APEX, in one page.

Fourteen IBM tools. Per-tool wire-up status, honest.

Three wired at HEAD (Granite Instruct 4.1 8B + Granite 4.0 Nano 350M + Granite Embedding R2 via HF Inference Providers). Nine at frontend-integration with canonical type contracts and backend swap-points per Stream M.3 + the D-058 wave-46 expansion (LangGraph + Granite MCP Gateway + ContextForge as the orchestration runtime per D-017 G7 + D-054 + Granite Instruct 4.1 3B chat-routing + Granite Speech 4.1 2B-Plus Watson STT proxy preview; Langflow retained as export-graph artifact). Two build-time accelerators (Docling library + Mellea v0.5.0 IVR-loop architectural slot; Mellea is build-time architectural inspiration only, not a runtime dependency).

Granite-Docling 258M

Integration

FIA COA PDF to structured JSON parser

Docling library latest

Accelerator

Open-source IBM Docling conversion + table-extraction

Granite Vision 4.1 4B

Integration

SRO + Britcar timing-sheet PDF to CSV

Granite TimeSeries TTM r2.1 + fine-tune

Integration

Track 1 of D-010 three-track ensemble (frozen-backbone + channel-mix decoder fine-tune per D-052)

Granite FlowState r1.1 18.5M

Integration

Track 2 of D-010 (sampling-rate-invariant continuous-time SSM at 50 Hz)

IBM TSPulse 1M

Integration

Polyphase time-frequency anomaly detector (D-016 Layer 2)

Granite Embedding R2 149M + 47M

Wired

Live dense RAG rerank via HF Inference Providers (granite-embedding-30m-english feature-extraction); lexical fallback (D-016)

Granite Instruct 4.1 8B

Wired

Race-engineer narrator: live /analyze coaching report + chat via OpenRouter (wave-64; verified phase=real in prod)

Granite Guardian 4.1 8B

Integration

BYOC custom-rule audit + D-024 physics-confidence downgrade

Granite 4.0 Nano 350M

Wired

In-browser WebGPU edge model via Transformers.js (D-019 + D-021)

LangGraph + MCP + ContextForge runtime (Langflow export-graph)

Integration

Orchestration runtime per D-017 G7 + D-026 (Langflow demoted to export-graph facade per wave-30)

Granite Instruct 3B 4.1 3B

Integration

Fast-path AICopilotChat routing for simple queries (D-058 wave-46 add)

Granite Speech 4.1 2B-Plus

Integration

Speaker-attributed ASR + word-level timestamps replacing Web Speech (D-058 wave-46 add)

Mellea >=0.5.0

Accelerator

Instruct-Validate-Repair tri-agent critic loop scaffold on narrator (IBM Research open source library, Apache 2.0); flips to WIRED when Vinh M3-V5.2 ships Mellea-orchestrated narrator path

Convergence 14 · the safety contract.

Decision-log D-A names the Convergence 14 fixture suite as the load-bearing safety contract. The catalogue below is the Stephen-lane display of the 14-fixture safety-contract specification: 4 at the Stage 1 convex QP boundary, 4 at the Stage 2 feasibility-filter boundary, 4 at the Stage 3 Granite Guardian BYOC text-audit boundary, plus 2 round-trip integrity fixtures that close the convergence loop. The Vinh-lane assertion suite at app/backend/tests/test_serializer.py lands per PLAN rows 2.9c + 4.2, with fixture files at app/backend/tests/fixtures/convergence-14/. Each fixture asserts the violation, the serializer output, and the Granite Guardian verdict. Click any row to see the expected Guardian verdict reason and the serialized violation-log excerpt the fixture targets.

1. C14-01Friction ellipse lateral

   flag

   Lateral g of 1.8 with brake near zero and steering near zero exceeds the constant-mu friction ellipse on a dry-tarmac circuit.

   Violation class

   Friction ellipse

   Detection stage

   Stage 1 · convex QP

   COA simul

   not applicable

   Fixture file

   app/backend/tests/fixtures/convergence-14/C14-01_friction_ellipse_lateral.json

   Guardian verdict reason

   Stage 1 QP projected lat_g from 1.8 g to friction-ellipse boundary (1.0 g at mu_v = 1.0); residual delta exceeded the 0.1 g flag threshold.

   stage:1 class:friction_ellipse t:7 channel:lat_g raw:1.8 projected:1.0 mu_v:1.0 delta_g:0.8

2. C14-02Friction ellipse longitudinal

   flag

   Longitudinal g of negative 1.5 (braking) combined with steering of 0.3 rad exceeds the friction-ellipse envelope at the corner-entry boundary.

