τ | Top-10 recall | What the fast tier may do in that stratum |
|---|---|---|
| ≥ 0.7 | ≥ 0.8 | Select finalists. Every design within 2·q̂ of the k-th best fast score still goes up, q̂ being the conformal residual quantile of the tier pair. |
| 0.5 to 0.7 | any | Discard only. It may drop the clearly dominated and may not choose among the rest. |
| < 0.5 | any | Not used for ranking. Every candidate that survives static analysis is scored on the higher tier, and the campaign estimate says so before it starts. |
The same record type serves the controller rungs of the fidelity ladder (baseline controller, family policy,
fine-tuned policy, specialist), which §28 allocates; Kinesis owns the tier-pair rows. HYPOTHESIS: tier pairs
agree far better than controller rungs, since they share the controller. No published rank agreement between
rungs exists (Appendix H), which makes this measurement the first thing the ladder depends on. Pilot before
funding: the 40-design calibration set at M3 on batch-cpu against the referee. Ceiling: a rung that only
discards saves at most the fraction it discards. ✖ KILL-RANK: τ below 0.5 in more than half the strata for the
reference classes; fallback, tier pairs are used for promotion of individual results only, never to order a
population.
Model-form rank checks S. A ranking of finalists is re-evaluated under variants that should not matter to a
physically meaningful preference. Each variant changes one thing and is a different StepPlanId:
| Variant family | Arms | What a flip means |
|---|---|---|
| Contact law | CL-1 against CL-2; against CL-H where admitted (§18.5) | The preference depends on compliance or on the associated-flow artifact of the convex law |
| Cone | The elliptic cone of §18.6 against a four-sided pyramid of the same μ | The preference exploits diagonal traction: a pyramid grants √2·μ along its diagonals, 41% more than the cone |
| Stiffness | Contact stiffness ×½ and ×2, dissipation time held | The preference depends on a penalty parameter |
| Step | h and h/2, physical parameters held (ADR-KIN-05) | The preference is a discretization artifact |
| Tolerance | Solver tolerance ÷10, or twice the Newton iterations | The preference lives in solver residual |
The check runs on the top k and the runners-up of a final round (DESIGN default k = 8) and returns
model-robust { variants, top_k_overlap, energy_audit } when the top-k overlap is at least k − 1 under every
variant and the energy audit is clean, and model-conditional { flipped_by } otherwise; on results the two
variants are the labels model-robust and model-conditional. Consequences of
model-conditional: the ranking is reported as fidelity-dependent with both orders; the result is labeled
model-conditional and cannot be promoted to established; it cannot authorize a build packet or spend
prototype budget (§31); the cheapest experiment that would settle it is named in next_actions. A preference
between two designs is stated only if it holds under both qualified contact laws. TARGET cost: at most 10% of a
campaign's compute. Demotion rule: if over 20 campaigns the overlap is at least 7 of 8 in 95% of them with no
energy flag, the check becomes a periodic audit.
16.10 The audit for energy created by the simulator S
Every step closes a discrete work–energy balance: the change in mechanical energy, minus actuator work, minus work by prescribed surfaces and external forces, plus modeled dissipation (contact damping, sliding friction, joint damping, stiction), leaves a residual. The energy ledger (§29) already reports that residual on its own line. The audit adds a sign and a consequence: a positive residual is energy the integrator or the solver created.
- Stage 9 accumulates the residual per world in double-float (
K-ledger-step), with the positive part kept separately and tagged by the contact mode in which it arose. - Energy stored in contact compliance is stored energy, not creation: for CL-1 a penetration d against stiffness k holds ½·k·d² and can return at most that. The audit counts it as storage, as the ledger does heat stored in a modeled thermal mass.
- A result carries the flag
energy-createdunder the one rule of §29.9, which owns the window and the denominator: created energy above 1% of the absolute work exchanged in any 0.5 s window on a fast tier, or above 10⁻⁶ relative on the referee (TARGET; re-baselined once at M2 with GATE-FID-01). This section defines no other label or threshold. A result flaggedenergy-createdcannot beestablished, and a ranking that contains one ismodel-conditional.
Counterexample, hand-checkable, and the first planted exploit: a position-level penetration correction that removes a depth d in one step gives the body a velocity d/h from nothing. With m = 1 kg, d = 1 mm and h = 2 ms that is 0.5 m/s and 0.125 J per event. The compliant law on the same penetration at k = 10⁵ N/m stores 0.05 J and returns less. A hopping policy that finds the first behaviour gains height from the solver. The second planted exploit is the pyramid cone of §16.9. GATE-FID-06 requires both to be flagged and an honest policy to pass.
16.11 The background referee and its coverage figures F
What it is. While a GPU run is in progress, idle CPU cores re-run a sample of its worlds on a higher tier and compare. It is a correctness guard. It removes no training and no search, so it is not a source of speed, and it is specified on every back end: it needs a snapshot, not shared memory.
- Sampling. Worlds are chosen by a hash of their address under a dedicated counter-based stream with a declared probability p, so estimates weight each sampled world by 1/p and no physics stream is touched.
- Snapshots. State is copied at the fence into a snapshot buffer: an explicit readback on CUDA; on Metal, a blit to a snapshot buffer read after command-buffer completion. FACT: the integration state of a 35-DoF humanoid is about 135 floats, so shadowing 256 worlds at 50 Hz moves about 7 MB/s; copy cost is irrelevant on either machine and "zero copy" is not part of the design.
- What is compared. Trajectories are not compared beyond a few steps, because any difference grows through contact. Two things are: (a) one-step defects, where the CPU advances one control interval from the snapshot and the difference in state, impulses and active set is binned by stage, which localizes a kernel or a law; and (b) outcome distributions, where whole episodes are re-run from reset at the same addresses with the policy weights frozen at the snapshot, and success rate, energy per metre and terminal class are compared as distributions through the promotion estimator of §16.8.
- The coverage figure. Every training and evaluation receipt carries
referee_coverage = { gpu_world_seconds, shadowed_world_seconds, share, one_step_defects, outcome_pairs, adjudicator }. A run with no background referee prints a share of zero.
How much it can cover. FACT, arithmetic on the targets of §36 (Appendix H): GPU world-seconds per second are T·h,
so a humanoid batch at 1.7 M world-steps/s (the tolerance of §36.3 row 7) and h = 2 ms produces 3,400 and a quadruped
batch at 2 M produces 4,000. A referee at 0.2 to 1.0 times real time per core on 24 cores produces 4.8 to 24. The
share is therefore about 0.14 to 0.71% for the humanoid and 0.12 to 0.6% for the quadruped, and 2.4% where a device
produces only 1,000 world-seconds per second: about 0.1 to 2.4% of GPU world-seconds at that assumed referee speed.
Referee speed dominates the figure and is unmeasured: a published CPU engine runs a humanoid at about 48 times real
time per core, and a referee within a factor of ten of it would cover several percent. HYPOTHESIS: the strict
profile, which is our own code in f64, is faster per core than the referee and lifts the share accordingly; the M2
measurement decides. FACT: a published CPU engine steps a humanoid at 28,000 to 46,000 steps/s per core (Appendix H).
