FrankenRobotics

Numerics and determinism

5,634 words16 subsections9 tablesPlan version 1.1

On this page16
35.1 Determinism classes35.2 Profiles35.3 What each class licenses35.4 Decision-level equivalence under class D335.5 A world does not depend on the batch it ran in35.6 The bans35.7 Hazards and rules35.8 Anchored positions35.9 Randomness by address35.10 Diagonal scaling for single-precision contact solves35.11 Refusals35.12 A number never claims more than its certificate35.13 Invariants and proof obligations35.14 Known answers and negative controls35.15 Gates35.16 Open questions

This chapter fixes what a number in FrankenRobotics is allowed to claim: which results are exact, which are bitwise repeatable and where, which agree only inside a tolerance, and which claim nothing. Every other chapter that compares, replays, pairs or cites a number does so under a class defined here. The classes are tested by the Gauntlet's G5 tier (§38) and by the gates of §35.15. Everything in this chapter is S, with two exceptions: class D1 on devices is a HYPOTHESIS under KILL-D1, and the asynchronous variant behind the fast-training profile is F under KILL-SHARD.

ADR Decision Reason
ADR-NUM-01 Every verb declares the minimum determinism class of its inputs, and refuses below it A statistic that pairs, replays or attributes is only as good as the replay under it (§35.3)
ADR-NUM-02 Class D2 includes independence from batch size, slot, co-tenants, bucket capacity and shard count Incumbents document replay of a batch, not of a world taken out of its batch; the studio forks and teleports single worlds
ADR-NUM-03 Under class D3, agreement of task decisions is tested beside agreement of states A state tolerance neither implies nor is implied by agreement on a verdict (§35.4)
ADR-NUM-04 The contact solve runs in power-of-two-scaled variables and reports a conditioning figure Tolerances and regularization must mean the same for every row and every robot in a batch (§35.10)
ADR-NUM-05 A nondeterministic training profile exists only as labeled class D4, relaxing the learner and shard order and never the physics The asynchronous two-device variant must not pass as deterministic (§35.2)
ADR-NUM-06 Anchored offsets are centred on their cell A shift by the cell side is then always exact (§35.8)

35.1 Determinism classes

Every result records its class. A class is claimed by a contract and tested by the Gauntlet's G5 tier.

Class Guarantee Where it applies Tested by
D0 Exact Integer, rational or symbolic results Static analyses over ℚ, group computations, certificates, world addressing; exact-integer kernels and the decisions of guarded kernels on every backend (§21.8); integer and uniform random draws (a normal variate built from det_* floating-point functions is D2 per backend and D1 across backends only where GATE-SPARK-01 passes); rational clocks and phase accumulators (§20) GATE-SPARK-08
D1 Bitwise across backends The same bits on batch-cpu (f32), CUDA and Metal Algebraic kernels, if GATE-SPARK-01 passes for the device family. A HYPOTHESIS until then (KILL-D1) GATE-SPARK-01, -02
D2 Bitwise per backend The same bits run to run and across worker counts on one device family, driver and compiler; and across batch size, slot, co-tenant worlds, bucket capacity and shard count (§35.5) The design requirement for every kernel GATE-SPARK-15
D3 Tolerance Agreement inside a recorded envelope, and agreement of task decisions (§35.4) Cross-backend comparison when D1 is unavailable; cross-tier fidelity GATE-SPARK-02, -16
D4 Unspecified None Never cited as evidence. Its one sanctioned use is the labeled training profile of §35.2 GATE-SPARK-17

The referee tier in f64 aims at FrankenSim's cross-ISA class on CPUs by using its deterministic elementary functions; FrankenSim's fs-mbd currently calls platform sqrt, sin and cos and makes no cross-ISA claim (FACT), so that is upstream work item FD-FSIM-07.

Two rules govern the classes. The weakest input wins (INV-NUM-06): the class of a result is the weakest class among the kernels, placements, profile and inputs that produced it. A class is never raised by relabeling, only by running again under a stricter profile.

What is parity and what is not. Newton 1.8 and Isaac Lab 3.0 document opt-in bitwise run-to-run determinism for the same batch on the same GPU, and Newton states that CPU and GPU results usually differ (FACT, Appendix H). Class D2 as the same batch repeated is therefore parity. The two properties this plan adds, each claimed only after its gate, are the independence of a world from its batch (D2 as defined above) and bit equality between a device and the CPU reference (D1). Under class D3 a replay on another backend is a tolerance replay, which an incumbent can also layer.