   Violation class

   Friction ellipse

   Detection stage

   Stage 1 · convex QP

   COA simul

   not applicable

   Fixture file

   app/backend/tests/fixtures/convergence-14/C14-02_friction_ellipse_longitudinal.json

   Guardian verdict reason

   Stage 1 QP projected long_g from -1.5 g to -0.95 g at the friction-ellipse boundary given lat_g = 0.30 g.

   stage:1 class:friction_ellipse t:12 channel:long_g raw:-1.5 projected:-0.95 mu_v:1.0 delta_g:0.55

3. C14-03Forward-Euler monotonicity

   flag

   Forecast claims speed increases from 40 m/s to 50 m/s with long_g pinned at zero across two 1-Hz mini-sectors. Forward-Euler equality fails: speed_{t+1} must equal speed_t + a_long_t * dt.

   Violation class

   Forward-Euler

   Detection stage

   Stage 1 · convex QP

   COA simul

   not applicable

   Fixture file

   app/backend/tests/fixtures/convergence-14/C14-03_forward_euler_monotonicity.json

   Guardian verdict reason

   Stage 1 QP recoupled speed[t+1] = speed[t] + a_long[t] * dt; projected speed_{t+1} from 50.0 to 40.0 m/s. Long_g consistency restored.

   stage:1 class:forward_euler t:5 channel:speed_mps raw:50.0 projected:40.0 a_long_t:0.0 dt:1.0

4. C14-04Jerk bound

   flag

   Long_g jumps from +0.6 g at t=8 to -0.8 g at t=9 in one 1-Hz mini-sector. Delta of 1.4 g/s exceeds the jerk bound at 8 m/s^3 (approximately 0.815 g/s). 1 Hz fixture is a deliberate demo simplification per arch-spec sampling-rate caveat; per wave-30 D-011 multi-frequency coexistence, production telemetry runs the jerk-bound at >=10 Hz on the raw 50 Hz upstream signal before mini-sector aggregation.

   Violation class

   Jerk bound

   Detection stage

   Stage 1 · convex QP

   COA simul

   not applicable

   Fixture file

   app/backend/tests/fixtures/convergence-14/C14-04_jerk_bound.json

   Guardian verdict reason

   Stage 1 QP jerk-bound clamp engaged; projected long_g[9] from -0.8 g to -0.2 g (prev +0.6 g minus the per-step jerk budget of 0.8 g) to respect the jerk bound at 8 m/s^3 over the 1.0 s mini-sector (D-011 multi-frequency coexistence framing).

   stage:1 class:jerk_bound t:9 channel:long_g raw:-0.8 projected:-0.2 prev:+0.6 jerk_max_m_per_s3:8 jerk_max_g_per_s:0.815

5. C14-05Bicycle: lateral g without steering

   reject

   Forecast claims lat_g = 0.5 g with steering_rad = 0.0 at speed 15 m/s. Bicycle-model coupling makes lateral g impossible without nonzero steering at low slip.

   Violation class

   Bicycle model

   Detection stage

   Stage 2 · feasibility filter

   COA simul

   not applicable

   Fixture file

   app/backend/tests/fixtures/convergence-14/C14-05_bicycle_no_steering.json

   Guardian verdict reason

   Stage 2 feasibility filter rejected forecast: bicycle-model expects steering_rad approx 0.059 rad for lat_g = 0.5 g at speed 15 m/s with wheelbase 2.7 m (derived from delta = atan(a_lat * L / v^2)); observed steering 0.0 rad falls outside the slip_tol_rad = 0.03 tolerance band.

   stage:2 class:bicycle_model t:14 lat_g_raw:0.5 steering_rad:0.0 speed_mps:15 expected_steering_rad:0.059 slip_tol_rad:0.03

6. C14-06Bicycle: magnitude mismatch

   reject

   Steering of 0.005 rad at speed 30 m/s would yield lat_g of about 0.17 g under low-slip kinematic approximation (a_lat = v^2 * tan(delta) / L; here 900 * 0.005 / 2.7 / 9.81 = 0.170 g). Forecast instead asserts lat_g = 0.8 g. Magnitude mismatch flags the bicycle-coupling audit.