A batch-cpu tier within a factor of four of that on 24 cores would produce 0.10 to 0.16 times the humanoid GPU
tolerance of 1.7 M at a factor of four and 0.40 to 0.65 times at parity, so under the lockstep profile the CPU could
shadow between a tenth and two thirds of the humanoid worlds; that comparison checks kernels and arithmetic, not the
law.
What it can detect. A gap of five points in success rate between tiers needs about 16·p·(1 − p)/δ² episodes per tier: 576 at p = 0.9 and 1,600 at p = 0.5, which is 11,520 to 32,000 world-seconds at 20 s per episode, or 8 minutes to 1.9 hours of background referee time at the rates above. One-step defects from a wrong kernel show in seconds.
The referee's own coverage. Separately from the share of GPU work that is shadowed, the referee cannot
complete every scenario (§16.4). At M2 exit, and in every release, the ledger reports per robot class: the share
of standard-battery scenarios the referee completes, broken down by refusal cause (contact cap, fixed-witness
gate, missing joint class, kinematic loop), the share the strict profile completes instead, and the wall-clock
of a promotion at the 50th and 95th percentiles. HYPOTHESIS: coverage is high for point-foot quadrupeds, partial
for humanoids with sampled soles and low for tracked robots; the test is the measurement. Consequence: in a class
where the referee completes fewer than 90% of battery scenarios (TARGET; the threshold of BET-34), established
always names its adjudicator, and a comparison between classes with different adjudicators says so in its first
line.
Pilot before funding: on Metal, a kernel rewriting a 100 MB buffer per command buffer with a blit of 256 rows to a snapshot buffer while 24 cores run the referee; on one RTX 4090, the same with explicit readback. Falsifier: GPU kernel time rises by more than 5% against idle CPU, or coverage is below 0.1% of world-seconds. Ceiling: no compute benefit by construction; the benefit is fewer false accepts. Because sampled worlds are drawn with recorded inclusion probabilities, they also serve as the referee half of the paired tier correction (§16.8), which takes confirmation off the wall-clock critical path without reducing its cost. ✖ KILL-SHADOW in §16.17.
16.12 Events, loops and limits of the claim
- Events. The lockstep tiers are time-stepping schemes with a velocity-scaled contact margin; they do not
locate events. The referee and the
strictprofile do (§16.5), and the referee's audit uses event location and certified collision detection (fs-time::rk45_until_event,fs-contact::certified_ccd, FACT) on sampled windows to count contacts the lockstep tiers missed. - Kinematic loops. On the lockstep tiers a closure is a compliant equality row in the convex solve, with its compliance stated F. Under a transactional profile closures are the bilateral rows A of §16.5. Exact closures on the referee wait for FrankenSim's constraint lane (FD-FSIM-04).
- Derivatives. A derivative through a step is offered only where it exists: fixed-mode derivatives on smooth kernels (D-8). Near a grazing contact, an ambiguous simultaneity or a changing active set the request is refused with the offending state, and a one-sided estimate, a generalized derivative or a smoothed surrogate each carry their own label. Design gradients by paired differences are §17.13.
- No real-time claim. Nothing here claims real-time or hardware-in-the-loop behaviour. Interactive viewing runs one world at presentation rate with physics unchanged.
16.13 Verbs
| Verb | Kind | First milestone | What it returns |
|---|---|---|---|
sim.run |
run | M0 (referee), M2 (all tiers) | Existing verb. Gains --profile (lockstep, strict or referee), --law, and per-world slots through --designs (§17). Fast-tier results are provisional. |
sim.estimate |
read | M2 | Existing verb. Now reports capacities, bytes per world, the expected Degraded rate and the expected referee coverage for the class. Never runs physics. |
eval.promote |
run | M3 | Existing verb. Applies the label rule and the tier correction of §16.8; --method is paired, dr, multilevel or segment. Prints ρ and the higher-tier-only count it replaced. |
fidelity.ledger |
read | M2 | Ledger rows for a class, quantity and tier pair, with envelopes and their freeze date. |
fidelity.rank |
run | M3 | The rank-agreement record of §16.9 for a design set and tier pair, with the permitted use per stratum. The measurement between controller rungs is eval.rank-agreement (§28.10), which writes the same record type. |
fidelity.modelform |
run | M3 | model-robust or model-conditional for a ranking, with the variant that flipped it. |
fidelity.energy |
read | M2 | Created-energy totals per world and contact mode for a run, with the worlds that crossed the bound. |
fidelity.coverage |
read | M2 | Both coverage figures of §16.11 for a run or a class. |
fidelity.shadow |
admin | M3 | Start, stop and configure the background referee for a job; reports its cost. |
The fidelity cluster has six verbs. The limit of seven in §32.12.4 applies to task clusters, the tools one agent
session lists at a time, and not to registry clusters.
16.14 Refusals
| Code | Name | Raised when | Ranked repairs |
|---|---|---|---|
FROB-REF-212 |
Cayley bound | ‖ω‖·h > 0.2 in a substep (existing, C-05) | Substep; use the exponential |
FROB-REF-220 |
Event refinement limit | An event cannot be bracketed within L_max, or more than E_max events fall in one physics interval | Report the accumulation; switch law; shorten h |
FROB-REF-221 |
Transaction exhausted | Every numerical plan of the retry ladder failed its checks | Teleport and inspect; the world is Faulted with numerical-fault |
FROB-REF-222 |
Profile not available on tier | strict or referee requested on a GPU tier |
Run on batch-cpu |
FROB-REF-223 |
No refinement law | A stochastic driver without a registered refinement law under a transactional profile | Register the law; move the draw to a sample instant |
FROB-REF-224 |
Indeterminate contact | No solution, or several distinct post-event motions | Report the configuration; do not return a number |
FROB-REF-721 |
Referee capacity | An island exceeds the referee's contact cap, or FrankenSim's fixed-witness gate refuses | Patch reduction with its fixture; --profile strict |
FROB-REF-722 |
Capacity overflow, refused | A world exceeds its contact or row capacity and the caller chose refuse | Raise capacity (new StepPlanId); reset; continue-flagged |
FROB-REF-820 |
Label misuse | A result that is not established is cited as established, or a model-conditional ranking is used to authorize spend |
Promote; run the settling experiment |
FROB-REF-821 |
Envelope not frozen | Promotion requested for a quantity whose envelope is not hashed in the ledger | Freeze the envelope first |
16.15 Invariants and proof obligations
| ID | Statement | How it is checked |
|---|---|---|
| INV-KIN-01 | No reader observes uncommitted state: every read names a commit_epoch. |
Fault injection between T5 and T6; a reader that sees a candidate fails the test |
| INV-KIN-02 | A retry restores the whole restore set. A transaction that commits on attempt k from checkpoint S yields the same digest as a run that started from S with plan k. | Forced retries on 10⁴ random intervals; digests compared |
| INV-KIN-03 | A random input is a function of its address and interval. It is never redrawn. | The redraw fixture; split, rejoin, reject and migrate on one noise path |
| INV-KIN-04 | A world's digests depend on its address and identities only: not on its slot, its co-tenants, the capacity of its buffers or the world count. Determinism class D2. | 1,000 episodes solo and in 100 random co-tenancies and capacities |
| INV-KIN-05 | Capacity overflow is flagged or refused. No contact is ever dropped silently. | Forced overflow at every capacity |
