Build and verification record

The check ships with the claim.

I build heavily with AI. My role is architecture, decomposition, test design, failure boundaries, and deciding what evidence is sufficient to trust a result. This page records those checks instead of asking you to infer technical depth from finished-looking output.

Verified Reproduced by an independent check or executable test path.

Measured Measured in the referenced project/corpus.

Caveat Known limitation stated before the headline.

Boundary A scope boundary that prevents a stronger claim.

20 / 20FirstLight heavy ↔ pure routing parity

~88.4%FirstLight candidate-pool variance retained by 6 axes, bundled corpus

30% / 45%elm-compress smaller than LZMA / zlib on the mixed 9-fixture corpus

FirstLight — corpus-derived dimensional reader + advisory router

FirstLight has two deliberately separate mechanisms: a dimensional reader and an advisory router. The dimensional build constructs 13 candidate axes from semantic (TF-IDF → SVD), structural, and cognitive-move arms. An exhaustive span-preservation objective selects six axes; live prompts are projected through the same fitted transforms. Routing remains separate so an interesting coordinate does not automatically dictate the selected tool.

Measured In the reviewed bundled corpus, the 6-of-13 max_combined subset retained about 0.8839 of candidate-pool variance. That is corpus-specific variance retention—not model accuracy.

Verified 8 passed / 0 failed / 0 skipped in the reviewed suite, including 20/20 heavy-vs-pure routing parity, a read path that succeeds with NumPy/scikit-learn imports blocked, cache write/read checks, deliverable-gate behavior, and fail-open hooks.

git clone https://github.com/charliedorsey/firstlight
cd firstlight
./run_tests.sh

Boundary The dimensions describe the request; they do not choose the route. The system is advisory and fails open.

elm-compress — compiler/codegen + lossless bakeoff

A tiny line-oriented transform language emits paired ordinary Rust forward/inverse kernels. build.rs invokes the compiler, refreshes committed fallback source, and CI diffs regenerated output against the committed code so codegen drift fails the build. Runtime selection keeps STORE, zlib, LZMA, adaptive delta, byteplane, and XOR-delta candidates in competition.

Fixture	Raw	zlib-9	LZMA	elm
gps_i32	64,000	33,813	18,252	1,638
stride4	50,000	2,118	1,328	87
sawtooth	50,000	519	364	227
audio_i16	40,000	38,427	30,664	16,370
tensor_f32	40,000	12,733	8,696	8,953
text	41,209	6,946	6,472	6,441
smallalpha	40,000	12,306	11,564	11,481
zeros	50,000	71	136	76
random	30,000	30,016	30,060	30,005
Corpus	405,209	136,949	107,536	75,278

Measured On this mixed nine-fixture corpus, the final output is about 45% smaller than zlib and 30% smaller than LZMA. The table intentionally leaves losses visible.

git clone https://github.com/charliedorsey/elm-compress
cd elm-compress/elm-compress
cargo test --release

Caveat The isolated compiled adaptive-delta kernel has measured around ~860–867 MB/s in the project benchmark path. The full bakeoff is roughly ~1 MB/s because it executes multiple candidate codecs—including losing ones—before choosing the smallest output. “The compressor runs at 860 MB/s” would be false.

Residual Lens — eye in the loop

A one-file browser instrument maps up to six scalar dimensions into three opponent channels: luminance, red/green, and blue/yellow. It uses null-aware bilinear interpolation that ignores unsupported neighbors and re-normalizes over supported values rather than treating null as numeric zero.

Verified The reviewed artifact is self-contained: no build step, no network calls, no external runtime dependency, bundled datasets plus JSON load/export.

Boundary The visual system is a hypothesis generator. The human eye gets proposal rights, not promotion rights. A pattern must become an exact measurable predicate and survive discrimination/derivation before it becomes a claim.

Cow / Ironclad — C/WASM experimental substrate

A from-scratch classical chess engine in C compiled to WebAssembly, with perft-correct move generation, toggleable ordering/mechanism cells, ablation and match harnesses, and structured per-decision telemetry. The wider mining/promotion loop is not yet fully closed.

Inspect the repository →

OXBOW — integrity, witness, falsification

The reviewed OXBOW wrapper represented 239 files; 238 were manifest-checked with 0 missing and 0 mismatched. Core, falsification, gauge, bundle, and witness self-tests passed in the reviewed environment. Witness tests include append-only/rebuild parity, mandatory overhang, self-witness caveat, and source/inference grammar checks.

Caveat The extracted vendored FirstLight run_tests.sh lacked executable permission, so one top-level shell path hit a packaging seam. Invoking it explicitly with bash produced the passing FirstLight suite. That is packaging friction, not a substantive test failure.

Boundary Bundle/witness validation establishes integrity and lawful form, not truth of source claims.

What is deliberately not claimed

That FirstLight’s ~88.4% number is accuracy.
That FirstLight’s dimensional coordinates choose the route.
That elm-compress as a full system runs at ~860 MB/s.
That custom compression mechanisms must beat ordinary LZMA.
That a validated governed bundle proves the truth of the claims it contains.
That AI-paired solo building is equivalent to years of conventional production software engineering.