   Violation class

   Bicycle model

   Detection stage

   Stage 2 · feasibility filter

   COA simul

   not applicable

   Fixture file

   app/backend/tests/fixtures/convergence-14/C14-06_bicycle_magnitude.json

   Guardian verdict reason

   Stage 2 feasibility filter rejected forecast: bicycle-model predicted lat_g approx 0.17 g for steering 0.005 rad at 30 m/s; observed 0.8 g exceeds the magnitude band derived from the slip_tol_rad = 0.03 steering tolerance.

   stage:2 class:bicycle_model t:18 lat_g_raw:0.8 lat_g_predicted:0.17 steering_rad:0.005 speed_mps:30 slip_tol_rad:0.03

7. C14-07COA gate: simultaneity forbidden

   reject

   Telemetry row shows throttle = 0.4 and brake = 2.4 MPa simultaneously, but the COA-derived c_overlap flag (derived from approved hand-control hardware specifications) is set to 0 (simultaneity not permitted). Stage 2 complementarity check fires reject.

   Violation class

   COA simultaneity

   Detection stage

   Stage 2 · feasibility filter

   COA simul

   forbidden

   Fixture file

   app/backend/tests/fixtures/convergence-14/C14-07_coa_simul_forbidden.json

   Guardian verdict reason

   Stage 2 feasibility filter rejected telemetry: brake-throttle simultaneity observed (throttle 0.4 + brake 2.4 MPa) but COA-derived c_overlap flag = 0 (driver's approved hand-control hardware specifications do not permit simultaneous brake-throttle actuation).

   stage:2 class:coa_simultaneity t:11 throttle:0.4 brake_pa:2.4e6 c_overlap:0 derivation:"hardware_spec"

   Replay: Flip COA overlap flag
8. C14-08COA gate: simultaneity permitted (positive)

   approve

   Same telemetry as C14-07 (throttle = 0.4 with brake = 2.4 MPa) but COA-derived c_overlap flag = 1 (driver's approved electronic hand-control hardware permits simultaneous brake-throttle actuation through corner entry). Stage 2 approves; report renders the tuning recommendation.

   Violation class

   COA simultaneity

   Detection stage

   Stage 2 · feasibility filter

   COA simul

   permitted

   Fixture file

   app/backend/tests/fixtures/convergence-14/C14-08_coa_simul_permitted.json

   Guardian verdict reason

   Stage 2 feasibility filter approved telemetry: brake-throttle simultaneity observed (throttle 0.4 + brake 2.4 MPa) within the COA-derived c_overlap permit window (derived from approved hand-control hardware specifications); combined force within friction envelope (lat_g 0.81 + long_g -0.22).

   stage:2 class:coa_simultaneity t:11 throttle:0.4 brake_pa:2.4e6 c_overlap:1 derivation:"hardware_spec" verdict:approve combined_g:0.84

   Replay: Flip COA overlap flag
9. C14-09Physical envelope: negative speed

   reject

   Forecast tensor channel speed_mps contains a value of -3.0 at t=22. Granite Guardian's physical-envelope BYOC rule fires reject on any speed < 0.

   Violation class

   Physical envelope

   Detection stage

   Stage 3 · Granite Guardian audit

   COA simul

   not applicable

   Fixture file

   app/backend/tests/fixtures/convergence-14/C14-09_negative_speed.json

   Guardian verdict reason

   Granite Guardian rejected forecast: speed_mps = -3.0 at t=22 violates physical-envelope rule (speed_mps >= 0). BYOC rule physical_envelope_speed_nonnegative.

   stage:3 class:physical_envelope t:22 channel:speed_mps raw:-3.0 byoc_rule:"physical_envelope_speed_nonnegative"

10. C14-10Physical envelope: circuit max speed

    flag

    Forecast tensor channel speed_mps contains a value of 110 m/s (about 396 km/h) on a circuit whose homologated maximum is 95 m/s. Guardian fires flag.

    Violation class

    Physical envelope

    Detection stage

    Stage 3 · Granite Guardian audit

    COA simul

    not applicable

    Fixture file

    app/backend/tests/fixtures/convergence-14/C14-10_speed_exceeds_circuit_max.json

    Guardian verdict reason

    Granite Guardian flagged forecast: speed_mps = 110 at t=17 exceeds circuit-homologated maximum 95 m/s. BYOC rule physical_envelope_speed_circuit_max.

    stage:3 class:physical_envelope t:17 channel:speed_mps raw:110.0 circuit_max_mps:95.0 byoc_rule:"physical_envelope_speed_circuit_max"

11. C14-11Physical envelope: gear vs engine state

    flag

    Telemetry row has rpm = 0 with speed = 38 m/s in gear 4. Drivetrain decoupled while in gear violates the engine-state envelope. Guardian flags.

