# ATOMiK Proof Cards This digest is for controlled proof review. It summarizes what each artifact proves, what it does not prove, and what language is safe to use. ## Proof Card: ATOMiK Desk v0.40-A UI Artifact Evidence label: HARDWARE_VALIDATED Artifact: `10-current-live-atomik-desk-v040a.png` (supersedes v0.39-K) What it proves: - The current ATOMiK Desk prototype/demo surface was captured live from `/dev/fb0` (fb2png) on the booted AX7020. - The UI direction is visually inspectable, and on-screen metrics are driven by real measured on-board data. What it does not prove: - Customer workload performance. - Production maturity or uptime. - Battery, heat, cooling, water, power-bill, or footprint outcomes. - Benchmark results. Safe claim: ATOMiK Desk v0.40-A is the current Zynq hardware UI artifact, captured live from the framebuffer. Caveat: Use with `10-current-live-atomik-desk-v040a.caption.md` and the label `HARDWARE_VALIDATED UI ARTIFACT - NOT A CUSTOMER WORKLOAD BENCHMARK`. ## Proof Card: Parallel-Bank Throughput on AX7020 Evidence label: LIVE_MEASURED Artifact: `PARALLEL_BANKS_HARDWARE_VALIDATED.md` (repo: `hardware/zynq/results/`) What it proves: - An 8-bank ATOMiK XOR accumulator scaled linearly with allocated banks on the AX7020, measured with an on-chip cycle counter. - 1/2/4/8 banks complete the same accumulation in 1/2/4/8x fewer cycles (about 100 to 800 Mdeltas/s at the 100 MHz sys clock). - The result was byte-identical across every bank count and matched a software recompute: the order-independent, lock-free shared accumulator, on silicon. What it does not prove: - A customer-workload speedup. - Battery, heat, cooling, water, power-bill, or footprint outcomes. - Production readiness. Safe claim: On the AX7020, an 8-bank ATOMiK accumulator scaled 1/2/4/8x with allocated banks, with a byte-identical result that matched software. Caveat: This is the measurement bench engine (0xF0021000), not the production application adapter. Quote throughput figures at the 100 MHz sys clock. ## Proof Card: Live Workloads Surface on AX7020 Evidence label: LIVE_MEASURED Artifact: `12-workloads-live-ax7020.png` (pairs with `PARALLEL_BANKS_HARDWARE_VALIDATED.md`) What it proves: - The Workloads surface (real-time telemetry aggregation) runs on the AX7020 over HDMI, captured live from `/dev/fb0`. - The on-screen throughput is live measured from the parallel-bank engine: 800 Mevents/s at 8 banks, 1/2/4/8 throughput scaling, byte-identical result across all bank configurations. - The first customer-facing workload demo proven end-to-end on hardware. What it does not prove: - Customer workload performance, production maturity, or downstream power/thermal outcomes. - An interactive session or USB keyboard/mouse input (both still in flight; this is a static capture). - That any on-screen value other than the throughput figures is measured (Resource Fabric lane values and top-bar temp/efficiency/predictive readouts are derived or scenario data). Safe claim: ATOMiK's Workloads surface runs on the AX7020 over HDMI showing live measured parallel-bank throughput (800 Mevents/s at 8 banks, byte-identical across configs). Caveat: Use with `12-workloads-live-ax7020.caption.md`. Scope the numbers to throughput only. Never present as an interactive demo or imply USB input. ## Proof Card: Verified Workload Savings + Self-Driving Demo (AX7020) Artifact: `13-workloads-telemetry-verified-ax7020.png` + `14-workloads-coalescing-selfdriving-ax7020.png` (captions alongside) What it proves: - The redesigned conventional-vs-ATOMiK Workloads surface runs on the AX7020 over HDMI, captured live from `/dev/fb0`. - Edge telemetry sync: 2,048 B conventional vs 160 B as deltas per tick — 92% less data moved — adapter-verified (exact 64-bit LOAD/ACCUM/READ round-trips at 0xF0020000, falsification-tested harness). - Control coalescing: 256 updates coalesce to 32 net writes — 87% fewer — adapter-verified, final state identical. - The demo SELF-DRIVES: the untethered board auto-cycles all three scenarios (two captures ~15s apart show different scenarios). What it does NOT prove: - Universal multipliers (92% is at 5% change density; 87% at this update locality — ratios scale with the pattern). Never blend with the retired 7,670x–916,000x range. - An interactive session or USB keyboard/mouse input (both still in flight). Self-driving = auto-cycling, not input. - Power, thermal, customer-outcome, or production-maturity claims. Fabric/top-bar values remain derived/scenario. Use with `13-...caption.md` / `14-...caption.md`. Scope every number to its pattern. ## Proof Card: Zynq HDMI Display (1080p30) Evidence label: HARDWARE_VALIDATED Artifact: `HDMI_1080P60_IMPOSSIBLE.md` (repo: `hardware/zynq/results/`) What it proves: - The ATOMiK Desk UI renders at 1080p30 over HDMI on the AX7020. - A fresh place-and-route shows 1080p60 fails minimum-pulse-width on the 737.5 MHz serializer. What it does not prove: - 1080p60 on this board (it is physically impossible on the soldered -2 part). Safe claim: ATOMiK renders at 1080p30 over HDMI on the AX7020. Caveat: Claim 1080p30 only. Never claim 1080p60. 