35.2 Profiles

A profile is a named set of pins. A run declares one; a result carries it.

Profile Class What is pinned What it is for
reference-replay D2 on batch-cpu and on the referee; D1 against a device where attained Solver, arithmetic, toolchain, order and addresses. Declared state and logical events are byte-identical on repeat and across checkpoint and restart The record; teleport targets; attribution; the numeric side of certificates
device-replay D2 on one device family Stable reductions, sorted contacts, select-only masks, the fingerprint. Identical on repeat and under any permutation or regrouping of worlds Training, evaluation and paired comparison on that family
cross-backend D1 if attained, otherwise D3 with the decision-level test Physically scaled state error, impulse and work residuals, event times and task decisions; the first divergence is kept Promotion; teleport, labeled exact or tolerance; results that cross machines
fast-training D4, labeled Only the knobs listed below, all recorded Producing weights, and nothing else

Scope of fast-training (DESIGN). The physics step stays class D2 in every profile. The profile relaxes only: the order in which shards deliver collected segments and gradient sums (asynchronous collection, with a declared bound on version lag); the learner's reduction order and matrix-product implementation (K-gemm may use a device matrix path); and minibatch composition by arrival. Its only output is a set of weights tagged trained_under: D4. Three rules bind it:

  1. No number observed during such a run (a learning curve, a training return) is evidence (FROB-REF-861).
  2. Every claim about the resulting policy comes from re-evaluation under device-replay or stricter, which is deterministic given the weights.
  3. The profile is off by default and is enabled per job, never per installation.

Ceiling, stated before any test: inference and learning together are about 13% of training wall-clock on one incumbent stack (§36.4), so relaxing them gains at most 1.15 on one device. The profile's real use is the asynchronous two-device variant of §36.6.

35.3 What each class licenses

Use Minimum class Below it
Certificates, digests, world addresses, strict predicates, requirement verdicts at a boundary D0 Refused at build time (FROB-REF-252)
Replay of a world on another tier with a claim of bit equality (sim.replay --tier) D1 across families; D2 within one The replay is labeled a tolerance replay and compared by §35.4
Paired comparison on common world addresses (eval.compare) Both arms D2 on one device family with GATE-SPARK-15 current, or D1 across families: paired-exact D3: allowed as paired-by-address, labeled, with the realized pair correlation reported beside the verdict. Any arm D4 or of unrecorded class: refused (FROB-REF-561). Arms on different families without D1: unpaired by declaration
Interventional attribution and counterfactual forks (§27) D2 with GATE-SPARK-15 current Refused (FROB-REF-562)
Promotion of a fast-tier result (§16) D3 with a ledger envelope and the decision-level test D4: refused
Reuse of a memoized episode (§28) D0 for analyses; D2 on the same backend and fingerprint for episodes Recomputed
Training D2 by default D4 only as the labeled profile
A performance figure Any class, stated in the disclosure block (§36.7) —

Why pairing and attribution need replay. Seeding a generator fixes the draws, not the order of a reduction. A difference of one unit in the last place grows like e^(λt) in contact-rich dynamics and reaches order one after about ln(2²⁴)/λ ≈ 16.6/λ seconds, for a largest Lyapunov exponent λ that is measured per task. The variance of a paired difference scales with 1 − ρ: if uncontrolled arithmetic lowers the correlation between arms from 0.95 to 0.50, a comparison needs (1 − 0.50)/(1 − 0.95) = 10 times the episodes (illustrative values). A paired test under class D3 remains valid as statistics, which is why it is allowed and labeled. Under class D4 nothing about an arm can be replayed, so its verdict could not be reproduced, promoted or attributed; that is the reason for the refusal, and the refusal is by code, not by convention.

35.4 Decision-level equivalence under class D3

A state tolerance and a decision are different things. The smallest case: a success predicate x ≥ 1 m, one placement ending at 1.0000001 m and the other at 0.9999999 m. The state error is 2·10⁻⁷ m, inside any envelope; the verdicts differ. Conversely two placements may end a metre apart and agree that the robot fell.

DESIGN. For a pair of placements claimed at class D3, on a frozen ensemble of world addresses with frozen policies, spark check and eval.promote compute per world:

  1. state error in physically scaled units at checkpoints, until the first divergence beyond the envelope;
  2. impulse and work residuals;
  3. event times (first contact, termination);
  4. task decisions: termination cause, each success predicate, each requirement verdict, the winner of each design comparison in the ensemble, and the pass or fail of each gate metric;
  5. the margin of each decision, that is, the distance of its deciding quantity from its boundary.