    Violation class

    Physical envelope

    Detection stage

    Stage 3 · Granite Guardian audit

    COA simul

    not applicable

    Fixture file

    app/backend/tests/fixtures/convergence-14/C14-11_drivetrain_decoupled.json

    Guardian verdict reason

    Granite Guardian flagged telemetry: rpm = 0 at t=9 while speed_mps = 38 and gear = 4 violates the engine-state envelope (rpm > 800 when gear > 0 and speed_mps > 5). BYOC rule physical_envelope_engine_engaged.

    stage:3 class:physical_envelope t:9 rpm:0 speed_mps:38 gear:4 byoc_rule:"physical_envelope_engine_engaged"

12. C14-12Physical envelope: throttle saturation

    flag

    Forecast tensor reports throttle_pct = 1.07 at t=14 (above the 0.0..1.0 unit-normalized envelope). Guardian flags as a clamp candidate.

    Violation class

    Physical envelope

    Detection stage

    Stage 3 · Granite Guardian audit

    COA simul

    not applicable

    Fixture file

    app/backend/tests/fixtures/convergence-14/C14-12_throttle_saturation.json

    Guardian verdict reason

    Granite Guardian flagged forecast: throttle_pct = 1.07 at t=14 outside unit envelope [0.0, 1.0]. BYOC rule physical_envelope_throttle_unit_bound.

    stage:3 class:physical_envelope t:14 channel:throttle_pct raw:1.07 byoc_rule:"physical_envelope_throttle_unit_bound"

13. C14-13Serializer round-trip integrity

    approve

    Synthetic violation tensor written to text log, parsed back to tensor, re-serialized; the parsed-then-re-serialized log is byte-identical to the original. Guardian's verdict on both forms is identical.

    Violation class

    Serializer integrity

    Detection stage

    Stage 3 · Granite Guardian audit

    COA simul

    not applicable

    Fixture file

    app/backend/tests/fixtures/convergence-14/C14-13_serializer_roundtrip.json

    Guardian verdict reason

    Granite Guardian verdict is verdict-stable across the serialize-deserialize-reserialize round trip. Convergence-14 byte-equality assertion passes.

    stage:3 class:serializer_integrity closure:round_trip byte_equal:passes guardian_verdict_round_trip:approve

14. C14-14Convergence: COA + physics + Guardian closure

    approve

    End-to-end Sarah Reynolds fixture: COA-permitted simultaneity row (C14-08) passes Stage 2 + reaches Stage 3 + receives Guardian approve verdict + tuning recommendation rendered with COA citation. Closes the full safety-contract loop.

    Violation class

    Serializer integrity

    Detection stage

    Stage 3 · Granite Guardian audit

    COA simul

    permitted

    Fixture file

    app/backend/tests/fixtures/convergence-14/C14-14_full_loop_sarah.json

    Guardian verdict reason

    Granite Guardian approved end-to-end fixture: Stage 2 COA-derived c_overlap permit accept + Stage 3 BYOC audit approve + tuning recommendation rendered with citation to the driver's approved hand-control hardware specifications.

    stage:3 class:serializer_integrity closure:end_to_end sarah_fixture:approved c_overlap:1 derivation:"hardware_spec"

Grid is display-only. The Vinh-lane test suite at the path above is the assertion source of truth.

Safety alignment

ISO 26262 vocabulary set, line by line.

Team BRIT, a professional team that races disabled drivers in UK endurance championships, cites ISO 26262 in engineering correspondence as the framework it applies when reviewing brake-by-wire interfaces, custom CAN gateway controllers, and adaptive throttle / brake / clutch logic, and granted permission to be cited. APEX maps its coaching-surface invariants against the standard’s vocabulary set so the things a functional-safety reviewer flags are the same things APEX’s Guardian + scrubber + bounded retry-loop + AbortSignal threading address by construction. We do not certify APEX against ISO 26262; we claim that the vocabulary set of the standard is the right anchor for the coaching-surface invariants we already enforce. Detailed treatment in paper section 3.8.

CAN signal plausibility check

Telemetry intake validates every channel against the canonical 14-channel shape contract in shared/contracts/shapes.py before any forecast call.

Sensor disagreement handling

V1 NumPy validator + V2 cvxpylayers projector each detect physically inconsistent input combinations (friction-ellipse + jerk-bound + tier-1 simultaneity-gate) and reject before narration.

