Aefinity AI

L24 Null Q7 · No tuning after the fact

Open challenge: the Hall of Divergence

QuestionHas anyone outside the lab reproduced these results, or broken them?

Run by us; no unaffiliated reproduction

01Why it matters

Everything on this site was run by us. The honest next step is for others to run it.

02Method, step by step

  1. Publish a table for verified reproductions and divergences.
  2. List only runs with a full log.
  3. Offer credit only, no cash or prize.

Method as recorded

A public table of verified reproductions and verified divergences for the CIS-2 spec (v0.3b). Credit only, no cash or prize. Only runs with a full log are listed. Reporters can also open a clean-room implementation PR or a spec-defect issue (CONTRIBUTING.md).

03Results, exact and sourced

  1. R1Every listed reproduction is our own run (two internal rows, 2026-09-22, own machines: Intel i5-5200U with AVX2 and Intel Celeron N4020 scalar, each 9 passed / 0 failed). No third-party reproduction is listed.

    The two rows above are internal Aefinity AI runs, not third-party reproductions

    Verbatim source cis2-spec / HALL-OF-DIVERGENCE.md

  2. R2No divergence has been reported as of this check (2026-10-12).

    *No divergence reported yet.*

    Verbatim source cis2-spec / HALL-OF-DIVERGENCE.md

04What it does NOT show

Anything about whether outsiders will succeed. Our own two rows only seed the table's format.

05Scope and caveats

  • A standing challenge, not a completed experiment.
  • No unaffiliated reproduction is recorded.
  • The two internal rows are run by us on our own machines (Debian 13, Linux 6.12.94); the source says they only seed the table's format.
  • The logs for both internal rows are not committed to the public repo; the page says a public log commit is a follow-up.
  • The quote wraps across a line break in the source (whitespace only).

06Outcome

Null

No effect found, or nothing yet to report.

07What we learned

As of 2026-10-12 no one outside the lab has reproduced or broken the CIS-2 results, and the page says so plainly. The only listed runs are our own, and their logs are not yet public.

08Check it yourself

Follow CONTRIBUTING.md to open a PR or issue. To reproduce, use the command under CIS-2 above.

  • L23
    Mixed

    Our test standard and the public experiment ledger

    We publish how we test, including what the method cannot prove, and we correct our own published claims through numbered errata.

    Q7

  • L25
    Mixed

    RANGER mock run: low-bit quantization ideas tested on synthetic tensors

    On synthetic data, one sub-hypothesis was falsified and two design constraints emerged: rotate only after splitting off the sparse outliers, and rotation helps 4-bit activations but not below 4 bits. A later test showed width and rotation overlap on outliers. These are mechanism-level findings, not claims about real models.

    Q9