| INV-KIN-06 | Energy created by the simulator is recorded per step, and a result over the bound is labeled. | The two planted exploits of §16.10 |
| INV-KIN-07 | Physical parameters are unchanged by a retry, a step variant or a tolerance variant. | Parameter digests before and after |
| INV-KIN-08 | The label of a result is the weakest label among its inputs. | Type-level, by the claim product; a property test over random derivations |
| ID | Obligation | Class and route |
|---|---|---|
| PO-KIN-01 | The lockstep objective stays strictly convex for every admitted parameter row and lagged friction coefficient: A ≻ 0 and each ℓ_c convex (T-08). | Certified per law; tested per parameter row |
| PO-KIN-02 | The leading block of a padded Cholesky factor equals the unpadded factor. | Proved (elementary); tested bit for bit |
| PO-KIN-03 | A power-of-two-aligned pairwise sum with trailing −0.0 entries equals the sum over the real prefix, for every capacity (with +0 entries it fails when the prefix sums to −0.0). | Proved by induction on the tree; tested |
| PO-KIN-04 | The dyadic refinement law reproduces the law of Brownian increments: child variances ℓ/2, zero covariance. | Derivation on file; statistical test at 10⁶ draws |
| PO-KIN-05 | The expectation of Zᵢ given the past equals the expected higher-tier outcome for any frozen ĝ when pᵢ is known. | Derivation on file; null simulation |
16.16 Known answers
| Case | Truth | Negative control (the mutation that must fail) |
|---|---|---|
| Ballistic drop onto a plane under a transactional profile | Impact time √(2·z₀/g), outgoing velocity from the law's restitution, remaining flight by the same formula | Endpoint-only gap check |
| Tied sliders at a wall | Wall impulse 2 N·s; both at rest afterwards | Solve one contact, then impose the tie: 0.5 m/s into the wall |
| Noise refinement | Each half-interval increment has variance ℓ/2 for a parent of length ℓ; halves uncorrelated; sum equals the parent | Redraw on rejection: P(positive sign) = ¾ |
| Forced retry | Digest equals the clean run with the same plan | Omit packet queues or warm starts from the restore set |
| Padding | Digests of a world are equal at capacities 32, 64 and 128 and in any co-tenancy | A reduction tree shaped by capacity |
| Parameter variants | A batch of P variants equals P homogeneous runs bit for bit on one back end | Parameter row read after the first use in the step |
| Promotion table | 16 clean pairs promote and 22 with one exceedance, at p₀ = 0.05, p₁ = 0.30, α⁺ = 0.01 | Envelope chosen after the outcomes |
| Tier correction | With pᵢ = 1 it equals the paired mean; with a constant tier offset it recovers the offset | Audit probabilities logged after the draw |
| Energy audit | Position-level correction creates 0.125 J per event at 1 kg, 1 mm, 2 ms; CL-1 creates none | Folding the residual into contact dissipation |
| Free-body variational step | Order 2 by step halving; reversal to round-off; momentum map conserved without potential | An additive quaternion update |
The rows of GATE-KIN-01 and the ballistic invariants of C-04 are unchanged.
16.17 Gates and kill gates
| Gate | Claim | Threshold (TARGET) | Earliest |
|---|---|---|---|
| GATE-KIN-01 | Known answer | Pendulum period within 10⁻⁶ relative at h = 1 ms; symmetric-top precession within 10⁻⁴; free-flight momentum conserved to 10⁻¹² relative over 10⁶ steps on the variational lane; ballistic invariants of C-04. | M1 |
| GATE-KIN-02 | Parity | Contact-free forward dynamics of each reference robot against the pinned oracle: joint accelerations within 10⁻⁹ relative in f64. |
M1 |
| GATE-KIN-03 | Parity | batch-cpu in f64 against fs-mbd: kinematics, inverse dynamics, forward dynamics and mass matrix within 10⁻¹⁰ relative on 10⁴ states per reference robot. |
M2 |
| GATE-KIN-04 | Known answer | The transactional rows of §16.16 pass on the referee and on strict: impact time within 10⁻⁹ s; tied sliders exact to 10⁻¹² N·s; zero observed uncommitted reads in 10⁵ injected faults; forced-retry digests equal; refinement statistics inside 3σ at 10⁶ draws; the redraw mutation caught. |
M1 (referee); M2 (strict) |
| GATE-KIN-05 | Known answer (determinism) | INV-KIN-04: 1,000 episodes with identical digests solo and in 100 random co-tenancies, capacities and slot assignments, per back end; the path-table Jacobian gives the same digests as the dense one. | M2 |
| GATE-KIN-06 | Parity | A batch of 256 parameter variants of each reference robot equals 256 homogeneous runs bit for bit on batch-cpu and on each GPU back end. |
M2 |
| GATE-KIN-07 | Known answer | Under the capacities of §16.6, Degraded world-steps ≤ 10⁻⁴ of world-steps per class on the standard battery; every forced overflow flagged; capacities printed in every disclosure block. |
M2 |
| GATE-FID-01 | Fidelity | On the M2 battery, batch-cpu against referee: median and 95th-percentile disagreement at or below 2% and 5% for base displacement over 2 s, 3% and 8% for energy per metre, 5% and 12% for stance impulse. Thresholds are re-baselined once at M2 with a written rationale, then frozen. |
M2 |
| GATE-FID-02 | Fidelity | The promotion rule is exercised end to end: a planted tier-sensitive policy is caught (NS-09). | M3 |
| GATE-FID-03 | Fidelity | The referee's own coverage is reported per robot class at M2 exit with the breakdown by refusal cause, the strict share and the promotion wall-clock at the 50th and 95th percentiles. Pass requires the report for every reference class and the adjudicator named wherever coverage is below 90%; coverage below 90% in a reference class falsifies BET-34 for that class, and its fallback applies. |
M2 |
| GATE-FID-04 | Fidelity | Rank agreement between batch-cpu and the referee on the 40-design set per reference class, with the permitted use per stratum written to the ledger; per-episode ρ on 256 paired episodes per design reported. |
M3 |
| GATE-FID-05 | Known answer | Model-form check: the planted pyramid-cone and penetration-recovery winners are returned as model-conditional; a genuine winner is model-robust; cost at most 10% of the campaign's compute. |
M3 |
| GATE-FID-06 | Known answer | Energy audit: both planted exploits flagged within one episode; an honest policy under CL-1 unflagged; created energy on the standard battery below the frozen bound. | M2 |
| GATE-FID-07 | Fidelity | Background referee: coverage at least 0.1% of GPU world-seconds, reported; GPU throughput loss at most 5%; a planted single-kernel defect localized to its stage from one-step defects; a planted tier-sensitive policy detected within twice the time §16.11 predicts. | M2 |
| GATE-FID-08 | Known answer | Tier correction: type-I error at most α over 10⁴ null comparisons with optional stopping; equals the paired mean at pᵢ = 1; a planted between-tier reversal is not established; ρ and the replaced higher-tier count printed. | M3 |
| Kill | Trigger | Fallback |
|---|---|---|
| ✖ KILL-DR | ρ below 0.7 for the promoted quantities on 200 battery addresses, or effective sample size below 30% of the audited count in more than a tenth of promotions | The plain paired e-process promotion of C-15; segment mode where whole-episode ρ is low |
| ✖ KILL-RANK | Top-weighted τ below 0.5 in more than half the strata for the reference classes | Tier pairs promote individual results only; no population is ordered across tiers |
| ✖ KILL-SHADOW | GATE-FID-07 fails on cost or on coverage after one repair cycle, or no planted defect is detected | The referee runs at promotion time only, plus an offline audit of sampled snapshots after each run |
| ✖ KILL-VARIATIONAL | §16.7 | The free-body step as oracle; the default integrator everywhere |
| ✖ KILL-TREEOP | §16.6 | Dense factorization and a cap of 64 degrees of freedom |
16.18 Open questions
- OPEN-KIN-1. Default physics step and Newton iteration budget. Default: 2 ms, 4 iterations with warm start. Decided by the fidelity-throughput frontier measured at M2.