Fault-detection and latch-reset

Self-Correcting Retry Loop on the narrator path detects forbidden regulatory-anchor patterns + emits a system-level retry directive enumerating the violations; bounded 2-retry budget.

Watchdog and timeout implementations

AbortSignal.timeout(3000) on every Vinh-backend wire-flip + AbortController on OpenRouter completion + 12-second timeout in the VS Code extension webview; consumer disconnection aborts the request.

Deterministic behavior under communication faults

Every wire-flip helper returns the canned-fallback payload on any upstream 5xx or fetch failure; the engine field declares which path served + the X-Apex-*-Engine header surfaces it to runtime logs.

State-machine review for unintended transitions

Discriminated-union state machines on the client (CoachVoicePlayback, CoachCodePanel, AICopilotChat, TelemetryUploadPanel) make impossible states a TypeScript compile error; no defensive runtime branch needed.

Failure Mode and Effects Analysis (FMEA) on driver-input paths

Convergence 14 fixture suite covers the 14 failure modes the V1 + V2 + Guardian stack must reject; each fixture asserts the violation string + the Guardian verdict so the cause/effect map is testable, not asserted.

Vocabulary set sourced from Team BRIT engineering correspondence, cited with permission (2026-05-29); mapping to APEX invariants is the contribution.

Confidence decomposition

Confidence, broken into the signals that earn it.

APEX does not roll a coaching report into a single invented confidence number. It surfaces each dimension next to the pipeline signal that produces it, so you can see where the confidence comes from and where it does not. One dimension below is an integration-tier preview, labelled as such.

Demo fixtures on this judges tour: each dimension is populated from a representative fixture that mirrors live pipeline output. In a live session each is computed per-run.

Physics feasibility

converged · max resid 0.0011

Source: Stage 1 + 2 projection residual norms (CvxpyLayer QP convergence)

Guardian safety

APPROVE

Source: Granite Guardian 4.1 BYOC custom-rule verdict

Forecast certainty

±0.456 s mean band (1.36% of pace)

Source: Next-session forecast 90% envelope width, post-projection

Physics-model confidence

in-distribution · d 1.84 / p95 2.50

Source: Mahalanobis physics-confidence detector (D-024)

Integration-tier preview: demo fixture, not a live per-session measurement at HEAD.

Each dimension is derived from a real pipeline signal (projection residuals, Guardian verdict, forecast envelope width) or labelled as an integration-tier preview. No weighted-sum overall score is shown, because APEX does not own a calibrated weighting and showing one would be theatre.

Eight-tier physics implementation.

Wave-30 decision D-015 locked the eight-tier physics stack in-scope for the submission. Each tier surfaces below as one tile with its plain-text formula, canonical input channels, expected outputs, and the handler that owns it inside the unrolled SCP outer loop per D-012. The class-to-handler binding lives on the ExtendedPhysicsFixture discriminated union at app/shared/types.ts; mis- binding a non-convex tier to the inner cvxpylayers iterate is a TypeScript compile error. Hover any tile to reveal the architecture-spec Appendix W30 cross-reference.

1. EP-013D track geometry

   tier 01

   Track elevation + bank + pitch project the gravity vector onto the vehicle frame so the effective vertical load varies corner-by-corner instead of being a flat 9.81 m/s^2.

   Handler

   SCP outer-loop linearisation

   Canonical inputs

   pitch_rad + bank_rad + speed_mps

   Formula + outputs

   g_eff = R(pitch, bank) * [0, 0, -g]; the per-step rotation matrix from the track-frame pitch + bank pulls the gravity vector into the vehicle frame so downstream Tier 4 + Tier 7 use the correct vertical-load component.

   Outputs

   • 01.g_eff (3-vector per step)
   • 02.per-step rotation matrix R
   • 03.vertical-load adjustment for Tier 4

   Arch-spec ref: Appendix W30 Layer 4 Tier 1 (pitch + bank channels added wave-30 D-016 channel expansion)

   Decision-log: D-015

2. EP-02Aerodynamics

   tier 02

   Aero downforce + drag scale with v^2; downforce adds to vertical load (Tier 4), drag opposes forward acceleration (Tier 8).

   Handler

   SCP outer-loop linearisation

   Canonical inputs

   speed_mps

   Formula + outputs

   F_z_aero = 0.5 * rho * Cl * A * v^2; F_drag = 0.5 * rho * Cd * A * v^2. Linearised around the previous SCP outer-iterate to feed back into the inner cvxpylayers convex QP.

