Aefinity AI

L31 Negative Q8 · Speed numbers

Audit of our own removed code: benchmark lines that measured nothing

QuestionDid an earlier version of our software print performance numbers it never measured?

Run by us; no unaffiliated reproduction

01Why it matters

This is where our provenance rule came from.

02Method, step by step

  1. Review the print statements of a deleted crate preserved in git history.
  2. Check whether they take format arguments.

Method as recorded

Review of the print statements in a deleted crate (antigravity-aegis, removed 2026-07-10) whose code is preserved in git history at commit 5b69327. Run by us.

03Results, exact and sourced

  1. R1The printed benchmark lines in the removed crate were fixed string literals, not measurements.

    All are `println!` calls with **no format arguments** — string literals asserting measurements that were never taken.

    Verbatim source alice-aegis / EVIDENCE_fabricated_metrics.md

04What it does NOT show

Anything about current code; no experiment was run.

05Scope and caveats

  • An audit of our own old code; no experiment was run.
  • The quote is verbatim, but it wraps across two lines in the source file.
  • The source also says the crate did not compile (9 errors), so the printed numbers could not have been produced even in principle.
  • The document quotes 7 lines from main.rs, including the three [BENCHMARK] lines (TTFT, TPS, Peak RSS). It is a self-review with no independent reproduction.

06Outcome

Negative

The claim or approach failed, or we retracted it.

07What we learned

Every published number needs a machine, a log and a source behind it. This audit is where that rule came from.

08Check it yourself

Read the audit (link above); the commit it cites is 5b69327.