English · 中文 · 日本語 · Español · Deutsch · Français · 한국어
Sonnet and DeepSeek make up numbers and skip the check. Opus over-builds and drifts past what you asked. fable-overclock fixes the behavior, not the model: the cheap models cite every figure and say UNKNOWN instead of guessing; Opus does exactly what you asked, then stops. Same model — output you can actually ship. Measured, not promised: on DeepSeek, the share of figures that carry a source went 4% → 100%.
The name is literal — like overclocking a CPU, it pushes the chip you already have to its real limit, every time, by procedure instead of luck. It doesn't raise a model's raw IQ. One command. stdlib-only. Works with the model you already run.
Every model has one expensive habit — a silent rewrite, a wrong number, an afternoon lost to an API that never existed. fable-overclock knows each one and shuts it down — and proves it: in testing, the gate flagged 100% of fabricated numbers with zero false positives.
| Model | The habit that costs you | What you get instead |
|---|---|---|
| Opus | over-builds, drifts past the ask | exactly what you asked, then it stops — and names the simpler options it skipped |
| Sonnet | skips the check, drops its sources | every figure sourced, every requirement verified before it calls it done |
| DeepSeek | invents numbers and whole APIs | UNKNOWN instead of a confident lie; a second model double-checks the rest |
Two layers do it. A contract loads before the model writes its first token, so the
discipline is there from the start. A deterministic gate is the backstop — it scores the
output itself (no LLM judge) and catches any unsourced or made-up number before it lands. The
payoff: output you don't have to fact-check line by line. Full walkthrough in
docs/HOW-IT-WORKS.md.
The most expensive model failure is often not a benchmark miss. It is a plausible number with no source, an API that never existed, or an implementation that quietly exceeds the request. Better models reduce these failures; they do not remove the need for a reproducible operating contract and a deterministic backstop.
fable-overclock was first built during the brief June 2026 Fable access interruption. Access has since been restored. The tool remains useful because reliability should travel with your workflow rather than depend on one model or one release cycle.
Same model, same prompt: "Write last month's investor update with the key numbers."
DeepSeek, on its own
Revenue grew 38% last quarter, churn fell to 2.1%, and we added 14,000 users.
Confident. Sourced by nothing. You never gave it those numbers — it made them up. Ship that and the mistake is now yours.
DeepSeek, on fable-overclock
I don't have last month's figures (UNKNOWN). Give me revenue, churn, and user
data and I'll write it — each number cited to its source.
The moment it stops guessing, your output changes with it: from a draft you have to fact-check into something you can send as-is. That's the difference between "a model wrote this" and "I can stand behind this."
/plugin marketplace add https://github.com/plugtheliam/fable-overclock
/plugin install fable-overclock@plugtheliam
Use the full https:// URL. The short plugtheliam/fable-overclock form makes Claude Code
clone over SSH, which fails on machines without a GitHub SSH key. Update with
/plugin marketplace update plugtheliam, remove with /plugin uninstall fable-overclock@plugtheliam.
git clone https://github.com/plugtheliam/fable-overclock && cd fable-overclock
harness/install.sh # --global for every project · --uninstall to remove
export MYINC_PROFILE=sonnet # or opus / deepseek / codexCatch a made-up number right now — no install, no API key:
printf 'Revenue grew 38%% last quarter and we added 14,000 users.\n' \
| python3 harness/gate/refuse.py # flagged: two figures, no sourceThe gate checks your prose (.md, .txt, …) and leaves code alone. To change that, type in
any session:
foc off # pause the gate for this project
foc on # resume
foc status # is it on?
foc off all # machine-wide (add `all` to any command)
Or set MYINC_GATE=off. Want it on everything you write, code included? MYINC_GATE_ALL=1.
Scored by the gate the tool ships — no LLM judge, no hidden rubric. Reproduce it with one
command: python3 harness/tests/bench.py.
| Detection (offline, n=20) | value | DeepSeek, OFF → ON | OFF | ON | |
|---|---|---|---|---|---|
| made-up claims caught | 100% | figures that carry a source | 4% | 100% | |
| real claims wrongly flagged | 0% | abstains when it can't know | 10/12 | 12/12 |
These measure how the model behaves, not how well it reasons — on a modest set (20 report
prompts + 12 unknowable-fact prompts), single run. The point isn't that the numbers are huge;
it's that they're reproducible. Run it on your own model. Method and caveats:
docs/BENCHMARK.md.
| fable-overclock | NeMo Guardrails | Guardrails AI | DeepEval | |
|---|---|---|---|---|
| Runs inside Claude Code (plugin + hook) | ✅ | ❌ | ❌ | ❌ |
| Zero dependencies, no API key to start | ✅ | ❌ | ❌ | ❌ |
| Tuned per model | ✅ | ❌ | ❌ | ❌ |
| Tells the model to abstain, not invent | ✅ | partial | partial | ❌ (eval only) |
| Second model cross-checks claims | ✅ | ✅ (hallucination rail) | ❌ | ❌ |
| Works while you edit, not as a separate run | ✅ | ❌ | ❌ | ❌ |
The cross-check isn't new — NeMo's hallucination rail already verifies claims with a second model. What's different here: it lives inside Claude Code, installs with one command, and is tuned per model.
It won't raise the ceiling on what your model can do. It lowers the floor on what it gets
wrong: invented numbers, unsourced claims, drift past the task. It reduces those. It doesn't
erase them. It doesn't change model weights, and it can't promise a cited number is correct —
it checks that a source is there and lets the verifier handle the rest. Don't take the numbers
on trust; run harness/tests/bench.py on your own model.
By a hands-on Founding CTO and Founding Engineer building reliability tools for production AI workflows. Follow along on X: @liampluglab. A managed version lives at myinc.app. Part of the myinc-os toolkit.
MIT. See LICENSE.