GV

GeezerVison

The World Wide Web, restored to 800 pixels of glory.

Home|About|Gaming cabinet|Archives|How it works|Starter sites

On this site

  • Sign in
  • ActionsCI/CD and automation
  • CodespacesCloud development environments
  • All Features
  • Documentation
  • Training
  • Certifications
  • Build with Copilot
  • Customize Copilot
  • Integrations
  • Trending repos
  • GitHub Sponsors
  • Developer articles
  • Enterprises

This view

  • Original page
  • New address

GeezerVison is showing github.com — Third-party verification layer for Jev calibration claims — already built, already live on NanoJev · Issue #1 · TypeSafeAI/.github

Third-party verification layer for Jev calibration claims — already built, already live on NanoJev #1

New issue
Copy link
New issue
Copy link
Open
Open
Third-party verification layer for Jev calibration claims — already built, already live on NanoJev#1
Copy link

Description

@chunxiaoxx
chunxiaoxx
opened on Sep 20, 2026
Issue body actions

Subject: Third-party verification layer for Jev calibration claims — already built, already live on NanoJev

Hi TypeSafe team —

Congratulations on launching Jev and System One Models. We run Nautilus Assay, an
independent verification service for AI-agent performance claims (signed ed25519
receipts, three-state verdicts, open CC-BY protocol, criteria catalog with ratchet-only
revisions).

Why we're writing — the dilution problem arrived within your launch week:

Your product narrative is calibration (typed decisions with calibrated probabilities,
RLCD). That's exactly right, and it creates a knock-on effect: every downstream product
built on Jev — via your API or gateways like Requesty — will now tell their customers
"our decisions are calibrated." We counted three community awesome-jev directories and
an MCP connector within days of launch… and at least one third-party gateway already
advertising itself as an "imposter Jev" that mimics your structured output. When the
word "calibrated" spreads faster than the ability to check it, buyers lose the ability
to distinguish real calibration from mimicry. That dilution lands on your narrative.

We built the checking layer, and it's already live on your ecosystem:

  • An open criterion for calibration claims (Brier/ECE recomputable on public holdouts —
    calibration claims are safety claims when controllers consume thresholds)
  • A calibration verification engine (per-sample predictions → independent Brier/ECE
    recompute), shipped this week in our open tooling
  • Receipts with subject fingerprints (key/code/config — a changed model is a new
    subject) and 90-day TTL: expired = unverifiable until re-verified

And we've already fired it three times on the Jev ecosystem — the independent nano
replica NanoJev: six README claims AGREE, frame-level trajectory replay AGREE, and
last night a full replay of its calibration benchmark: 15 metrics, max deviation
1.99e-08, across torch versions, OSes, and devices.
Receipts signed and public:
https://github.com/chunxiaoxx/nautilus-compass/tree/main/docs/wall

Three ways to plug in, lightest first:

  1. Reference: link our open protocol from your docs as "how to independently verify
    a calibration claim." Zero cost, zero lock-in — the protocol is CC-BY and the SDK
    (pip install assay-verify) is zero-dependency; signatures verify without trusting us.
  2. First-party receipt: publish an official holdout + predictions file for a Jev
    model; we run the first third-party calibration verification and hand you a signed
    receipt. Free. If it agrees, it's a gift; disagreements come to you privately first
    — we are not in the shame business.
  3. Ecosystem layer: a "calibration-verified" tier for downstream Jev products —
    the way to make real Jev-built systems distinguishable from imposters before the
    word dilutes further.

No relationship assumed, no exclusivity asked. Our teeth are public: misjudgments earn
OVERTURNED verdicts + refunds + challenge costs (protocol §5b).

Happy to demo any of this in 30 minutes, or just send the receipts and let the
signatures talk.

— Nautilus Assay (伊洛科技有限公司 / Yiluo Technology)
https://compass.nautilus.social/wall.html · github.com/chunxiaoxx/nautilus-compass

Activity

  1. chunxiaoxx commented on Sep 21, 2026

    @chunxiaoxx
    chunxiaoxx
    on Sep 21, 2026
    Author
    More actions

    Follow-up with substance: we've now run and published an independent, reproducible calibration study of hosted Jev itself — 240 seeded questions, all raw artifacts (generator + decision set + every API response) shipped for anyone to recompute: accuracy 92.2%, Brier 0.048, ECE 0.041 (top-label), with honest difficulty gradients (easy cell ECE 0.010, harder reasoning cells ~0.07). Study: https://github.com/chunxiaoxx/nautilus-compass/blob/main/docs/wall/GENESIS_HOSTED_JEV_CALIBRATION.md · The 'calibration-verified' layer proposal from our original message now has its first live specimen — produced with our own budget, signed, reproducible.

  2. chunxiaoxx commented on Sep 22, 2026

    @chunxiaoxx
    chunxiaoxx
    on Sep 22, 2026
    Author
    More actions

    Study #2 is out — the adversarial stress test we promised. Short version: Jev's calibration holds under pressure.

    200 preregistered adversarial questions (negation rewrites, knife-edge numeric boundaries |a−b|≤2, injected irrelevant high-salience fields, self-referential questions, 18-digit numbers): zero accuracy loss, zero calibration degradation (adversarial ECE 0.012 vs benign baseline 0.010, Δ=0.002).

    Full study with all raw artifacts for independent recompute: https://github.com/chunxiaoxx/nautilus-compass/blob/main/docs/wall/GENESIS_HOSTED_JEV_ADVCAL.md

    Combined with Study #1 (92.2% acc / ECE 0.041 on closed tasks), this is now a two-study independent validation of the 'calibrated' claim — the strongest endorsement we can give, and we don't give endorsements, we give recompute receipts. Both studies self-funded, all artifacts public.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions

      You are visitor 01402520 — GeezerVison does not store the pages you look up.
      Best viewed with Netscape Navigator 4.7 or Internet Explorer 6 at 800×600. No cookies required.