Independent contract review
Challenge command fields, transitions, reset/error precedence, accumulator lifecycle, memory behavior, and ambiguous language.
The calibrated workload is complete. Accelerator nodes 0–217 and the explicit host tail pass tested inputs; a faster command-equivalent runner, COCO-wide accuracy, independent review, and command-delivery decision are next.
Challenge command fields, transitions, reset/error precedence, accumulator lifecycle, memory behavior, and ambiguous language.
Re-derive expected fixture outcomes, inspect provenance and masking rules, and propose missing positive, boundary, or negative cases.
Build a faster backend that matches existing fixtures and multi-input outputs exactly, then measure COCO-wide command-path accuracy. Review package size and command transport.
See the next slice →Add adversarial cases, independent comparisons, and minimized regressions for rounding, saturation, layouts, activations, and accumulation.
HASLAB is deliberately layered so a contribution can be deep and useful without spanning the complete system.
Export auditing, operator coverage, calibration, accuracy budgets, preprocessing, and host-tail boundaries.
Deterministic lowering, shape and layout logic, tiling, SRAM planning, packages, errors, and simulator execution.
Narrow conformance-driven modules, assertions, differential testing, formal properties, and reproducible tool flows.
Resource probes, memory mapping, board evaluation, process assumptions, SRAM, DFT, physical verification, and packaging.
A contribution should explain why the change exists, which contract it affects, and what evidence shows it is correct.
Start with the contribution guide and current development plan. For interface or numerical changes, expect a short architecture decision record and evidence that keeps the golden oracle independent.