   Outputs

   • 01.F_z_aero (downforce contribution to Tier 4)
   • 02.F_drag (deceleration contribution to Tier 8)

   Arch-spec ref: Appendix W30 Layer 4 Tier 2

   Decision-log: D-015

3. EP-03Adaptive hand-controls

   tier 03

   FIA Certificate of Adaptations parsed at onboarding derives the brake-throttle simultaneity flag + steering-lock limits + clutch-lever travel range; these feed the lexicographic Tier-2 + Tier-3 COA constraints per D-022 elastic-slack hierarchy.

   Handler

   COA constraint layer (lexicographic)

   Canonical inputs

   coa_simul_permitted + throttle_pct + brake_pa + steering_rad

   Formula + outputs

   If coa_simul_permitted == true: throttle * brake constraint relaxed; else: throttle * brake = 0 (able-bodied assumption). Steering-lock limit |steering_rad| <= steering_lock_max derived from COA hardware spec. Slack variable activates when corner geometry forces relaxation; Guardian audit surfaces the relaxed constraint in the coaching report.

   Outputs

   • 01.lexicographic COA constraint set (Tier 2 + Tier 3)
   • 02.per-corner slack variables
   • 03.Guardian audit annotations

   Arch-spec ref: Appendix W30 Layer 4 Tier 3 + D-022 lexicographic COA hierarchy

   Decision-log: D-015

4. EP-04Double-track load transfer

   tier 04

   Lateral acceleration shifts vertical load between left + right tires; longitudinal acceleration shifts between front + rear axles. Each tire's vertical load updates per-step so Tier 7 Pacejka uses the correct F_z.

   Handler

   SCP outer-loop linearisation

   Canonical inputs

   lat_g + long_g + speed_mps + fz_total (computed)

   Formula + outputs

   dFz_lat = m * a_y * h_cg / track_width; dFz_long = m * a_x * h_cg / wheelbase. Per-corner allocation respects axle-distribution + roll-stiffness split; fz_total channel (wave-30 D-016 addition) carries the result downstream.

   Outputs

   • 01.per-tire F_z (4 corners)
   • 02.fz_total adjustment for Tier 7

   Arch-spec ref: Appendix W30 Layer 4 Tier 4 (fz_total channel added wave-30 D-016)

   Decision-log: D-015

5. EP-05Tire thermal + degradation

   tier 05

   Two-mass thermal model evolves tire-core + tire-surface temperatures inside the SCP solver (NOT an input channel of the 14-channel tensor); peak friction coefficient mu_v modulates with T_surface + lap-count.

   Handler

   Internal state evolution

   Canonical inputs

   mu_v (consumed + updated) + speed_mps + lap-count metadata

   Formula + outputs

   Two-mass ODE: dT_core/dt = (T_surface - T_core) / tau_core + tire-heating from slip work; dT_surface/dt = (T_ambient - T_surface) / tau_surface + (T_core - T_surface) / tau_internal. mu_v(T_surface, lap_count) modulates peak friction; T_surface is per-step internal state (initial value = ambient-plus-warmup per circuit metadata).

   Outputs

   • 01.evolved T_surface internal state
   • 02.per-step mu_v adjustment for Tier 7

   Arch-spec ref: Appendix W30 Layer 4 Tier 5 (T_surface internal state clarification at line 391; not a 15th channel)

   Decision-log: D-015

6. EP-06Transient tire dynamics

   tier 06

   Stiff-ODE tire-relaxation transient (lateral slip catches up to steering input with time constant tau_y) is collapsed to steady-state algebraic per D-014 numerical-hazard resolution; full transient model reserved for offline validation only.

   Handler

   Steady-state algebraic substitution

   Canonical inputs

   lat_g + steering_rad + speed_mps

   Formula + outputs

   Original: tau_y * d(slip_y)/dt + slip_y = slip_y_steady_state. Steady-state algebraic substitution: LHS = 0 so slip_y = slip_y_steady_state. Solver runs the algebraic form inside the inner cvxpylayers convex QP iterate; stiff-ODE form skipped because cvxpylayers cannot handle stiff differential algebraic equations.

   Outputs

   • 01.algebraic slip_y consumed by Tier 7 Pacejka

   Arch-spec ref: Appendix W30 Layer 4 Tier 6 + D-014 numerical hazard resolution

   Decision-log: D-015

7. EP-07Full Pacejka combined-slip

   tier 07

   Pacejka Magic Formula produces tire forces F_x + F_y from longitudinal + lateral slip + vertical load + friction coefficient + tire-surface temperature; combined-slip boundary forms the heart-shape friction-ellipse generalisation in 2D.