720p60 is the achievable true-60Hz path. ## Proof Card: Linux Userspace-to-FPGA Validation Evidence label: HARDWARE_VALIDATED Artifacts: `LINUX_USERSPACE_PROOF.md`, `LINUX_USERSPACE_PROOF_SUMMARY.md`, `ZYNQ_BASELINE.md` What it proves: - ATOMiK primitive path was exercised through Linux userspace to FPGA. - 16/16 algebraic property checks passed on the documented Zynq configuration. - MMIO ordering requirement was discovered and documented. What it does not prove: - Customer workload value. - Battery, thermal, cooling, water, footprint, or power-bill outcomes. - Production readiness. - Universal speedup. Safe claim: ATOMiK algebraic property checks passed through a Linux userspace-to-FPGA path on the documented Zynq configuration. Caveat: Use with environment details and the `fence iorw, iorw` plus dummy STATUS read ordering requirement. ## Proof Card: Zynq Frozen Baseline Evidence label: HARDWARE_VALIDATED / baseline context Artifact: `ZYNQ_BASELINE.md` What it proves: - The zynq-linux-v1 baseline records board, CPU, Linux, OpenSBI, rootfs, toolchain, boot configuration, checksums, and pass result. - The documented baseline is the frozen Linux/userspace hardware-path validation baseline. What it does not prove: - v0.39-K UI screenshot metrics. - v0.33-D AX7020 performance matrix results. - Future SD-boot workload validation. Safe claim: zynq-linux-v1 is the frozen Linux userspace-to-FPGA validation baseline for the 16/16 algebraic checks. Caveat: Do not blend this baseline with later UI or SD-boot bring-up artifacts. ## Proof Card: AX7020 Raw Matrix Evidence label: LIVE_MEASURED Artifacts: `perf_matrix_ax7020_20260509.txt`, `perf_matrix_ax7020_20260509.csv`, `perf_matrix_ax7020_20260509.summary.json` What it proves: - A board-run four-way matrix exists for software, direct hardware, batched hardware, and profiled/coalesced paths. - The matrix includes workload-specific wins and losses. What it does not prove: - Universal speedup. - Customer workload results. - Battery, heat, cooling, water, power-bill, footprint, or production outcomes. Safe claim: The AX7020 matrix shows workload-specific behavior across software, direct, batched, and profiled paths. Caveat: Do not isolate the biggest number without the row, workload, baseline, and caveat. ## Proof Card: AX7020 Interpretation Evidence label: LIVE_MEASURED interpretation Artifacts: `20260509_matrix_interpretation.md`, `AX7020_SUMMARY.md` What it proves: - The defensible story is workload-specific: ATOMiK can win when batching/coalescing/personality rules reduce redundant work. - Naive hardware access can lose. - Larger or non-fit rows can lose to software. What it does not prove: - That ATOMiK is always faster. - That hardware alone is the value. - That downstream customer outcomes are measured. Safe claim: ATOMiK wins when the workload lets the architecture compound; naive hardware access and non-fit workloads can lose. Caveat: Use with the raw matrix and the known limitations around SYNC repeat behavior, AGENT small-workload overhead, and first-run mmap overhead. ## Proof Card: Claims Registry Snapshot Evidence label: Claims registry Artifact: `claims_registry_snapshot.yaml` What it proves: - Current public-safe claims have labels, artifacts, caveats, and notes. - v0.39-K, AX7020, Linux userspace validation, formal proof work, SD boot build artifacts, and concept visuals are separated by label. What it does not prove: - The technical claim itself without the referenced artifact. - That external or future artifacts are included in this folder. Safe claim: The packaged claims registry maps each public-safe claim to an evidence label, artifact, and caveat. Caveat: Update the registry before adding any new Zynq workload claim. ## Proof Card: Evidence Labels Evidence label: Evidence framework Artifact: `evidence-labels.md` What it proves: - The packet has a governance system for separating live measured, hardware validated, software validated, synthesis validated, build artifact, formal proof, projected, conceptual, and roadmap claims. What it does not prove: - That any claim is true without its artifact. Safe claim: ATOMiK uses evidence labels to keep proof, projected outcomes, concept visuals, and roadmap work separate. Caveat: Labels must travel with the claim in decks, website copy, and follow-up materials.