imernest
Search

imernest

Software, systems, experiments, and things I learned building them.

Now

Bench
BuildingWeft — a .NET 10 physical-plan compiler under two floating-point contracts.InvestigatingWhy an A/A control returned 19.99% variance before a fix and 0.06–0.21% after, and how much of the remaining spread is still the instrument.
Latest measurementSame detector, same machine, same ggml anchor: 3.23–3.31× slower under Strict, 2.9751× faster under Inference.

Featured essay

Measurement methodology, Correctness

The instrument does not announce its own failure

Seven instances across five projects where a measurement reported success while measuring nothing.

Read the essay

14 min

Project · active

Weft

A .NET 10 compiler that lowers a semantic graph of physical plans to four backends, under two floating-point contracts.

C#

Flowchart, top to bottom. A C# fluent frontend builds a Graph through GraphBuilder, which becomes the semantic IR in Weft.Ir. Weft.Analysis computes value ranges, effects and access structure; Weft.Optimizer runs elementwise fusion, region merge and evidence-gated device placement. The result is a Physical Plan, .weftplan.json, which feeds pluggable renderers: IL, C# source, Metal, and a CUDA C artifact. Those reach a thin runtime that binds parameters and runs, a developer doing inspect and A/B, or a remote GPU built off-machine.

Lab · SEP 17 2026

A result that survives a contaminated box

Machine
RTX 5090, Vast.ai instance 51353366, driver 580.173.02
Commit
`gabilan/sam3-ggml` `ewi1963/f16-vectorized-converts` (`4f023a25`, `14a301d5`, `429236ef`)
frame, base → f16c
89.43 → 82.16 ms, −7.27 ms (−8.1%), 3/3
encode, base → f16c
53.57 → 48.22 ms, −5.35 ms (−10.0%), 3/3
prop, base → f16c
35.85 → 34.11 ms, −1.74 ms
other_cpus across rounds
3.59 → 10.20
base frame spread across rounds
89.43 → 89.27 → 89.56 (0.29 ms)
f16c frame spread across rounds
82.16 → 82.67 (0.57 ms)
Bit-identical records, base vs f16c
80/80, worst-cell IoU 1.000000, max|Δlogit| 0

Every run read gate=violated while interference varied threefold. The arms did not drift with it: frame 89.43 → 82.16 ms, encode 53.57 → 48.22 ms, 3/3 on the sign, 80/80 bit-identical.

A green check is a claim about a check, not about the thing it was supposed to check.

The instrument does not announce its own failure

Writing

All writing
  1. —
    Correctness, Compilers and code generation

    The oracle is the method

    Three unrelated systems, one shared discipline: nothing counts until something else can disagree with it.

    12 min

  2. —
    Measurement methodology, Compilers and code generation

    "Not recorded" is not zero

    Two projects, at opposite ends of the same pipeline, both decided to withhold a number rather than caption it.

    11 min

  3. —
    Measurement methodology, Audio

    What the listening caught

    Six measurement errors in one project, four of which no amount of reading would have found.

    10 min

Measurement

Kiln

9

0.1.0 → 0.6.0. Publicly verifiable.

Contradiction

SAM3 kernels

Strict contract 3.23–3.31x slower / Inference contract 2.9751x faster

Same detector, same machine, same ggml anchor. The posture is the missing denominator — see BENCHMARK_LEDGER.md weft-sam3-strict.

Contradiction

Inference gateway

Rate limit Bounds the rate / Worst-case accrual Bounds the total

A per-hour price cap does not bound spend over time. The design commits worst-case accrual at creation and persists it immediately.