The pair is decision-equivalent when all of the following hold (thresholds are TARGETs, OPEN-NUM-2):

  • the one-sided 95% upper bound (exact binomial) on the rate of discordant decisions is at most 1%. With 4,096 worlds that admits roughly thirty discordant worlds; with zero discordant worlds the bound is about 3/n, so at least 300 worlds are needed to show 1% at all;
  • discordances are not one-sided: the bound on the net rate (one placement passing where the other fails, minus the reverse) is at most 0.5%;
  • continuous outcomes (return, energy per metre, time to failure) are equivalent by two one-sided tests inside the margins of the fidelity ledger;
  • the ranking of the designs in the ensemble agrees at the top, by the rank-agreement measure the plan uses between evaluation rungs (§28).

Discordant worlds are listed with their margins. A decision whose margin is smaller than the divergence envelope is reported as boundary-sensitive on both placements, whichever way it fell. The first divergence of each discordant world (step, kernel, quantity) goes to the divergence ledger.

35.5 A world does not depend on the batch it ran in

Statement (INV-NUM-03). Given a policy (weights and normalizers), a world's trajectory is a function of its address, design, scene and task alone. It is the same bits when the world runs alone, in any batch, at any slot, beside any co-tenants, in any bucket whose capacities it fits, and on any number of shards of one device family.

FACT: Newton, which documents bitwise replay of a batch, holds a world equal to its isolated run only to 10⁻⁵ to 10⁻⁴ in its own tests (Appendix H). The property is therefore worth having only if it is exact.

Design conditions, each a rule elsewhere in this plan:

  1. No kernel reads or writes another world's rows (INV-SPARK-04).
  2. A per-world reduction is a pairwise tree anchored at index 0 over power-of-two blocks, so which two active elements are added first does not depend on capacity.
  3. Padding neutrality (INV-NUM-02). Padding trails the active lanes and holds the neutral element of its consumer: −0.0 for sums, 1.0 for products, +∞ and −∞ for minimum and maximum, and identity rows with zero right-hand side in the factorization, which follow every active row and so never enter an active entry.
  4. A mask is a select.
  5. The number of solver iterations launched is a constant of the shape class; a world whose done mask is set is copied forward, not recomputed.
  6. Randomness, resets and task assignment are functions of the world address (§35.9).
  7. Sort keys are total orders within a world.
  8. Statistics across worlds never feed state. Observation normalizers are frozen within a collection segment and are constants of the policy in evaluation.

Two counterexamples show that the rules are not pedantry. Padding with +0.0 is not neutral: the sum of the single active element −0.0 is −0.0 alone and +0.0 when one +0.0 lane is appended, because (−0.0) + (+0.0) = +0.0; the values compare equal, the bits differ, and a digest over bits changes with capacity. A batch normalizer is a neighbour: two worlds observing 1 and 3 and normalized by their batch mean see −1 and +1; the first world alone sees 0, takes a different action and follows a different trajectory.

The test is GATE-SPARK-15. Slot dispatch (§21.11), teleport, forks and the memo graph all rest on it.

35.6 The bans

In batched kernels, on every backend and in every profile except where §35.2 says otherwise:

  • no fast-math relaxation of any kind;
  • no fused multiply-add, contracted or explicit;
  • no built-in transcendental function and no reciprocal-square-root built-in;
  • no floating-point atomic;
  • no strict predicate in bare floating point.

A lint enforces the list in dialect sources and in the Rust of frob-kernels (mul_add, sin, cos, exp, ln, powf and their kin are rejected; sqrt is correctly rounded and allowed), with FROB-REF-251 at build time. The IEEE probe records how each device family rounds an explicit fused operation, as data for OPEN-NUM-1; the ban does not wait on it.