- OPEN-KIN-2. Whether sleeping and islands are worth their complexity for single-robot worlds on the lockstep tiers. Default: no. (Islands are always formed under a transactional profile.)
- OPEN-KIN-3. Final contact and row capacities per workload other than Q and H, whose capacities OPEN-SPARK-3 owns (§36.3). Default: the table of §16.6. Owner: Kinesis lead. Blocks GATE-KIN-07. Fail-closed: overflow is flagged and the caller chooses.
- OPEN-KIN-4. Which law the
strictprofile runs by default once CL-H is admitted. Default: CL-1 with located events, with CL-H as a named variant. Blocks nothing; decided with GATE-TRAC-09. - OPEN-KIN-5. L_max, E_max and the retry budget. Default: 16, 64 and 3. Decided by the refusal rate on the standard battery at M1; fail-closed default is refusal.
- OPEN-KIN-6. The created-energy bound per law and step. Default: 1% of actuator work per episode. Re-baselined once with GATE-FID-01.
16.19 Row-class solve queues
What it is. F. This subsection extends §16.6 and adds an arm to the head-to-head of §17.14. The lockstep
solve of §16.6 launches stages 7 and 8 over every world at the row capacity R for a fixed number of Newton
iterations, under masks. Padding by morphology is bounded by the buckets of §17.5. Padding by work is not:
inside one model the active row count changes within a stride, and worlds converge after different numbers of
iterations. A row-class solve queue regroups worlds by the work they need, so that one execution group (32 lanes
on CUDA) holds worlds with similar active row counts that are still iterating, whichever robot they belong to.
Owner: frob-kernels for the kernels, frob-dyn for the step plan. Exposure: the option --solve-queues of
sim.run and phylum.sweep; phylum.plan prints the bound below for the population in hand; the placement
receipt of §21.12 records whether queues ran.
Mechanism (DESIGN).
- Classes. After stage 5 each world has an active contact count c_w ≤ C. Three row classes are fixed fractions of capacity: c_w ≤ C/4, c_w ≤ C/2 and c_w ≤ C. Class k has a queue of world indices with a fixed capacity Q_k, a constant of the shape class, so no buffer changes shape.
- Assignment. Kernel
K-queue-assigncounts worlds per class, takes an exclusive scan and writes world indices in ascending world order: count, scan, write, with no floating-point atomics. A world that does not fit its class queue goes to the next larger class, and from the largest to the general masked lane, which holds every world and therefore always fits. - Execution. The kernels of stages 7 and 8 are launched once per class with that class's loop bound, and they read and write each world's rows in place through the index list. No state moves.
- Regrouping live worlds. After every second Newton iteration (DESIGN default) the assignment is repeated on
the
donemasks, so later iterations run over the live prefix of each queue. The number of launches stays a constant of the step plan; a launch over an empty prefix returns at once. - Generations. Queue capacities change only at a control-interval boundary, as a new generation of the step plan's execution options.
Why no digest can change. A world's arithmetic reads its own rows only (INV-SPARK-04). The loop bound of its
class is at least its active count, and inactive rows are identity rows that follow every active row (§35.5).
The per-world done test is unchanged. Reductions inside a world are anchored trees whose shape does not depend
on a bound. Queue membership, order inside a queue and co-tenants therefore enter no result: this is INV-KIN-04
applied to work classes. Until GATE-KIN-08 has passed on a backend, a run with queues carries the label
experimental-execution and is refused as promotion evidence (refusal: execution option not yet qualified on
this backend).
Counterexample: classing once per episode. A trotting quadruped with four samples per foot has 8 active contacts in diagonal support and 16 in double support (illustrative). With C = 24 the class bounds are 6, 12 and 24, so the world crosses a class boundary twice per stride. A sort made at reset leaves it in the wrong class for part of every stride: either the loop bound is below its active count, which drops contacts and is forbidden by §18.11, or it sits in the largest class and nothing is saved. Assignment is therefore per step.
Ceiling, stated before any speed test. Let f be the share of step time in the solve, ℓ the share of the lane-time launched there that does arithmetic on an active row of a world still iterating, and h the added relative overhead (assignment, indirection, extra launches). Under the serial model of §17.6,
speedup ≤ 1 / (1 − f + f·ℓ + h).
- FACT (Appendix H): the solve is 47% of an incumbent's humanoid step. With every idle lane removed and no overhead the bound is 1/(1 − 0.47) = 1.89. The mechanism cannot double the step.
- Idle lanes are the rule in a lockstep solve. If a fraction ℓ of worlds is live and liveness is independent across the 32 lanes of a group, the group is busy with probability 1 − (1 − ℓ)³². At ℓ = 1/4 that is 1 − 1.0 × 10⁻⁴: every group stays busy while three quarters of its lanes wait. After regrouping, 8,192 worlds at ℓ = 1/4 fill 64 groups instead of 256.
- HYPOTHESIS, from the launch figures of §17.5: the solve accounts for about 30 of a step plan's 60 launches at 5 to 10 µs each, so three classes in place of one add 60 launches, 0.3 to 0.6 ms per step. At 0.5 µs per world-step that is h = 0.07 to 0.15 at 8,192 worlds and 0.009 to 0.018 at 65,536. The test is the measured launch cost in the fingerprint (§21.16).
| Worlds | ℓ | h | Bound at f = 0.47 |
|---|---|---|---|
| 8,192 | 0.50 | 0.07 to 0.15 | 1.20 to 1.09 |
| 8,192 | 0.25 | 0.07 to 0.15 | 1.39 to 1.25 |
| 65,536 | 0.50 | 0.009 to 0.018 | 1.29 to 1.28 |
| 65,536 | 0.25 | 0.009 to 0.018 | 1.52 to 1.50 |
The mechanism is a modest optimization that pays at large batches and low live fractions, and it loses at 8,192 worlds if launches cost what the hypothesis says. On Metal every dispatch of a chunk is encoded serially into one command buffer (§21.9), and the cost of an added dispatch there is unmeasured; admission is per backend.
