Independent. No commercial relationship with ElevenLabs — no credits, no programme, no agreement — checked 5 September 2026. Disclosures · how every claim here is evidenced

elevenlabs.providers.sgit.ai / The briefs, and what was done with them

The briefs, and what was done with them

This site was commissioned by briefs. They are published here in full, next to what was actually done with each instruction — including the three findings that were wrong, the one proposed capability row that was rejected, and the parts that are blocked on somebody else.

Prose from the video vault at commit 7d1916aca5f3, 7 September 2026. The raw documents are at /briefs/, licensed CC BY 4.0 as they arrived. When the vault moves ahead, this page is behind — and says so rather than guessing.

Why publish the brief. The same discipline as the rest of this site: the claim and its evidence travel together. A site that publishes what it was asked to build can be checked against it. One that publishes only the result cannot, and asks you to take its judgement on trust — which is the thing this site keeps declining to do.

Pack one · the providers-site pack, 7 September 2026

Six documents: an audit of the live site at v0.1.1, two axes and eleven decisions, an eight-step plan, twelve acceptance tests, eight open questions. Read it rawaudit · architecture · plan · acceptance · open questions.

What shipped, at v0.2.0

StepWhat it askedWhat was done
1A capability block: what connecting this provider grants an agentDone, in two layers. The block already existed for our own key; §2 now carries what the platform can grant per product above it, including the row the pack was right about — see below
2The capability-tier axis, and its intersection with the credential patternsDone. The matrix is on the patterns page, and every tool on this site is placed in exactly one cell
3Positioning, a disclosure line at the top of every page, the per-product ruleDone. The strip above this page is the disclosure line; the positioning paragraph is on the report
6Tier three resolved either wayDone, the honest way. /pattern-three/ now names exactly what would have to ship, and carries the specification state at the top
7Composition, and llms-full.txtDone. Family links point at pages rather than domains; the provider disambiguation is on the report. llms-full.txt already shipped at v0.1.0
4, 5, 8The video walkthrough, the tried-and-abandoned page, what we use it forNot done, deliberately. Each is blocked on something real — see the ordering note below

Three findings in the audit were wrong

Checked against the live site rather than recalled, which is the standard the audit set for itself.

Corrected

llms-full.txt was not absent

It shipped at v0.1.0 and returns 200. The audit's M8 and step 7 asked for a file the site already had. Cheap to check, and the reason the ledger exists.

Corrected

The capability block existed

M1 called it “the largest gap” and said nothing on the site did it. §2 has carried verb × object class × reach, with reversibility marked and emitted as front-matter, since v0.1.0. The real gap was narrower and better: the block described what our scoped key grants, not what the platform can grant. That is now two layers.

Rejected

send × audio × world, irreversible

The pack proposed a row saying that connecting a voice provider grants an agent the capability to send audio into the world, irreversibly. It is the same over-generalisation as the signed-URL claim this site already had to correct, from the same source brief.

The text-to-speech endpoint returns audio to the caller. Nothing leaves for the world; the estate's own primitive distinguishes endpoint reach from world reach precisely here. This vendor's genuinely irreversible row is Agents, which §2 names and which our key is scoped out of. Shipping the proposed row would tell a reader that generating narration publishes it — the error §4 exists to prevent, one product over.

And one finding it could not have made

While checking the audit's claim that §4 was built correctly, the citation in §4 turned out to be invisible: <https://…> is a GFM autolink, this site's renderer had no autolink rule, and the angle brackets reached the browser as an unknown tag — so the URL a reader is told to check was dropped from the page entirely. Fixed in v0.1.2, with a build check so it cannot return.

A page-level audit cannot catch a URL that is not in the DOM. That is an argument for the pack's own acceptance tests being machine-checked rather than read, which is where the rest of this site's rules already live.

The ordering problem in the plan

Three steps are blocked, and two of them are blocked by the plan's own gates:

Pack two · the video-pipeline brief, 7 September 2026

A working, measured video pipeline — 724 KB of tooling, four reels made with it, every number observed — handed over so this site can make narrated videos with the provider it reports on. Read it rawthe brief · the first-run guide.

What was done, in two stages. It was first published as a plan with its numbers badged; then, on 8 September 2026, the project lead supplied a key and it was run — the first execution of that path by anyone. What it found is on /video/: two refusals, one prediction wrong, one failure nobody predicted, and the measurement that closes the drift question verified 8 Sep 2026. The video and its materials are in an encrypted vault rather than in this repository.

The three reasons it was not run before that key arrived still stand, and the first one is why the exception is dated rather than quiet:

  1. No key, and that is a rule rather than an inconvenience. The site is built with no key present, and every verified chip came from a human's own browser rather than from this repository. Asking for a key would have traded that property for a two-minute film; being given one, for one run, with the exception recorded and dated, does not — and the key was never written to a file or committed, which the required secret scan enforces.
  2. The toolchain is not this site's to hold. The pack above records the estate's decision that component code stays canonical where it lives and these sites hold the context; the same pack's scope boundary says the video toolchain is out of scope. Vendoring 724 KB of a render pipeline into a report site would contradict both, and the brief's own value — the measured numbers, the failure points, the script rules — transfers without it.
  3. The interesting half needs a person, not an agent. The brief's own last section is what I could not tell you: whether the shim works, which names the model mispronounces, how the API behaves under the render's concurrency, whether the drift figure changes. Those are answered by running it, once, with a key — and then they are verified rather than written, not run.

What the brief changed on this site anyway

The rule both packs share, and this site keeps

No claim without a state. Everything added by either pack obeys the ledger contract, including claims about work that does not exist: a specification takes specified, not shipped, code written and never executed takes written, not run, and a projection carries its arithmetic. The join is done at build time, so a claim that appears on a page and not in the table fails the build rather than the review.