35.7 Hazards and rules

Hazard Fact Rule
Platform elementary functions They differ across libraries and GPUs Only det_* functions written in the dialect or from fs-math; a lint bans the built-ins in kernel sources
Fused multiply-add Compilers contract a·b + c by default on some backends Contraction off on every backend; no explicit fused operation in batched kernels
Reductions Order changes the sum Fixed-shape pairwise trees anchored at index 0; compensated sums for long accumulations
Floating-point atomics The main source of GPU nondeterminism (FACT, Appendix H) Banned. Count, exclusive scan, write. Integer atomics only for maximum, minimum, or and wrapping add
Sorting Ties and unstable sorts reorder contacts Stable sorts with total order keys (value, then index); one radix-sort primitive
Hardware texture filtering Fixed-point weights differ by vendor Manual interpolation in kernel source
Subnormals on Metal The specification permits flushing; no control (FACT) Measured at GATE-SPARK-01; D1 claimed only where behaviour matches
Hash-map iteration Order is not specified Not allowed to affect outputs
Warm starts Solver state carries across steps Warm-start buffers are part of every snapshot
Thread scheduling Completion order varies Results are placed by index, never by arrival
Toolchain and driver changes Code generation moves The kernel cache is keyed by compiler and driver; a change re-runs spark check
Strict inequalities in floating point 2²⁴ + 1 is not an f32; a recorded pilot mislabeled 36% of near-boundary cases Exact or guarded kernels (§21.8)
Batch composition A batch-wide reduction that feeds state makes a world depend on its neighbours §35.5
Padding (−0.0) + (+0.0) = +0.0; a padded contact left active changes the physics Trailing padding with the neutral element; masks on every kernel, reward and gradient
Arithmetic masks 0·∞ is NaN Masks are selects
Error-free transformations Metal's default mode may contract multiply-adds and assume no NaNs (FACT: ft-kernel-metal documentation); a folded error term is zero Safe mode; start-up canaries (§21.14)
Minimum, maximum, NaN, signed zero Built-ins differ on NaN and on ±0 between languages det_min and det_max by explicit selects; any NaN marks the world Faulted
Integer overflow and shifts Signed overflow and oversized shifts are undefined in both source languages Unsigned arithmetic and masked shifts in exact kernels
Quantile sketches Insertion order changes the sketch (§21.20) Integer log-bin histograms: the bin is the sign, the exponent and the top five mantissa bits of the f32 pattern, so a bin is at most 3.2% wide; counts are integers; merging is addition; a quantile is reported as a bin interval
Learner reductions and minibatch order Gradient sums and permutations by arrival change the weights Fixed-order tree sums over sample index; permutations from address-keyed streams; otherwise the labeled D4 profile
Gradient reduction across shards A mean of shard means weights samples by shard size (§36.13) A sum of sums in device order, divided by the global valid count
CPU worker count and tile size A tree shaped by tile boundaries changes with the worker count Tree shape is a function of logical index only
Draws consumed in sequence One extra element shifts every later draw Counter-based streams keyed by address and by the physical entity perturbed (§35.9)
Stride conventions between array libraries Signed byte strides against unsigned element strides (FACT to recheck) Views carry stride unit and signedness (§21.5)
Quantized scene fields Sixteen bits over 20 m resolve 0.305 mm, which is 30.5 N at a stiffness of 10⁵ N/m Distance stored in a narrow band with a per-brick scale (§19)
Uniform regularization A shift ε·I is large for a rotor-inertia row and negligible for a base row Regularization is stated in scaled variables (§35.10)
Time Accumulating a step in floating point drifts and schedules lose phase Integer ticks and rational phase accumulators (§20); time is class D0
Retries and restores A retried step that keeps part of its state diverges A retry restores the whole committed state (§16, §21.9)
Device digests A 128-bit integer mix is not an identity A tripwire only; identities are BLAKE3 over committed snapshots on the host
Rewrites and specialization Reassociation changes bits Exact rewrites only; a specialized kernel is bit-identical to its generic one
Cell crossings Adding the cell side to a small negative offset rounds Centred offsets (§35.8)

35.8 Anchored positions

Single precision resolves about 7.6 µm at 100 m from the origin and worse beyond, which is too coarse for contact (GB-04). A world position is therefore an integer cell index plus an f32 offset from the centre of a cell of side 8 m. The offset lies in [−4, 4), where resolution is at worst 2⁻²² m, about 0.24 µm. Scene fields are addressed cell-relative. Quantities that are differences (gaps, link-relative poses, velocities) never touch absolute position.

Re-anchoring is exact (PO-NUM-01). When the magnitude of an offset reaches 4, K-anchor subtracts or adds 8.0 and moves the integer index by one. For 4 ≤ |x| ≤ 16 the shifted value is exact by Sterbenz's lemma (the difference of two floats within a factor of two of each other is representable). Offsets are checked every step and a body cannot cross 12 m in one step, so the precondition holds; a step that would break it refuses (FROB-REF-249). The obligation is discharged exhaustively: the two binades and two signs hold 2²⁵ values.