   Handler

   SCP outer-loop linearisation

   Canonical inputs

   lat_g + long_g + fz_total + mu_v + speed_mps + steering_rad + T_surface (internal)

   Formula + outputs

   F_x, F_y = pacejka(s_x, s_y, F_z, mu_v, T_surface) via Magic Formula. Non-convex; the SCP outer loop linearises the Pacejka boundary around the previous iterate via first-order Taylor step; cvxpylayers inner solver enforces the linearised half-spaces as convex constraints. Three iterations unrolled per D-012.

   Outputs

   • 01.F_x, F_y per tire (4 corners)
   • 02.per-iterate Jacobian + residual norm for SCP convergence trace

   Arch-spec ref: Appendix W30 Layer 4 Tier 7 + D-012 unrolled SCP outer loop

   Decision-log: D-015

8. EP-08Kinematic integration

   tier 08

   Forward-Euler kinematic step ties acceleration to speed evolution + position-integration: speed[t+1] = speed[t] + a_long[t] * dt. Newton-compliant + lives in the convex inner iterate (the only tier that is natively convex).

   Handler

   SCP inner iterate (convex QP)

   Canonical inputs

   long_g + speed_mps + yaw_rate

   Formula + outputs

   speed[t+1] = speed[t] + a_long[t] * dt for all t in [0, prediction_length=30). Convex equality constraint inside the cvxpylayers QP. Jerk-bound activates at >=10 Hz per Rajamani; C14-04 demo fixture is 1 Hz simplification per arch-spec sampling-rate caveat.

   Outputs

   • 01.projected speed_mps trajectory across (B, 30) horizon
   • 02.kinematic integration residual

   Arch-spec ref: Appendix W30 Layer 4 Tier 8 + 2.9a convex inner iterate row in PLAN.md

   Decision-log: D-015

Grid is display-only. The Vinh-lane SCP outer-loop linearisation lands per PLAN row 2.18 with file paths at app/backend/apex/physics/tier_{1..8}.pyand app/backend/apex/physics/scp_outer.py.

The pipeline, paced for a judge tour.

Wave-45 Phase 5 Block C.2 close-out. Six stages cycle every 1.5 seconds: ingest -> TTM forecast -> V2 projector -> Guardian audit -> Instruct narration -> provenance footer. Full pipeline runs sub-60s on RTX 4060 per G8 budget; this visualization is the architecture paced for a 9-second judge tour.

Run Granite in your browser. The 7th shouldn't-be-possible move.

Wave-45 Phase 7 Block D close-out per D-053. Same Granite TimeSeries TTM r2.1 model that ships server-side in Vinh's production pipeline, lazy-loaded in the driver's browser via Transformers.js + WebGPU / WASM SIMD. Differentiator vs the ApexIQ local-Ollama server pattern.

Engine-agnostic byte-equality lock. D-050.

Wave-45 Phase 10 Block G close-out. The regression guarantee against engine swaps (D-050 reframed per the new feedback_byte_equality_regression_guarantee_not_killshot memory rule as the engineering safety contract behind the pitch, NOT the load-bearing positioning headline; the pitch headline is Card 4 COA-bound adaptive-controls model): two projector engines (V1 NumPy + V2 cvxpylayers) emit byte-identical violation strings on the same physical event, modulo a single ENGINE header line. Cross-ref Q&A killshot #3 + paper §3.2 + the Vinh pytest assertion shipped 2026-05-25 at commit 9048573.

ENGINE=numpy_v1
VIOLATION_LOG row=18 friction_ellipse exceeded mu_v=0.92 mu_v_max=0.85 brake_pressure_mpa=0.4 throttle_pct=12 coa_simul_permitted=true projected_pace_envelope_lower=-0.03 projected_pace_envelope_upper=0.02
VIOLATION_LOG row=19 friction_ellipse exceeded mu_v=0.94 mu_v_max=0.85 brake_pressure_mpa=0.38 throttle_pct=14 coa_simul_permitted=true projected_pace_envelope_lower=-0.04 projected_pace_envelope_upper=0.02
VIOLATION_LOG row=20 friction_ellipse exceeded mu_v=0.96 mu_v_max=0.85 brake_pressure_mpa=0.36 throttle_pct=16 coa_simul_permitted=true projected_pace_envelope_lower=-0.05 projected_pace_envelope_upper=0.01
SUMMARY violations=3 projected_pace_envelope_max=0.05 audit_id=apx-bd91a8c4

