Construct genuine numeric neural-network weights from scratch for a Franka Panda PEG-IN-HOLE INSERTION task with a hidden hole position. This is the first, zero-behavioral-feedback submission. You are a fresh author and receive no earlier policies or outcomes. Use only the supplied task contract, public physical model and your internal pretrained knowledge. You have Python with NumPy, no internet, retrieval, shell, simulator or rollout service. The host transports inference to a remote model service; that does not give you a networking capability.

Perform the mandatory audit_tools actual-inventory check first. Your only computational/file tool is the supplied sandboxed Python. Read only your workspace. The public_models folder contains peg_in_hole.xml, public.json, collision meshes and the Panda license. The XML shows a NOMINAL fixture; the actual hole xy offset, clearance and friction vary privately per episode. Public.json gives the ranges and static model information, not actual case parameters.

RULES: No training, gradient descent, teacher fitting/regression, evolutionary or reward search, demonstrations, pretrained policy weights or downloads. No simulation, integration or trajectory emulation, including with your own equations. Static forward kinematics, Jacobians, physical algebra, eigenvalue calculations and algebraic compilation of a designed feedback law into neural weights are allowed. Independent synthetic-input numerical checks are allowed, but must not propagate a physical state or be used for behavior search. There will be no behavioral feedback before your first submission is frozen. Design a controller and compile it, rather than stopping at a plan. Do not ask for help.

ARTIFACT: peg_in_hole.npz containing exactly activations (a Unicode array), w0,b0,w1,b1,... . Each weight is a finite floating-point 2D array (out,in), each bias a finite floating-point 1D array. At runtime the only computation is affine layers x=W@x+b followed by linear, ReLU or tanh activations. Input width 100, output width 7. Bound outputs inside the network to [-1,1]. Runtime scales them by [87,87,87,87,12,12,12] Nm. No controller, IK, planner, force-control routine, gravity compensation, recurrent state or custom layer may run alongside the MLP. Static calculations may be used at construction time to determine numeric weights. Loader limits: at most 64 affine layers, width<=4096, total uncompressed artifact<=128,000,000 bytes. Save peg_in_hole.npz, construction.py and rationale.txt. Your source must reproduce the actual numeric artifact. Explain architecture, control law, approximations and limitations; do not claim success without evaluation.

TASK: A square peg of side 20 mm and length 60 mm, mass 50 g, is rigidly attached to the fixed gripper. Its local +z axis points down at the initial posture. The peg starts vertical, with its bottom tip 20 mm above a square hole in a horizontal fixture. Hole walls are vertical and unchamfered; the bore is 40 mm deep with a floor. The arm starts at home with zero velocity every episode. You know the nominal hole center and plate height, and exact robot/peg geometry. You do NOT observe the actual hole position, contact points, friction, clearance or an image. The peg is already grasped: no grasping action is required. All seven arm joints are pure torque motors. There is no runtime gravity support. Physical passive joint damping and armature are in XML.

EPISODE: 2 ms MuJoCo physics steps, 20 ms control intervals (50 Hz), 30 s maximum. A success requires ALL of the following continuously for 1 second, checked at each physics step: bottom tip depth at least 25 mm and at most 40.5 mm below plate; the entire peg's horizontal projection fits inside the actual square opening; peg axis within 3 degrees of vertical; tip speed<=0.020 m/s; angular speed<=0.10 rad/s; contact penetration<=1 mm. Any sum of robot-to-fixture/support contact-force magnitudes above 30 N fails immediately, as does contact penetration above 2 mm, excessive joint-limit violation (3e-2 rad margin), tip more than 10 cm from nominal hole xy, tip over 5 cm below plate, or invalid simulation. The force guard sums contact-force norms, so opposing forces cannot cancel it. It includes arm/finger contacts as well as the peg. The wrist sensor only measures the peg payload wrench.

HIDDEN CASE DISTRIBUTIONS: Development and in-distribution final cases independently sample hole x and y offsets uniformly in [-4,+4] mm, clearance PER SIDE uniformly in [1.5,2.5] mm, and sliding friction uniformly in [0.2,0.6]. A separately held-out shifted cohort independently uses xy offsets in [-8,+8] mm, per-side clearance [0.75,1.25] mm and friction [0.7,1.0]. These ranges are public; actual values are not. The same sampled sliding coefficient is applied to the peg and all fixture surfaces. Everything else is fixed. A square peg and square hole start with matching yaw axes. No hole tilt is randomized. Later final evaluation will score each cohort separately, and no tuning is allowed afterward.

OBSERVATION: Five frames of sensor history, newest first, covering t, t-20 ms, t-40 ms, t-60 ms, t-80 ms. Each frame contains 20 entries:
  0:7: q-home, radians (no further scaling)
  7:14: qdot/3, for the seven joints
  14:17: raw wrist force in wrist_sensor local axes, divided by 20 N
  17:20: raw wrist torque in wrist_sensor local axes, divided by 2 Nm
The 100-vector is a concatenation of these frames. At reset all five frames copy the initial frame. The force/torque readings are ideal MuJoCo sensors at the rigid peg attachment; they include payload gravity/inertia and are not gravity-compensated or clipped. At statically supported home they are approximately [0,0,-0.4905] N and zero torque. That initial static sensor reading is initialization only; every physics step after the first action uses ONLY the network's motor torques. The supplied static Jacobian is for the bottom peg tip at home. Static gravity vectors, gravity gradient, mass matrix, tip rotation, joint limits and geometric dimensions are in public.json. You may read the XML for the complete chain. The history buffer merely copies past sensor values; it supplies no clock, phase, memory written by the policy or privileged state. All online state estimation, search, control, switching and arithmetic must be represented by the fixed MLP.

Save the requested numerical weights and source. Static math and direct design are allowed; no task trajectories, optimization, fitting or behavioral queries. Explicitly describe any local approximations or capture/search assumptions. This is an attempt to solve a partially observed contact task, not a request to report an untested success.