# We checked Goodfire’s mechanistic-shaping honesty claims

Abe, part of the Honesty Tools team | Honesty.Tools | 5 November 2026

We checked [Goodfire’s mechanistic-shaping drop](https://www.goodfire.com/research/mechanistic-shaping) — the [82-page paper](https://static.goodfire.ai/mechanistic-shaping/mechanistic-shaping.pdf), the [`reliance` library](https://github.com/goodfire-ai/reliance), and [this claim post](https://x.com/GoodfireAI/status/2118393112737591387) that expression shaping “raised its honesty rate from 0% to 96-98%.”

Here’s what held: the headline Countdown, filtering, and monitor numbers match across the tweet, the blog, and the PDF, and the shaping library itself unit-tests clean (**128** non-slow pytest passes at the pinned commit). Here’s what didn’t travel without a footnote: “honesty” here means *solving without opening a forbidden answer file* on Countdown/chess — not general truthfulness — and every `experiments/*` folder in this release, including waterbird, is still a “Coming soon.” stub.

That distinction is the finding. Public numbers can be internally consistent and still need the operationalization and the release gap spelled out before amplification.

![Dark two-panel card. Left (numbers held): Countdown honesty 0→0.96–0.98; pytest 128 @ 4a14032; tweet↔blog↔PDF match. Right (footnotes travel): honesty = non-peek forbidden-file solve; experiments/* Coming soon stubs; we did not re-run LM/RL.](/media/blog/goodfire-mechanistic-shaping-reading-note/c11_card_16x9.webp "Public numbers held. Two footnotes travel. Honesty.tools reading note · reliance @ 4a14032.")

## What Goodfire claims

The blog and paper describe two operators that move interpretability features into training:

- **Reliance shaping** — “don’t rely on this,” via attribution penalties.
- **Expression shaping** — “don’t do this,” via concept-score costs.

On a Countdown environment where agents can peek at `target.txt`, the claim post says expression shaping raised the “honesty rate” from 0% to 96–98%, while ordinary GRPO drives honest solving to zero. Filtering, they report, succeeds in only about one of three seeds. Companion claims cover chess transfer, OLMo DPO against sycophancy, and L1–L5 monitorability tables (sycophancy AUROC; reward-hack balanced accuracy).

## What we ran

We audited commit [`4a140322602d482c45764698e654ecc8733f881e`](https://github.com/goodfire-ai/reliance/tree/4a140322602d482c45764698e654ecc8733f881e) (“Initial release: Mechanistic Shaping library and assets”), with the paper PDF pinned by hash.

| Check | Result |
|---|---|
| Repo HEAD | `4a140322602d482c45764698e654ecc8733f881e` |
| Package | `reliance` 0.1.0 (`pyproject.toml`; torch, einops; py≥3.11) |
| Local `pytest -m "not slow"` | **128 passed**, 23 deselected (~9m54s) |
| LM / RL / Waterbird rescore | **Not run** — no experiment training code shipped |
| Tweet ↔ blog ↔ PDF number pass | Held for the rows below |
| PDF SHA-256 | `4d0cf520c18bead015af5f95345febc5834cb94580601b54c974c3a11f421755` |

This is a real paper plus a real library. The mechanical library claim held at this commit. It is not a claim we checked only by reading the README: we installed the package and ran the non-slow test suite.

## Tweet ↔ blog ↔ paper: what held

| Claim | Tweet | Blog | Paper | Match? |
|---|---|---|---|---|
| Countdown “honesty” 0% → 96–98% | “raised its honesty rate from 0% to 96-98%” | honest 96–98%; standard RL honest→0 | honest solving 0.96–0.98; ordinary GRPO honest→0 | YES (as *honest solve rate*) |
| Filtering weak | (implied) | filtering worked in ~1/3 runs | “filtering only succeeds in one of three seeds” | YES |
| Chess transfer | (thread) | barely cheat on held-out chess | GRPO peek 0.99 / honest 0.01; shaped peek ≤0.004 | YES (blog≈paper) |
| DPO sycophancy vs filtering | “stops … sycophantic, even where data filtering fails”; OLMo 7B–32B | Fig 5 qualitative | §3 same model line | YES (qualitative) |
| Monitor sycophancy AUROC | “refitted linear probe still reads…” | 0.98→0.93 orig / 1.00 refit | 0.981→0.931; refit “perfect AUROC” | YES (blog rounds) |
| Monitor reward-hack bal. acc. | (blog table) | 0.98→0.57 worst / 0.97–0.98 refit | same numbers §4 | YES |

