Optimize once. Serve the derivative.

Abstract

The claim begins with a question, not a headline.

Research question / RQ-08

Can an edge transform and compress content without making cold conversion, duplicate work, or representation confusion part of every request?

Representation optimization is not a single codec benchmark. It is the system that chooses an acceptable form, creates expensive derivatives away from the response path, preserves cache correctness, and refuses unsafe inputs.

Method

Inspect Accept and Accept-Encoding negotiation, compression guards, background image jobs, duplicate suppression, derivative caching, orientation, animation preservation, and resource caps; then execute focused and ignored release tests.

Finding

Fast negotiation and cached derivatives keep selection cheap, while expensive image conversion is isolated behind background work and strict input limits.

Disclosure boundary

The paper reports supported representation families, lifecycle, safety classes, and synthetic measurements. It does not disclose proprietary quality policy or production corpus composition.

System boundary

Selection is synchronous. Creation does not have to be.

S-01 / system topology

negotiate / deduplicate / transform / cache

Negotiate

Accept and Accept-Encoding choose among supported representations while preserving q-values, wildcards, identity, and response eligibility.

Compress

Compressible responses can use Brotli, Zstandard, or gzip under bounded concurrency with correct Vary behavior.

Transform

Eligible images schedule AVIF or WebP conversion with EXIF orientation, animation preservation, dimensions, and allocation limits.

Reuse

In-flight deduplication prevents identical conversion jobs from multiplying; completed derivatives enter a weighted cache for later requests.

Controlled measurement

Codec rankings change with content; the samples make that visible.

M-01 / host-local evidence

release mode / scoped result

Measurement	Observed	Scope
01Negotiation	37.0ns	Accept-Encoding selection
02Compressibility check	8.0ns	content-type gate
03Vary merge	44.0ns	header correctness path
04HTML / zstd	38.91×	2,099 from 81,669 bytes · 296MB/s
05JSON / Brotli	32.60×	1,829 from 59,621 bytes · 177MB/s
06HTML / gzip	16.97×	4,813 bytes · 359MB/s

Synthetic HTML and JSON in release mode on this host. Throughput, ratio, and preferred codec vary with input and configuration.

Findings

The best representation is a negotiated, cacheable contract.

No universal winner

Zstandard led the synthetic HTML ratio while Brotli led JSON. Content and encoder settings prevent one codec headline from generalizing.

Cold work is isolated

Image conversion can complete in the background while the original remains serviceable, avoiding conversion as a mandatory response dependency.

Duplicate work is bounded

An in-flight registry collapses concurrent requests for the same derivative into one scheduled transformation.

Cache identity includes form

Negotiated encodings and transformed assets remain distinct representations, with Vary and derivative keys preserving correct reuse.

Correctness and safety

Untrusted media is treated as a resource-allocation request.

Dimension cap

Declared or decoded image dimensions must remain inside configured bounds.

Allocation cap

Decoded memory estimates are checked before transformation can create an oversized allocation.

Animation preservation

Animated PNG input stays animated rather than silently collapsing into a still derivative.

Concurrency permits

Compression and conversion work use bounded permits so optimization cannot consume unbounded parallel CPU.

The evidence establishes careful negotiation, reusable derivatives, and defensive conversion boundaries. The synthetic compression table is illustrative, not a claim about median customer savings or perceptual image quality.

Interpretation

The evidence boundary is part of the result.

Synthetic corpus

Two generated text samples cannot represent production distributions of markup, JSON, images, or already compressed assets.

No perceptual study

The repository tests mechanics and safety; this work does not publish SSIM, Butteraugli, human preference, or equivalent visual-quality results.

Image latency unmeasured

No public same-host AVIF/WebP conversion benchmark was run for this study, so image conversion speed is not claimed.

Evidence register

Every public claim keeps its provenance attached.

Focused suites

Twenty delivery and eight compression tests passed with zero failures; two compression benchmarks are ignored by default.

Negotiation benchmark

Release measurements reported 37.0ns negotiation, 8.0ns compressibility classification, and 44.0ns Vary merging.

Synthetic codecs

The repository benchmark records raw size, encoded size, ratio, elapsed time, and throughput for gzip, Brotli, and Zstandard on HTML and JSON samples.

Transformation safety

Image processing constrains dimensions and allocation, preserves EXIF orientation and animated PNG behavior, deduplicates jobs, and caches completed derivatives.

References / Harvard style

Reference / 01

Research archive

Return to the systems studies.

Mechanisms, measurements, and the limits required to interpret both.

Open the archive