docs/VALIDATION_GUIDE.md

# Evidence and validation guide

This guide separates original research evidence, the checks performed during packaging, and the work needed for a physical prototype. Repeating frozen cases does not create a new independent benchmark population.

## Evidence levels

| Level | Supported statement | Required interpretation |
|---|---|---|
| Mathematical contract | Atomwise parity factorization preserves the supplied integer operator | Only admitted shapes, scales, epochs and representable coefficient signs; character-family minimality |
| Encoded software | Packed frames and executable arithmetic match dense references | Does not establish electrical hardware, timing or throughput |
| Native software | Independently compiled C/C++ checks match the reference | FORK state-machine reference is not SystemVerilog simulation |
| Trained-model experiment | Frozen format-trained digit model runs; sharing matches dense integer output | Small architecture-trained task; no modern-LLM coverage evidence |
| Scenario calculations | Costs, analog error, KV payload and roofline quantities are explicit | Assumptions and numerical calculations, not measurements |
| Hardware | RTL and integration specifications supplied | No completed HDL/synthesis/FPGA/ASIC validation in the source |

## Original result records

- `research/results/python_tests.xml`: 200 passing Python tests. The test source remains under `research/tests/`.
- `research/results/native_reference.json`: 18 exact FORK cases from six output sizes and three backpressure settings, plus malformed-frame rejection.
- `research/results/d4_reference.json`: six exact generic INT4/INT8 native-reference cases; cycle counts are an arithmetic schedule formula.
- `research/results/c_loader_test.txt` and `c_driver_test.txt`: compiled C smoke checks, covering pre-write frame rejection and held-output consumption/sign extension.
- `research/results/clean_reproduction_report.json`: nine successful software/check stages, with six generated experiment CSVs matching the original bytes.
- `research/results/frozen_digits_replay.json`: all 270 frozen held-out integer vectors match, with 263 correct classifications and verified frame/data hashes.
- `research/results/rtl_status.json`: explicitly records zero HDL/synthesis runs, no bitstream or layout, no integrated ECC, and no D4 bridge integration.

These records are preserved byte-for-byte; paths within them may refer to their original generation environment. Do not edit old logs to make them appear to have run on your machine.

## Packaging-specific checks

`docs/PREPARATION_REPORT.json` records current checks, including source archive equality, nested hash verification, result schemas/card configuration paths, documentation links, uploader offline checks and fresh software replications where executed. It records skipped checks explicitly. Preparation is complete when the files and checks are consistent; this does not make incomplete hardware stages complete.

Two successful original compiler logs are intentionally zero bytes because compilation produced no text. They remain valid original artifacts. An empty compiler log alone is not a test-pass certificate; use its stage return code and subsequent executable result.

## Verify without running experiments

From the top-level directory:

```bash
python scripts/verify_package.py
python research/verify_release.py
```

The first command authenticates the outer payload against its own manifest and checksums. The second checks the preserved source research. Hashes prove integrity relative to the manifests; they are not author signatures. Extra local caches, environments or upload receipts are not authenticated research artifacts.

## Reproduce without modifying frozen evidence

Install the research package in an isolated environment and run its clean-copy launcher:

```bash
cd research
python -m venv .venv
# Activate .venv for your platform.
python -m pip install -e ".[research]"
python reproduce.py --output ../../SAELQIR_reproduction_01
```

GCC/G++ are required for native references and C smoke checks. A new output directory is required. `--python-only` skips the native stages and labels them as skipped. Optional `--include-training` needs the additional `.[probe]` dependencies and reruns the original architecture-training recipe. Training on the released protocol is replication on a known split, not a new holdout study.

The original source manifest copied into a reproduction directory describes the input snapshot. Generated figures, records or environment strings can differ and are not retroactively authenticated by that input manifest. Preserve the new `reproduction_report.json` and all failed-stage logs.

## Compare correctly

The equality claim concerns shared versus pre-shared **integer** weights and outputs. It does not establish equivalence with unquantized original neural weights. First-layer frame savings must not be reported as whole-model savings. The 1.4545 ratio concerns counted serial work; configuration, host transfer, stalls and achieved clock determine device latency.

The small digit task does not prove generic model-quality superiority: INT4 has one more correct label, whereas the atlas has better NLL. Report both. Analog perturbation and manufacturing-volume tables must retain their simulation/assumption labels. The MHD numerical identity checks are a scientific workload and do not validate SRAM physics.

## What a future physical release must add

At minimum: actual RTL simulation results, synthesis/resource reports, timing closure, board and bitstream identity, measured complete timing and energy, failure cases, useful-model coverage, matched quality, and a strong native comparison. D4 needs separate integration/verification; ECC and platform MMIO/ordering cannot be inferred from the current software references.

Use `docs/BENCHMARK_PLAN.md` to define those boundaries before measuring. Do not attach future hardware results to the original source version without versioning the changed implementation and preserving the old records.