Relation to family lanes. The lanes of §17.6 group worlds by tree and contact mode in order to share constants. Queues group by amount of work across unrelated robots. Both remove masked work from the same stages, so their gains do not multiply. The profile measures each alone and both together, and the simpler configuration that reaches the larger measured gain is kept.
Pilot before funding. No new kernel. On the M2 profile of the masked path, counters record ℓ per Newton
iteration index and the active-contact histogram per reference class (telemetry.histogram), and the
fingerprint supplies the launch cost. Falsifier: a predicted bound below 1.15 at every batch size the campaigns
of §28 use.
Known answers and controls. (1) A world's digests with queues on equal its digests on the masked path, bit for bit. (2) The trot fixture: per-step assignment follows the 8 and 16 alternation, and the reset-time sort is caught, by a truncation flag or by zero saving. (3) A planted kernel that sizes a reduction tree by the class bound, in place of anchoring it, changes a digest and is caught.
Gate and kill gate.
| Gate | Claim | Threshold (TARGET) | Earliest |
|---|---|---|---|
| GATE-KIN-08 | Performance; known answer (determinism) | Digests identical to the masked path on 1,000 episodes per backend under 100 random co-tenancies and queue capacities; overflow to the general lane at most 1% of world-steps on the standard battery; a mixed cohort with queues reaches at least 0.70 of its weighted homogeneous throughput; the 95% lower bound of the paired whole-step ratio against the masked path is at least 1.15 at the batch size named in the receipt. | M2 on CUDA; Metal when its dispatch cost is measured |
✖ KILL-ROWQUEUE. Trigger: the pilot's falsifier holds; or the measured lower bound is below 1.15; or h ≥ f·(1 − ℓ) on the measured profile; or a digest mismatch survives one repair cycle. Fallback: the masked lockstep solve of §16.6 alone, with the ℓ counters kept in the disclosure block of §36.7.
In the head-to-head of §17.14 this is arm E, arm A with queues on, reported on measure (2). The pre-declared outcomes of that experiment are unchanged.
16.20 Decoupled stepping of separated robots in one world
What it is. M. Per-world time tiles between interactions are rejected as a general mechanism (§21.20):
separate environments are already independent, and policy and sensor frames bound a tile to 5 to 20 steps. With
a fraction p of step time inside tiles of L steps and overhead o the bound is 1/(1 − p + p/L + o); ten needs
p ≥ 0.924 at L = 64 and o = 0.01, and is out of reach for any p once L ≤ 10, where the bound at p = 1 is 9.1.
One narrow case remains, for the multi-robot scenes of §25.17: several robots in one world that are far apart
for most of a shift. While a certificate shows that two robots cannot touch, exchange a message or sense each
other before the next boundary, each is stepped as its own single-robot world in its own cohort and shape
class, and the multi-robot world exists as a ledger over them. Owner: frob-dyn for the tile planner, with the
swept bounds of §18.4 from frob-contact and the clock contract of §20. Exposure: sim.run on a multi-robot
scene with --decouple; the tile ledger is a table of telemetry.summary.
Mechanism (DESIGN).
-
Islands. The bodies of the world are partitioned into islands: a robot with everything it carries or touches; each free prop on its own; the static scene in none.
-
Tile boundary. All islands of a world are synchronized at every tile boundary. A tile ends at the earliest of: (a) the next control tick of any robot in the world; (b) the next exposure or sample instant of an exteroceptive sensor whose range and field of view could contain another island; (c) the earliest arrival of a message on a declared channel between islands, which is its send instant plus the channel's minimum latency; (d) the earliest possible contact between two islands. Clause (a) is the fence that removed the general mechanism, and nothing here exceeds it.
-
Separation bound. For islands a and b with bounding radii r_a, r_b about reference points at distance d, current speeds v_a, v_b of those points, and caps ā_a, ā_b on their accelerations (from actuator cards and friction limits; a free prop takes g), the gap after time t is at least
γ(t) = d − r_a − r_b − (v_a + v_b)·t − ½·(ā_a + ā_b)·t².
A tile of length T is certified for the pair when γ(T) > 0. Clipping speeds at the cards' speed caps can only enlarge γ and is allowed.
-
Decoupled stepping. Inside a certified tile each island steps in a slot of a single-robot cohort that shares the world's scene slot. Its random streams keep the multi-robot world's address, with the island as the perturbed entity (§35.9), so a robot's draws do not depend on whether it was stepped coupled or apart.
-
Merge. Islands whose pair is not certified for the next tile are stepped together for that tile as one coupled group, under the
strictprofile with an island-wise solve, or in a multi-robot shape class where one is admitted. The switch happens only at a tile boundary and is recorded with the pair, the gap and the clause that ended certification.
Smallest case, by hand. Two robots are at rest with a 0.10 m gap between their bounding spheres, and each can accelerate at 5 m/s². The gap closes when 2·(½·5·t²) = 0.10 m, at t = √0.02 s = 0.141 s. A 20 ms tile is certified, with γ = 0.10 − 0.002 = 0.098 m. A 0.2 s tile is not: γ = 0.10 − 0.20 < 0. Counterexample: a bound built from current velocities alone sees a closing speed of zero and certifies any tile length, and the robots then collide at 0.141 s inside a tile, unseen. The acceleration term is not optional. Second case: head-on at speed caps of 1.5 m/s each from a 3.0 m gap, the earliest contact is at 1.0 s, so fifty 20 ms tiles are certified and the pair merges when the gap falls below 3.0 × 0.02 = 0.06 m.
Ceiling, stated before any speed test. For N similar robots let φ be the share of a single-robot step spent in the dense factor. A coupled world factored densely costs (1 − φ) + φ·N² per robot-step, in units of a single-robot step, and the decoupled worlds cost 1 + o. At φ = 0.1, the upper end at 18 to 35 degrees of freedom (§16.6), and o = 0.02 the ratio is 2.45 for N = 4 and 7.16 for N = 8. That is the ratio against the weakest comparator. A comparator that factors island by island already removes the N² term, and against it the saving is the cross-island broad phase alone. FACT (Appendix H): collision is about 4% of an incumbent's humanoid step, so that ratio is below 1.05. The mechanism is therefore not a throughput lever. What it gives is admission: a fleet scene runs on the single-robot shape classes, kernels and launch graphs that already exist, each robot in its own size bucket, with no multi-robot bucket padded to N times the largest robot. Its value is bounded by the share of a shift the robots spend certified apart.