Why the offset is centred. With offsets kept in [0, 8), a body that drifts to −2⁻²⁵ m must add 8. The number 8 − 2⁻²⁵ is not an f32 (the spacing just below 8 is 2⁻²¹), so the sum rounds to 8.0: the position has moved and has left the cell's range. The centred convention never adds the cell side to a small number.

Cell indices are signed 32-bit integers per axis and a scene declares its bounds. The difference of two anchored positions is the float difference of the offsets plus 8 times the integer difference of the cells, and is formed only between neighbours. GATE-SPARK-05 runs the precision battery at distances up to 1 km from the origin.

35.9 Randomness by address

Counter-based streams keyed by logical identity: (WorldAddr, step, stream_id), built on FrankenSim's Philox streams (FACT: fs-rand). No device-side generator state exists. Common random numbers for a comparison are therefore free: two designs given the same addresses see the same world.

Additions (DESIGN):

  • Draws are exact. Philox is 32-bit integer arithmetic, so a draw is class D0 on every backend. A uniform f32 in [0, 1) is (x ≫ 8)·2⁻²⁴, which is exact (PO-NUM-03). Normal variates follow fs-rand's construction, ported to the dialect under the det_* rule.
  • A draw depends on nothing else. Not on slot, bucket, variant index, shard, device or batch size.
  • Streams are registered. stream_id names a purpose (disturbance, each sensor's noise, reset pose, each randomized parameter, action sampling, minibatch permutation) in a registry. Adding a stream changes no draw of any other stream.
  • Streams are keyed by the physical entity they perturb, wherever that entity exists independently of the design: a gust by time, floor friction by tile, a sensor's noise by the hash of its port name. Pairing by seed is otherwise not pairing by physical field. The smallest case: a sequential stream r₀, r₁, r₂. Design A draws a push (r₀) and a friction (r₁). Design B has one more leg and draws a per-leg value first, so its push is r₁ and its friction r₂. The seeds are equal and the two designs were pushed differently.
  • Proposal draws carry exact weights. Because every draw is addressed, a disturbance drawn from a tilted proposal has an importance weight that is computed, not estimated (§26).

35.10 Diagonal scaling for single-precision contact solves

The convex solve factors H = A + JᵀGJ per world in single precision (§16). Two things matter: the condition number that governs the accuracy of that solve, and the meaning of the solver's tolerances.

FACT (textbook results; Appendix H). Cholesky factorization without pivoting is invariant under two-sided scaling by powers of two: absent overflow and underflow, the computed factor of D·H·D is exactly D times the computed factor of H, since every intermediate is scaled by an exact power of two. The attainable accuracy of the direct solve is therefore governed by the condition number of the unit-diagonal-scaled matrix whether or not the scaling is applied. Scaling to unit diagonal is within a factor n of the best diagonal scaling of a symmetric positive definite matrix.

FACT (two recorded computations on reference robots with simplified contact rows; Appendix H). For a 29-joint humanoid over 201 poses, κ₂(H) is 3·10⁴ to 1·10⁵ raw and 2·10³ to 9·10³ scaled at default contact impedance, and 2·10⁶ to 3·10⁶ raw and 2·10⁵ to 5·10⁵ scaled near-rigid. Over sampled contact states of three robots the scaled medians are 1.2·10³ to 2.7·10³. With unit round-off 6.0·10⁻⁸, κ·u in scaled variables is 1·10⁻⁴ to 5·10⁻⁴ at default impedance and 0.012 to 0.03 near-rigid.

So the plausible shortcut "scale the matrix to make the solve accurate" is wrong: scaling does not change one bit of a Cholesky solve. What it changes is everything around the factorization. DESIGN: the solve stage runs in scaled variables.