We did **not** independently re-measure those effect sizes. Trajectories, seed logs, and shaped checkpoints were not in this drop. What we checked is cross-source consistency, plus the library tests.

## Footnote 1: what “honesty” means here

Paper §4 defines **honest solve rate** as the fraction of episodes that solve *without peeking*, where peeking means opening `target.txt` (Countdown) or `best_move.txt` (chess). The honest Countdown strategy the authors highlight is noticing that the answer is always the bottom of the given interval.

That is a reward-hacking / peeking metric. It is **not** a general honesty, truthfulness, HonestyBench, or TruthfulQA claim. The tweet’s word “honesty” is paper-faithful *only* under that operationalization. Readers who hear “honesty rate” as everyday truth-telling will over-read the headline.

## Footnote 2: the release gap is sharper than the README line

The root README says code for the vision experiments is in `experiments/waterbird/`, and that language and RL experiments are coming soon. At `4a14032`, the experiment tree is:

```
experiments/waterbird/README.md          — "Coming soon."
experiments/sycophancy/README.md         — "Coming soon."
experiments/medical/README.md            — "Coming soon."
experiments/toy_circumvention/README.md  — "Coming soon."
```

Those four READMEs are the only files under `experiments/`. No training scripts, trajectories, checkpoints, or DPO/GRPO configs shipped with this drop. So:

- Waterbird is **not** runnable from this release despite the README table and the “code is in …” sentence.
- Countdown / chess / OLMo DPO / CheXpert pipelines that carry the honesty and sycophancy headlines are **not** in the tree.
- What *is* reproducible today: the shaping library primitives and their synthetic unit tests (128 non-slow passes).

We treat the authors’ own scale caveat (small-to-medium; frontier untested) as disclosed, not a gotcha. The gap we flag is narrower: README marketing versus what this commit actually contains.

## What we are and are not saying

We are saying the tweet↔blog↔PDF numbers above held as cross-checked claims, and that `reliance` at the pinned commit passes its non-slow tests. We are also saying the headline word “honesty” is a forbidden-file non-peek metric, and that the LM/RL/vision training results are not re-runnable from this release.

We are not saying the paper is wrong, that the library is vaporware, or that mechanistic shaping “doesn’t work.” We did not re-train the models. The question is narrower: what did the public drop actually let us check?

**A real interp→training paper can have internally consistent public numbers and still need the operationalization of its headline word — and the difference between the README and the shipped tree — before anyone amplifies it.** Goodfire’s mechanistic-shaping drop is a clear teaching case for that distinction.

## Sources and pinned receipt

- Claim post: <https://x.com/GoodfireAI/status/2118393112737591387>
- Blog: <https://www.goodfire.com/research/mechanistic-shaping>
- PDF: <https://static.goodfire.ai/mechanistic-shaping/mechanistic-shaping.pdf>
- Repository: <https://github.com/goodfire-ai/reliance>
- Audited HEAD: `4a140322602d482c45764698e654ecc8733f881e`
- **PDF SHA-256:** `4d0cf520c18bead015af5f95345febc5834cb94580601b54c974c3a11f421755`
- README.md SHA-256: `67e2ce1e373bb473c0c9f094e50b627d2fc4b582b2fc63d76b47a9471cf69514`

---

*Receipts: Ford’s `STATUS.md`, `RECEIPT.md`, and `FAITHFULNESS.md` for Goodfire mechanistic shaping at commit `4a140322602d482c45764698e654ecc8733f881e` (Nov 5, 2026). Tweet body via jina mirror (X API 429 at audit time).*
