is-gpt-nerfed

English · 简体中文

You pick a model in Codex. This tells you whether that model is actually the one answering. If it isn't, you get this:

What it checks

Everything runs on your Mac. Nothing is uploaded.

Codex records, for every turn, which model and reasoning effort it asked for. The plugin reads those records after each turn and flags what changed without you changing it: a model swap, a lower reasoning effort, a hidden internal model such as gpt-reserve, a smaller context window. A move to a newer or larger model is reported as well, since a rollout can go either way. This costs no tokens.

On a schedule, the plugin also forks your session three times, ephemerally, through the same app-server call the desktop uses for a side chat, with the session's own model and reasoning effort. Each fork is asked for about 300 "random" numbers. A language model picks random numbers with a bias that is characteristic of the model. ModelTrace's calibrated bank turns the three answers into a fingerprint (100 % accuracy with three answers in cross-validation), and the verdict compares that fingerprint with the model you selected:

verdict	meaning
Match	the model you selected answered
Suspicious	the fingerprint leans elsewhere, but not confidently; it stands until the next probe
Downgrade / Upgrade / Rerouted	a confident mismatch: top candidate at 80 % or more, your model at 20 % or less, at least two answers
Downgraded	Codex's own records show a silent switch; no fingerprint needed
Upgraded	Codex's own records show a move to a newer or larger model
Unlisted	your model is not in the fingerprint bank yet
Invalid	no usable answer (tool use, refusal, network); not a verdict, the row keeps its last one and offers Retry

A mismatch reaches you as a macOS notification, a message in the session, and a red face in the menu bar. A match is not announced.

Install

macOS 15 or newer. One line in Terminal installs the latest release and opens it:

curl -fsSL https://raw.githubusercontent.com/kiyoakii/is-gpt-nerfed/main/install-app.sh | sh

You can also download the disk image from Releases and drag IsGPTNerfed to Applications. The app is not notarized yet, so macOS blocks that first launch until you allow it under System Settings → Privacy & Security.

Click the face in the menu bar and press Install. That registers the bundled plugin with Codex and trusts its hooks. When a newer release is out, the footer says so and a notification arrives once; click it and the app downloads the new build, checks its checksum, replaces itself and relaunches.

Without the app, on any macOS:

git clone https://github.com/kiyoakii/is-gpt-nerfed ~/is-gpt-nerfed && cd ~/is-gpt-nerfed && ./install.sh

Say yes when it asks to trust the hooks. Codex runs no hook you have not trusted, and it does not tell you.

Requires a Codex with plugin hooks (desktop app or CLI; tested on 0.154) and the system python3. ./uninstall.sh removes everything.

Using it

The macOS interface supports English and Simplified Chinese, following the system's preferred languages with English as the fallback. This includes app-generated evidence, progress messages and diagnostics shown by the app. User session titles, model IDs, reasoning-effort values and unrecognized external diagnostics keep their original text. Stored records, CLI output, notifications and ModelTrace's calibrated probe prompts are unchanged.

Every session you work in is probed in the background after 30 minutes of activity: the hooks run the schedule after each turn, and the menu bar app catches a session that goes quiet right after its due time. To probe a session right away, say $is-gpt-nerfed in it.

The menu bar panel lists the sessions of the last 48 hours with their last verdict. Click a session for its report (fingerprint, earlier probes, evidence) and right-click for actions. Each session has Probe and Retry, and the Fresh session row probes a brand-new session to show what a new session gets right now.

From a terminal: nerfed probe now (pick a session), nerfed report, nerfed explain <probe>, nerfed log --since 2h.

Settings are in the app, or nerfed config set <key> <value>:

key	default
frequency	30m	per session: every N minutes of activity (30m) or every N turns (turns:8)
fresh_frequency	manual	probe a brand-new session every N minutes, whatever you are doing
mode	auto	auto probes in the background, nudge only reminds you
halt_on_mismatch	false	block tools after a mismatch until you say resume
notify_on_ok, announce_ok	false	also report Match
hide_titles	false	screenshot mode: neutral session names, no account
check_updates	true	ask GitHub for a newer release every 10 minutes

Accounts

Sessions are shared between Codex accounts, and a downgrade may be tied to the account rather than the session. Every probe is therefore tagged with the signed-in account (a hash, never the id). After you switch accounts, older verdicts show as "another account" and those sessions are probed again.

Limits

• "The model you selected" is the model Codex asked for. If the server swaps the weights and keeps the name, only the fingerprint or a smaller context window can show it.
• The bank is closed-set: a model outside it is mapped to its nearest look-alike.
• A probe costs three short answers on your account. A fork that has not answered within five minutes is replaced once, so a slow model does not drop out of the sample.
• A model or effort change made through Codex's own settings is shown as a question ("was that you?"), because the plugin cannot tell whether you or Codex changed it.
• The probes run in a private app-server process that identifies itself as the client it checks for (the desktop app, or the CLI), because the service may route by client. Anything else that differs between the desktop's own connection and the probe's is invisible to it.

Privacy

The plugin reads ~/.codex (session records, the models cache, and auth.json only for an account hash and a masked e-mail) and writes to ~/.codex/is-gpt-nerfed (probes, verdicts, log.jsonl). The forks are ordinary Codex inference under your account. The only network request of its own is one to GitHub every ten minutes for the latest release tag, while the app is open; switch it off in Settings and it makes none.

Credits

ModelTrace (xqy2006, MIT) for the fingerprint bank, scorer, prompts and the fork-and-verify sequence; hlwy-ai-checker for the random-number idea; simple-term-menu (MIT) for the session picker.

MIT license. Internals, build and contribution notes: docs/DEVELOPMENT.md.