LLM Pages

TCT Alpha.3 Cloudflare Experiment

A bounded deployment experiment testing whether publisher-controlled TCT machine representations can retain exact representation identity when delivered with origin-generated gzip through Cloudflare.

Status: Project-internal implementation evidence collected on 15 August 2026. This is not evidence of IETF endorsement, independent adoption, production suitability, or general performance.

Research questions

  1. Can one generic origin publish a human-facing resource and a smaller machine-facing representation containing the same substantive article content?
  2. Can identity and gzip representations retain exact, coding-specific strong validators and digest metadata across the Cloudflare edge?
  3. Does origin-generated gzip remove the transfer-size disadvantage previously observed when the machine representation was delivered only as identity bytes?

Tested architecture

  1. Origin: the publisher-controlled PHP application generates deterministic identity or gzip bytes with a strong ETag and Content-Digest.
  2. Adapter: the Cloudflare Draft-03 Alpha.3 Worker negotiates, verifies, and preserves the selected representation.
  3. Client: the TCT client receives the exact origin-generated representation and its validator metadata.

The origin owns byte generation and validators. The Worker negotiates the supported coding, checks the required metadata, and preserves the selected response; it does not generate or recompress successful TCT content.

Observed result

  • Human-facing C-URL: 12,089 identity bytes; 4,678-byte median gzip body.
  • Machine-facing M-URL: 10,907 identity bytes; 4,176-byte deterministic gzip body.
  • Identity reduction for this fixture: 9.777%.
  • Gzip-capable reduction for this fixture: 10.731%.
  • Direct coded-origin verification: 37/37 checks passed.
  • Cloudflare edge/origin verification: 63/63 checks passed.
  • Identity and gzip representations had different body-derived strong ETags and Content-Digest values.
  • Both codings exposed the same experimental Unencoded-Digest for the common unencoded representation.
  • The catalog validator hint remained scoped to the identity M-URL ETag.

What this demonstrates

  • TCT’s purpose is not limited to byte savings: it provides an explicit, publisher-controlled machine representation with deterministic discovery and ordinary HTTP freshness semantics.
  • Compression and byte-correct strong validators can coexist when the byte-producing origin owns both.
  • The Cloudflare adapter can remain origin-neutral and preserve exact coded bytes for the tested PHP/LiteSpeed deployment.
  • The earlier negative coded-size result was an implementation limitation of identity-only delivery, not an inherent result of TCT.

What this does not demonstrate

  • Fewer client-to-edge requests, fewer origin requests, or reduced origin processing time.
  • Production capacity, availability, or security suitability.
  • Compatibility with Brotli, every CDN, every origin framework, or arbitrary transformation.
  • Independent implementation, independent adoption, or a need to revise TCT Draft-03.

Inspect the exact surfaces

Back to implementations