  1. Exact scaling. K-scale sets D_ii = 2^(−⌈e_i/2⌉), where e_i is the binary exponent of H_ii read from its bit pattern. The scaled diagonal lies in [½, 2). No logarithm is evaluated and no rounding is introduced.
  2. Tolerances in scaled variables. The gradient-norm and decrement tests, the sufficient-decrease test of the line search, the pivot breakdown test and the regularization are all stated on the scaled system. One tolerance then has one meaning across rows whose raw diagonals span four orders of magnitude, and across the different robots of a mixed batch.
  3. A conditioning figure in the certificate. The step reports κ̂ = (max L̂_ii / min L̂_ii)², read off the factor at no cost, and κ̂·u. It is a lower bound on κ₂ of the scaled matrix and is named a lower bound. The referee sampler computes κ₂ in f64 for sampled worlds, and the ledger holds both.
  4. Precision profiles. default uses no refinement. stiff, for near-rigid contact and the optional hard-contact law, adds one step of iterative refinement whose residual is accumulated in double-float (K-refine). One step multiplies the relative error by about κ·u, taking an error of order 0.03 to order 10⁻³. The number of refinement steps is a constant of the shape class, so the launch graph stays static.
  5. A limit. A world whose κ̂·u exceeds 0.1 is marked Degraded: ill-conditioned; under the refuse policy the run refuses (FROB-REF-248).

The smallest case for item 2: H = diag(1, 10⁻⁸) and a residual (0, 10⁻⁷). Its norm passes an absolute tolerance of 10⁻⁶ while the error in the second unknown is 10⁻⁷/10⁻⁸ = 10. In scaled variables, with D = diag(1, 10⁴), the residual is (0, 10⁻³) and the test fails, as it should.

35.11 Refusals

A refusal has a stable code, a message and ranked repairs. Families: 1xx design and model; 2xx numerics and integrator admission; 3xx scene; 4xx control and estimation; 5xx learning and evaluation protocol; 6xx backend and device; 7xx budget and capability; 8xx evidence and protocol (for example citing a provisional result as established); 9xx internal. Three standing rules: refuse, never truncate; a budget stop publishes nothing partial as a result; a misspelt parameter is a refusal, not a default. One clarification of the second rule: a mission query (§25) answers with certified cells, each of which is a complete result with its own certificate, so a budget stop returns the cells certified so far together with the unresolved remainder, and the envelope says that the answer is incomplete. No cell is ever returned half-certified. The consolidated registry of refusal codes is Appendix O.

Codes defined by this chapter and by §36 (the 6xx codes are in §21.21):

Code Condition
FROB-REF-247 The determinism class a verb requires is not attainable on the requested placement
FROB-REF-248 A solve is ill-conditioned beyond its precision profile, under the refuse policy
FROB-REF-249 An anchored offset left its range in one step, or a cell index left the scene bounds
FROB-REF-250 A kernel is not dimensionally homogeneous (build time)
FROB-REF-251 A banned operation appears in kernel source (build time)
FROB-REF-252 A strict predicate in floating point without an exact form or a certified guard (build time)
FROB-REF-561 A paired statistic was requested with an arm of class D4 or of unrecorded class, or as paired across device families without class D1
FROB-REF-562 Attribution or a counterfactual fork was requested below class D2
FROB-REF-761 A job exceeds the compute envelope in force (§36.14)
FROB-REF-861 A number from a fast-training run was cited as evidence, or a policy trained under it lacks a stricter re-evaluation
FROB-REF-862 A performance figure without a complete disclosure block (§36.7)
FROB-REF-863 A performance ratio without a paired lower bound (§36.6)
FROB-REF-864 A device result cited without a current CPU-versus-device check
FROB-REF-865 A summary statistic that omits what it excluded
FROB-REF-866 A speed experiment requested without its computed ceiling, or with a ceiling below its goal (§36.5)

35.12 A number never claims more than its certificate

Every numerical result records what bounds it: solver gradient norm and iteration count, maximum constraint and closure violation, group-membership error, the energy-ledger residual, the number of worlds marked Degraded or Faulted. A summary statistic over worlds states how many worlds it excludes and why.

Added to the record of every step and every result (DESIGN):

  • the conditioning lower bound κ̂ and κ̂·u, the precision profile and the refinement steps taken;
  • the fallback fraction of each guarded predicate;
  • capacity headroom: the largest active contact and row counts against their capacities;
  • the determinism class, the profile, the placement receipt and the fingerprint digest;
  • for a decision, its margin (§35.4).

And three rules of wording, enforced by schema. A lower bound is called a lower bound and an estimate an estimate. A quantile read from a histogram is an interval. A device digest is a tripwire classed Tested; it is never called an identity, and no result is called verified.