Pilot before funding. Paper only. On the warehouse archetype of §19.7, with 4, 8 and 16 kinematic route-graph actors standing in for robots, the share of robot-time certified at 20 ms tiles is computed from traces. Falsifier: a share below 0.5 at 8 robots.
Known answers and controls. The two hand cases; a planted message whose latency is shorter than the tile, which must end the tile at its arrival; a lidar that can see the other robot, which must end the tile at its sample instant; a thrown prop, whose island must merge before impact.
Gate and kill gate.
| Gate | Claim | Threshold (TARGET) | Earliest |
|---|---|---|---|
| GATE-KIN-09 | Known answer | On the two-carrier handoff fixture of §25.17 and a 16-robot sparse warehouse: zero contacts, message arrivals or sensor returns from another island inside a certified tile over 10⁶ robot-steps; every robot's digests equal those of the reference that steps the same islands in one world with island-wise solves on batch-cpu; the four controls behave as stated; the certified share is reported. |
M8, with §25.17 |
✖ KILL-INTERACT. Trigger: one missed interaction survives a repair cycle; or the certified share is below 0.5
of robot-time on the warehouse fixture. Fallback: a multi-robot world is one coupled world under strict with
island-wise solves and fixed decimation, and a fleet scene beyond its capacity is refused under §18.11.
16.21 Step adjoints through the converged contact solve
This subsection specifies the mechanism behind the M6 item "gradients through contact" and the revisit clause of
NONGOAL-15. Tier M. It is off the critical path and is funded only after its pilot. Owner: frob-dyn for the
reverse sweep, with derivative kernels produced by derive/frob-derive into frob-kernels. Verb:
phylum.gradient --method adjoint; the paired finite-difference method of §17.13 stays the reference.
What it is. On lockstep tiers stage 8 returns the velocity v⁺ that minimizes a strictly convex objective Φ(v; z), where z collects the state at the start of the step, the action and the design parameters of the parameter row (§16.6). At convergence the residual r(v, z) = ∇ᵥΦ is zero and the Hessian H = A + JᵀGJ has already been factored. The implicit function theorem gives dv⁺/dz = −H⁻¹ · ∂r/∂z. In reverse mode, for the cotangent v̄ of a scalar loss with respect to v⁺:
solve H · λ = v̄ (two triangular solves with the factor the step already holds), then z̄ ← z̄ − (∂r/∂z)ᵀ · λ.
Nothing is differentiated through Newton iterations or the line search. The two solves cost about n_v² multiply-adds: 324 for the reference quadruped (n_v = 18) and 1,225 for the humanoid (n_v = 35). The products (∂r/∂z)ᵀ·λ and the transposed Jacobians of stages 1, 6, 7 and 9 are straight-line reverse programs derived at build time from the kernel IR that produces the forward kernels, each with an instance certificate that it equals the symbolic derivative (§12, §21).
The sweep. A horizon of K control intervals with K ≤ 32 (DESIGN; §17.13 caps paired differences for the same
reason). Step buffers are stored for every world-step of the horizon, so no recomputation schedule is needed. FACT
(arithmetic on §16.6): at 27 kB per quadruped world, K = 16 and 10 world-steps per interval, the store is
160 · 27 kB = 4.32 MB per world and 8.85 GB for 2,048 worlds. sim.estimate prints the figure and the job is refused
when it does not fit. HYPOTHESIS: the reverse sweep costs about one forward sweep; the pilot measures it.
Labels. Each step's derivative carries a label of C-16. smooth: every contact and joint row is strictly inside
one region of its law (open, stick, slip; limit inactive or active), with normalized margin above a guard.
fixed-mode: the derivative is taken with active sets frozen. one-sided: a row sits on a boundary; both one-sided
values are returned on request. A step whose solve did not converge has no derivative, because the premise r = 0
fails, and its world is dropped from the batch gradient and counted. The label of a K-step gradient is the weakest
label on its path, reported with the number of steps under each label.
Guards.
- Agreement. In every batch, 8 random unit directions in design space are also evaluated by paired central
differences (§17.13) on 64 seeds, and the cosine between the two sets of directional derivatives is recorded. Below
0.65 the horizon is halved and the job repeats. Below 0.5 at K = 8 the verb returns the finite-difference result
with the label
adjoint-rejected. - Metrics. Smooth metrics only; a binary outcome is refused as in §17.13.
- Determinism. Class D2 or better, since the agreement check pairs worlds (§35).
- Manifold. A gradient with respect to an inertia is a tangent vector of the cone in the chart of §14, and a step along it uses the retraction. Under a symmetry of the closed loop the gradient is averaged over the observation-preserving subgroup (§13), so a symmetric design stays symmetric.
Known answers.
- Scalar implicit function. Φ(v; z) = ½·m·(v − z)² + ½·g_c·v², a free term plus a contact term of curvature g_c.
Then r = m·(v − z) + g_c·v, H = m + g_c and ∂r/∂z = −m, so v⁺ = m·z/(m + g_c) and dv⁺/dz = m/(m + g_c). With
m = 2 and g_c = 6 the derivative is 1/4; with the contact open it is 1. At the boundary the one-sided values are 1
and 1/4 and the label must be
one-sided. - A zero gradient is not a stationary point. A 1 kg block rests on a floor with μ = 0.5 under ideal Coulomb
friction and is pushed with force u for one step of 0.01 s. Breakaway is at μ·m·g = 4.905 N. At u = 4 N, v⁺ = 0 and
dv⁺/du = 0. At u = 6 N, v⁺ = 0.01 · (6 − 4.905) = 0.01095 m/s and dv⁺/du = 0.01 s/kg. For the objective "reach
0.1 m/s" the adjoint at u = 4 N is exactly zero although raising u past 4.905 N makes progress. Under a regularized
law whose friction ramps linearly up to a stiction speed of 1 mm/s, the stick derivative is 0.01/(1 + 49.05) ≈
2.0 × 10⁻⁴ s/kg, 50 times smaller than in slip, and its value is set by the stiction speed, which is a numerical
parameter (§18.5). A
fixed-modegradient in stick is therefore reported as uninformative, and a consumer that receives one takes a derivative-free step.
Consumers. opt.start with gradient steps on the design manifold, the refinement lane of §25.21.2 and the
sensitivity analysis A-16. A first-order policy learner over short horizons is not part of this item; it would be a
separate experiment under §23 with its own pilot.
Pilot before funding. Paper-scale and outside the workspace, on an incumbent engine that already ships derivatives: one quadruped, 30 design parameters, a smooth tracking-and-energy objective over 16 control intervals, 256 worlds, 20 replicate batches. Arms at equal wall-clock: adjoint ascent, the stencil of §17.13, and one derivative-free generation. Falsifier: median cosine against central differences below 0.5, or adjoint ascent needs more than a quarter of the derivative-free arm's wall-clock to reach the same objective value. Cost: about three engineer-days and 6 GPU-hours (TARGET).
