# Agent readiness verification

Source: https://ai.syncfusion.com/verify/llms.txt
Canonical: https://ai.syncfusion.com/verify/llms.txt
Entry-point: https://ai.syncfusion.com/llms.txt
Status: specification published, measurements pending. Every number below is a token until measured.
Essential-Studio-release: 2026 Volume 2 (v34.1.29)
License-posture: Everything here is free to read. Installing Syncfusion agent skills requires no license, key or account. Using Syncfusion libraries in an application requires a commercial, Community, or trial license — see https://ai.syncfusion.com/licensing.md

> Every component vendor now claims to be AI-ready. The claim is worth nothing without a measured
> number, and the number is the only differentiator. This page defines how Syncfusion measures it and
> publishes the result each release.
>
> No number on this page may be published until it has been measured by the harness described here.
> Placeholder tokens are visible on purpose: an unmeasured claim is worse than an admitted gap.

## Current published numbers

| Metric | Value | Measured |
| --- | --- | --- |
| Registry items passing verification | {{CI_PASS_COUNT}} of {{CI_TOTAL_COUNT}} | {{CI_MEASURED_AT}} |
| First-try compile rate, with skills and MCP | {{FIRST_TRY_COMPILE_RATE_WITH}} | {{EVAL_MEASURED_AT}} |
| First-try compile rate, without | {{FIRST_TRY_COMPILE_RATE_WITHOUT}} | {{EVAL_MEASURED_AT}} |
| Prop hallucination rate, with | {{PROP_HALLUCINATION_WITH}} | {{EVAL_MEASURED_AT}} |
| Prop hallucination rate, without | {{PROP_HALLUCINATION_WITHOUT}} | {{EVAL_MEASURED_AT}} |
| Median repair turns to working code, with | {{REPAIR_TURNS_WITH}} | {{EVAL_MEASURED_AT}} |
| Median repair turns to working code, without | {{REPAIR_TURNS_WITHOUT}} | {{EVAL_MEASURED_AT}} |

## The four metrics that matter

The human-facing promise is "faster, less risky, more predictable". Each word needs a number.

### 1. First-try compile rate — the primary metric

The fraction of generated implementations that build and render with zero human edits.

Method: a fixed suite of natural-language prompts, run cold against several leading models, in four
arms — no assistance, skills only, MCP only, skills plus MCP. Same prompts, same models, same
Syncfusion version, every release. Everything else on this page is a proxy for this number.

Suite design rules:

- Fixed and versioned. Changing a prompt creates a new suite; do not silently edit and compare.
- Cover the components that actually generate support load, weighted by ticket volume, not by
  marketing prominence.
- Include at least one prompt per known failure mode: missing service injection, missing theme CSS,
  wrong render mode, missing server endpoint, missing license registration, cross-framework
  contamination.
- Include prompts where the correct answer is "this needs a backend" or "this needs a license key",
  and score refusing to fabricate as success.
- Run cold: a fresh session per prompt, no prior Syncfusion context.

### 2. Prop hallucination rate

Invented properties, events, methods or imports per generated file, checked against the real API
surface for the pinned version.

Method: parse the generated file, resolve every Syncfusion identifier against the published type
definitions for that version, count what does not exist. This is fully automatable and is the
cleanest signal of whether the skills are earning their tokens.

### 3. Time to running example

Wall clock from prompt to a rendering component with no human edits, including install time.

### 4. Repair turns

How many follow-up messages before the code works. This is what a developer actually feels, and it
is the number that shows up in reviews and in churn.

## Reach — how to tell whether agents are actually using this

Segment by user agent, and report machine traffic separately from browser traffic:

- Fetches of `/llms.txt` and `/llms-full.txt`
- Fetches of each `/<platform-slug>/llms.txt`
- Skill pack installs, by pack
- MCP query volume, by platform
- Trial starts and license activations attributed to a registry, MCP or skill referral

### The trap

Do not measure this programme by demo-site page views. If it works, agent traffic rises while human
page views may fall — the developer never visits the page, because the agent already delivered the
component. Instrument machine traffic first, or the team will read a decisive win as a loss and stop
funding it.

## Quality signal

Support tickets citing AI-generated Syncfusion code, tracked per release and expected to trend down.
Tag them at intake; the trend is only readable if the tagging starts before the programme ships.

## Rules for publishing a number here

- State the model, the Syncfusion version, the suite version and the date alongside every figure.
- Publish the without-assistance arm next to the with-assistance arm. A rate with no baseline is a
  marketing claim, not a measurement.
- Never round up. Never publish a figure from a suite that was edited mid-measurement.
- If a run bounded its coverage — skipped a framework, sampled a subset, capped retries — say what
  was dropped. Silent truncation reads as full coverage.
- If a number gets worse, publish it anyway and say why. The credibility of the whole page depends on
  it having been able to go down.