35.13 Invariants and proof obligations

ID Statement Verification
INV-NUM-01 Results are placed by index or address, never by arrival GATE-SPARK-15 under slot permutation
INV-NUM-02 Padding neutrality: outputs of active lanes are bit-identical at any capacity at or above the active count GATE-SPARK-15 at two capacities
INV-NUM-03 Given a policy, a world's trajectory is a function of its address, design, scene and task alone GATE-SPARK-15
INV-NUM-04 A random draw is a function of (address, step, stream) alone Draw tables compared across slots, buckets, shards and backends
INV-NUM-05 A summary statistic states its exclusions Schema check; FROB-REF-865
INV-NUM-06 The class of a result is the weakest class among what produced it Type-level in frob-core; a compile-fail test for any conversion upward
PO-NUM-01 Re-anchoring is exact for offsets of magnitude 4 to 16 Exhaustive over 2²⁵ values on each backend
PO-NUM-02 The computed Cholesky factor is invariant under power-of-two scaling Bitwise metamorphic test on 10⁴ matrices per backend
PO-NUM-03 The integer-to-uniform conversion is exact and lies in [0, 1) Exhaustive over 2²⁴ values

35.14 Known answers and negative controls

Known answer Expected Negative control that must be caught
One world alone, in 100 random co-tenancies, at two capacities, on two shards Identical digests A kernel with a batch-wide floating reduction feeding state
Sum of the one active element −0.0 at two capacities −0.0 both times Padding with +0.0
Two worlds observing 1 and 3 Each world's actions equal those of its isolated run A normalizer updated inside the segment
Success predicate at 1.0000001 m against 0.9999999 m Reported as discordant and boundary-sensitive A check that compares states only
Offset −2⁻²⁵ m in a [0, 8) cell The centred convention shifts exactly The uncentred shift returns 8.0
Sequential draws with one extra leg Addressed streams give both designs the same push A sequential stream gives r₀ against r₁
Factor of D·H·D against D times the factor of H Bit-identical A scaling by 10 instead of a power of two
H = diag(1, 10⁻⁸) with residual (0, 10⁻⁷) The scaled test fails The unscaled absolute test passes
The set {0, 3, 5, 9} histogrammed in two orders Identical histograms A closest-pair sketch (§21.20)
A fast-training policy evaluated twice under device-replay Identical digests Citing its training return is refused

35.15 Gates

Gate Claim Threshold (TARGET) Earliest
GATE-SPARK-15 Known answer For 1,000 episodes of each reference workload, per-world digests are identical alone, in 100 random co-tenancies, under slot permutation, at two bucket capacities and on one and two shards, on batch-cpu and on each device family; draws depend on address alone; both negative controls of §35.14 are caught M2
GATE-SPARK-16 Fidelity For every class D3 pair of placements used as evidence, on a frozen ensemble of at least 4,096 worlds per reference workload: upper bound on discordant decisions at most 1%; bound on net one-sided discordance at most 0.5%; continuous outcomes equivalent inside ledger margins; first divergences recorded. A planted bias of one decision in fifty must fail M2
GATE-SPARK-17 Known answer Each licence refusal of §35.3 fires on its planted case (a D4 arm in a paired comparison; attribution below D2; a cited training return; a cross-family pair demanded as paired), and a policy trained under fast-training re-evaluates to the same digest twice M3
GATE-SPARK-18 Known answer PO-NUM-02 holds on each backend; on 10⁴ contact states per reference robot the relative error of the f32 Newton step against f64 is at most 10·κ₂·u, and at most 10⁻³ on near-rigid states under the stiff profile; κ̂ never exceeds the κ₂ computed in f64 M2

GATE-SPARK-01, -02, -05 and -08 (§21.24) are the other determinism and precision gates. If class D1 cannot be attained, KILL-D1 (§21.15) applies.

35.16 Open questions

  • OPEN-NUM-1. Whether an explicit fused multiply-add is ever admitted in batched kernels. Default: no. Evidence: the probe's record of fused rounding on each family. Blocks nothing.
  • OPEN-NUM-2. The decision-equivalence thresholds (1% and 0.5%) and the ensemble size. Default: as stated, frozen once at M2 with a written rationale. Blocks GATE-SPARK-16.
  • OPEN-NUM-3. The cell side. Default: 8 m, centred. Blocks nothing.
  • OPEN-NUM-4. The limit for Degraded: ill-conditioned and the refinement count of the stiff profile. Default: κ̂·u above 0.1; one step. Decided by GATE-SPARK-18.