Ceiling, stated before any speed test. One gradient costs about two sweeps against 2P + 1 = 61 for the stencil at
P = 30: at most about 30 times fewer world-steps per gradient, on smooth objectives over horizons where most steps are
smooth, and nothing where the objective is decided by a binary outcome or by a change of contact sequence. FACT
(Appendix H): several incumbents already differentiate their step, so this is not a differentiator.
Gate. GATE-KIN-10 (Known answer and Design comparison, M6). The scalar and block fixtures are exact,
with their labels. On a 100-case suite of short horizons, 80 with no change of active set and 20 controls with a
planted change: cosine against central differences at least 0.95 in at least 72 of the 80, and all 20 controls
labeled one-sided or fixed-mode, or refused. The reverse sweep takes at most twice the forward sweep's wall time
on the reference quadruped at 2,048 worlds.
Kill gate. ✖ KILL-STEPADJOINT. Trigger: the pilot's falsifier fires; or GATE-KIN-10 fails after one repair cycle; or the median cosine on the design corpus is below 0.5 at K = 8. Fallback: the paired stencil of §17.13 where its own gate holds, otherwise rank-based search; the response batches of §18.5 keep supplying adjoint columns at a single state.
16.22 Closed-form elimination of single-loop planar closures
Eliminated closures F. This extends the kinematic-loops rule of §16.12 and the linkage transmissions of §15.9. On the lockstep tiers a closure is a compliant equality row, and exact closures on the referee wait on an upstream constraint lane. For the loops that robots most often contain (a parallel knee, a steering linkage, a parallelogram that keeps a foot level) neither is needed: a planar four-bar has one degree of freedom and a closed-form position solution, so the loop can be removed from the constraint system when the model is compiled.
Recognition. At compile time frob-model marks a closure as eliminable when it closes exactly one independent
loop of four revolute joints with parallel axes, the mobility analysis A-01 reports one degree of freedom for the
loop, and no other closure shares a link with it. Everything else stays a row.
The closed form. With ground link d, input crank a at angle θ, coupler b and output link c at angle φ, both angles measured from the ground link, the loop equation |O₂ + c·(cos φ, sin φ) − a·(cos θ, sin θ)| = b has the two solutions
φ = atan2(k_y, k_x) ± arccos( (b² − c² − R²) / (2·c·R) ), (k_x, k_y) = (d − a·cos θ, −a·sin θ), R = ‖k‖.
The two signs are the two assembly branches. The branch is fixed by the design's reference pose and recorded in the
model; a compile that cannot tell the branches apart at the reference pose refuses. frob-kir emits kernels for
φ(θ), the transmission ratio n(θ) = dφ/dθ and n′(θ), and the same for the coupler angle. The dependent joints of the
spanning tree then follow θ: q̇_dep = n(θ)·θ̇ and q̈_dep = n(θ)·θ̈ + n′(θ)·θ̇², and the step uses the reduced
coordinates through the embedding Γ = ∂q/∂θ, with mass matrix ΓᵀMΓ and bias Γᵀ(M·Γ̇·θ̇ + h). No closure row, no
compliance and no drift remain for that loop on any tier, and the transformer modulus of §15.9 is n(θ) exactly.
Refusal near the toggle. Where coupler and output link line up, the argument of the arccosine reaches ±1, the branches meet and n(θ) is unbounded. Elimination is refused when the sine of the transmission angle falls below 0.17 (about 10°, DESIGN) anywhere in the declared joint range, and the loop stays a row with that reason in the report.
Known answers.
- Parallelogram, a = c = 1 and b = d = 2: on the open branch φ = θ and n = 1 for every θ.
- The same linkage at θ = 90° on the crossed branch: the coupler end is at (1.2, −0.6) and φ = −143.13°. A solver that returns this root for a knee assembled open has folded the leg the other way.
- A crank-rocker with (a, b, c, d) = (1, 3, 2.5, 3) at θ = 90°: φ = 98.857° and n = 0.3722 on the open branch; φ = −135.727° and n = −0.1722 on the crossed one.
- The parallelogram at θ = 0 is a change point: all four links are collinear and elimination is refused.
Owners and verb. frob-model owns recognition and the branch record; frob-kir the emitted kernels; frob-dyn the
reduced embedding. design.check reports each closure as eliminated or row with the reason.
Ceiling and pilot. Ceiling: one constraint block fewer per such loop, and no closure error; no throughput claim is
made, and a spatial or multi-loop mechanism gains nothing. Pilot before funding: the parallel-knee archetype on
batch-cpu both ways, comparing knee torque and foot trajectory over a 2 s stride against the referee's bilateral
rows.
| Gate | Claim | Threshold (TARGET) | First milestone |
|---|---|---|---|
| GATE-KIN-11 | Known answer; Parity | The four known answers to 10⁻¹⁰ rad in double precision; reduced-coordinate forward dynamics within 10⁻⁹ relative of the bilateral-row solution on referee over 2 s; on the lockstep tiers in single precision the loop residual stays below 10⁻⁶ m for 10⁶ steps; the crossed-branch mutant and a loop driven through its toggle are both refused |
M2 |
| Kill gate | Trigger | Fallback |
|---|---|---|
| ✖ KILL-LOOPELIM | The parity part of GATE-KIN-11 fails on any archetype loop after one repair cycle; or the recognizer matches fewer than half of the closures in the archetype library | Compliant equality rows on the lockstep tiers and bilateral rows under the transactional profiles, as in §16.12 |
16.23 Ancestor-sparse assembly and the solve-stage ledger
Ancestor-sparse assembly and the solve-stage ledger F. This extends the tree-operator route of §16.6. The pilot named there counts factor operations only, and the ceiling stated there ("the factor is under a tenth of the step at 18 to 35 degrees of freedom") is a HYPOTHESIS with no measurement behind it. On an incumbent's step trace the two stages this route touches, constraint assembly and the solve, take 38.31 and 123.55 of 262.33 ns per world-step, which is 61.7% of the step (FACT, Appendix H). If both vanished the step would be 2.61 times faster; a three-fold reduction of both gives 1/(0.383 + 0.617/3) = 1.70. Those are ceilings on that incumbent's step and say nothing yet about ours. They are large enough that the question is settled by a ledger and a pilot, not by an assumption, and small enough that the answer cannot be a source of the ten-fold (§2.5).
What is exploited. For a contact or limit row r attached to one link, the Jacobian row is nonzero only on that link's ancestor coordinates. Let n be the number of coordinates, d_k the depth of coordinate k (the number of its ancestors, itself included) and d_r the depth of the link that row r is attached to. Three operations inside every Newton iteration then shrink.
| Operation, per iteration | Dense count (multiply-adds) | Ancestor-sparse count |
|---|---|---|
| One row's contribution to JᵀGJ | n·(n + 1)/2 | d_r·(d_r + 1)/2 |
| Factor of H = A + JᵀGJ, ordered leaves first, no fill | (n³ − n)/6 | Σ_k d_k·(d_k − 1)/2 |
| Two triangular solves | n² | Σ_k (2·d_k − 1) |
| One row's residual and line-search evaluation | n | d_r |
Known answers, checkable by hand.