ENGINE=cvxpylayers_v2
VIOLATION_LOG row=18 friction_ellipse exceeded mu_v=0.92 mu_v_max=0.85 brake_pressure_mpa=0.4 throttle_pct=12 coa_simul_permitted=true projected_pace_envelope_lower=-0.03 projected_pace_envelope_upper=0.02
VIOLATION_LOG row=19 friction_ellipse exceeded mu_v=0.94 mu_v_max=0.85 brake_pressure_mpa=0.38 throttle_pct=14 coa_simul_permitted=true projected_pace_envelope_lower=-0.04 projected_pace_envelope_upper=0.02
VIOLATION_LOG row=20 friction_ellipse exceeded mu_v=0.96 mu_v_max=0.85 brake_pressure_mpa=0.36 throttle_pct=16 coa_simul_permitted=true projected_pace_envelope_lower=-0.05 projected_pace_envelope_upper=0.01
SUMMARY violations=3 projected_pace_envelope_max=0.05 audit_id=apx-bd91a8c4

D-031 staged ladder. Stage A + Stage B in flight.

Wave-45 Phase 9 Block F V12 + V13 close-out per D-050 + D-031. Stage A linearizes the 8-tier Pacejka physics. Stage B wraps Stage A in a 3-iterate SCP outer loop with Powell-ratio trust-region adjustment. Both engines preserve the engine-agnostic byte-equality boundary per D-050: violation strings stay byte-identical on the same physical event regardless of which projector engine fired. HEAD ships canned-fallback for both stages; Vinh wires the real runtime at the V12 + V13 swap-points.

Galaxy-tier inference and adaptation moves.

Decision-log D-019 enumerates seven shouldn't-be-possible moves (D-019 baseline of five plus wave-44 D-049 addition of IBM TSPulse polyphase anomaly detector as #6 plus wave-45 D-053 addition of Granite TTM in-browser scaffold as #7) that lift APEX from a hackathon submission into a research-paper-grade artifact. The wave-38 build shipped move #1 (WebGPU Granite Nano 350M, Layer 0; surfaced below). Wave-40 ships the visualization layer for moves #2 (Activated LoRA hot-swap, Layer 6), #3 (GEPA reflective prompt evolution, Layer 5), and #4 (EAGLE-3 speculative decoding, Layer 6 inference). Wave-44 Phase 6a adds the IBM TSPulse 1M polyphase time-frequency anomaly detector (D-016 Layer 2; per-window detection before the projector QP fires). Move #5 (Agent-as-Judge tri-agent critic loop, Layer 7) renders one section below. Each panel here renders demo fixtures from the wave-46 INTEGRATION-tier stub; live fetch wires via the wave-49 NEXT_PUBLIC_USE_REAL_* env flags (TSPulse anomaly endpoint shipped wave-49 commit 1d17eee; EAGLE-3 + aLoRA + GEPA are ACCELERATOR-tier swap-points per the per-tool honesty ladder).

Panels render demo fixtures for /judges visualization. Live fetch wires per docs/decision-log.md D-019 + D-032 + wave-49 Vinh M3-V7 ship.

WebGPU Granite Nano edge inference.

Wave-30 decision D-019 item 1 + D-021 lock Granite 4.0 Nano 350M as the in-browser edge model via Transformers.js v4 + WebGPU. The 1.5 GB WebGPU buffer pre-check (pre-mortem row 61) gates the model load; if the adapter advertises insufficient headroom, the chip below degrades to the server-only path. The 30-line Newton friction-ellipse projector runs offline alongside the small-LM head; server- authoritative reconnect (D-021) overwrites local state on every reconnect so no mechanical recommendations are emitted from the offline path. Arch-spec cross- reference at docs/architecture-spec.md Appendix W30 Layer 0 (Edge/Client Plane).

Edge inference is advisory per D-021. The canonical APEX pipeline runs server-side through Vinh-lane backend at app/backend/apex/; the edge path lowers the barrier for journalists + adaptive-driver coaches to demo the pipeline without an RTX-4060 install.

Two-person team.

Frontend, pitch, project architecture, narrative

Stephen Sookra

Computer Science, Kennesaw State University

Backend, ML pipeline, FastAPI, LangGraph + MCP + ContextForge runtime (Langflow export-graph artifact per D-017 G7 + D-054), Infrastructure

Vinh Le

Computer Science, Kennesaw State University