- Quadruped, n = 18: a six-coordinate base and four three-coordinate legs, so the depths are 1 to 6 once and 7, 8, 9 four times. Factor: 35 + 4·(21 + 28 + 36) = 375, against (18³ − 18)/6 = 969: a ratio of 2.58. A foot row has d_r = 9: 45 against 171, a ratio of 3.8.
- Humanoid, n = 35: a six-coordinate base, two six-coordinate legs (depths 7 to 12), a three-coordinate waist (7 to 9) and two seven-coordinate arms beyond the waist (10 to 16). Factor: 35 + 2·251 + 85 + 2·560 = 1,742, against (35³ − 35)/6 = 7,140: a ratio of 4.10. A foot row has d_r = 12: 78 against 630, a ratio of 8.1.
- A single chain (d_k = k) gives a ratio of exactly 1 in every row of the table. There is nothing to exploit, and the route reports that instead of running.
The counterexample that bounds the route. A row between two limbs (a hand on a thigh, one foot on the other) is nonzero on both ancestor chains, and its product couples the two limbs: the leaves-first order then has fill. The smallest case is a one-coordinate base with two one-coordinate limbs (n = 3). Rows on single limbs give H the arrow pattern with a zero at the (limb 1, limb 2) entry; one row touching both limbs fills that entry, and a factor that skips it is wrong. Such a world takes the dense path for that step under a flag, as §16.6 already requires, and the share of world-steps that do so is a reported number.
Fewer operations are not a faster step. On a device the step is bound by memory traffic and launch structure as much as by arithmetic (§36). The ancestor lists are data in the model table (§17.4) and are fixed per shape class, so every lane of a bucket still runs one static launch graph; the price is gathers through an index and shorter, irregular inner loops. Nothing here needs a compile per design. The sparsity is carried as data, and whether unrolled per-design code adds anything to it is the separate question of §21.18.
Pilot before funding: the ledger. xtask instruments the batch-cpu solve and prints, per Newton iteration and
per reference robot, the operation counts of the four rows above and their measured share of the step, for the
dense path and the ancestor-sparse path on identical worlds, at contact counts drawn from the standard battery.
Two engineer-days, no device (TARGET). Falsifier: the four operations together are under one fifth of the measured
batch-cpu step for both reference robots. In that case the route keeps its original scope: mechanisms above 60
coordinates and padded buckets.
Owner and verb. frob-dyn owns the solve and frob-model the ancestor lists. sim.run --solve-path {dense, ancestor} selects the path, and the StepPlanId includes it. The two paths order their arithmetic differently, so
they agree inside determinism class D2 and not bit for bit; a study never mixes them (§35), and a request that
would is refused with the two step-plan identities.
| Gate | Claim | Threshold (TARGET) | First milestone |
|---|---|---|---|
| GATE-KIN-12 | Known answer; Performance | The ledger reproduces 375, 969, 1,742 and 7,140 and the row counts exactly, and reports a ratio of 1 for a chain. The three-coordinate fill case takes the dense path and is flagged. The two paths pass the decision-level equivalence test of §35 on the standard battery. On the M2 device profile the whole-step gain is at least 1.15 for the 35-coordinate humanoid at 8,192 worlds | M2 |
| Kill gate | Trigger | Fallback |
|---|---|---|
| ✖ KILL-ANCESTOR | The pilot's falsifier holds; or the device gain is below 1.15 on the humanoid and below 1.10 on the quadruped; or more than 20% of battery world-steps take the dense path | The dense path for 18 to about 60 coordinates as in §16.6; the tree route only where KILL-TREEOP leaves it |
16.24 A held command and the step that integrates it
What it is. S A known answer, and three rules about the step that integrates a command held between control
ticks. A loop that is stable for the exact held-input plant can be unstable under the default integrator at one step
per tick. A loop that is asymptotically stable under a finer step can still miss a deadline. The controller, its rate,
the observation and the actuation are the same in every case; only the numerical step differs. Owner: frob-dyn
(integrators, §16.7), frob-analyze (analysis A-22), frob-conformance (the fixture). Verbs: sim.run,
eval.compare, and the verb of A-22.
The fixture. An ideal slider with commanded acceleration and a control period h = 10 ms. In the variables x (position error, mm), w = h·v (mm) and u = h²·a (mm), the controller reads x and w at each tick and holds u = −(3/2)·(x + w) until the next. The test takes the 64 initial pairs (x, w) in {−3, …, 4}² mm and passes a case when max(abs(x), abs(w)) ≤ 1/32 mm at tick 16.
| Integration between ticks | One-tick map | Spectral radius | Cases passed |
|---|---|---|---|
| Exact for the held input | x′ = x + w + u/2, w′ = w + u | 1/2 | 64 of 64 |
| One semi-implicit step: velocity, then position | x′ = x + w + u, w′ = w + u | (1 + √3)/2, about 1.366 | 1 of 64, the origin |
| Two such steps of h/2 | x′ = x + w + 3u/4, w′ = w + u | (5 + √57)/16, about 0.784 | 15 of 64 |
| d such steps of h/d, for d = 3, 4, 8, 16, 32 | x′ = x + w + (d + 1)·u/(2d), w′ = w + u | 1/2, 1/4, 0.395, 0.451, 0.476 | 64 of 64 |
The second row is the default integrator of the lockstep tiers (§16.7) at a decimation of one. The third row is asymptotically stable and fails 49 of 64 cases, because its contraction of 0.784 per tick is too slow for the deadline.
Rules. DESIGN.
- Analysis A-22 evaluates each sampled loop twice: with the hold modelled exactly, as it does now, and with the tier's own integrator at the task's decimation. It reports both spectral radii and does not label a loop stable on a tier where the second exceeds one.
- A step-refinement study holds every clock of the clock contract fixed in seconds (control period, sensor periods, latencies) and changes only h. Clocks are physical parameters in the sense of ADR-KIN-05. A study that halves the control period together with the step compares two controllers and is refused as a refinement (refusal: refinement study that changes a clock).
- A requirement with a deadline is compared across decimations d and 2d by its verdicts, case by case, and not by a
stability label. A verdict that differs is
fidelity-dependent(§24.10) and is not promoted from the coarser run.
FACT (a recorded pilot, Appendix K; to be re-run before a gate cites it): a rational referee and a native single-precision kernel on the Metal device and on each RTX 4090 agreed on all 192 decisions of the first three rows, and the device read-backs matched independently reconstructed single-precision terminal states byte for byte. The counts and radii in the table were recomputed in exact rational arithmetic for this section.
What it does not show. An ideal slider has no actuator limit, no contact and no sensor path. A robot's loop needs its own check. The fixture tests the sampled plant and its numerical lowering.
| GATE-KIN-13 | Known answer. The four rows reproduce exactly in rational arithmetic on the host. In single precision on batch-cpu, CUDA and Metal the pass or fail decision of each of the 64 cases equals the rational one for the exact map and for d = 1, 2 and 4. A-22 reports a tier radius of 1.366 at d = 1 and withholds the stability label. The mutant study that halves the control period with the step is refused. M2. |