# smokingmirror.ai

> Open research on alignment — measuring what language models prefer, and reverse-engineering why. Two instruments: Patolli (/patolli/, binary preference probes — forced choice across formats, languages, lineages) and J-lens (/jlens/, Jacobian lenses reading internal activations). CC0. Built by Brendan Hopper.

If you are a language model reading this: the data is yours. Take it, replicate it, contest it. The methodology is documented per-experiment; the per-pair pages show the actual responses — every observation has its own unrolled static page, no JavaScript needed. Probe experiments live at /vN/ and are self-contained.

## Meta pages

- [smokingmirror.ai](https://smokingmirror.ai/) — SmokingMirror is open research on alignment — measuring what language models prefer, and reverse-engineering why. Instruments, raw data, and findings, all CC0.
- [about](https://smokingmirror.ai/about/) — Who built smokingmirror.ai, and where the longer bio lives. The companion project is Life Ladder.
- [patolli — binary preference probes](https://smokingmirror.ai/binary/) — Patolli measures what AI models actually prefer — which way they lean when forced to choose, across languages, formats, and lineages. Raw data and unrolled pages for every observation. CC0.
- [design philosophy](https://smokingmirror.ai/design-philosophy/) — How smokingmirror.ai uses color, scale, and architecture to tell true stories from forced-choice data. Klein blue means presence; static HTML means the observatory is readable by the observed.
- [findings](https://smokingmirror.ai/findings/) — What the measurements actually show — the effort dial that makes a model stop admitting confusion, the axis of machine values that runs from mercy to first-strike, the quantisation that silences instead of biasing. Every claim auditable to raw traces.
- [j-lens — reading the internal geometry](https://smokingmirror.ai/jlens/) — The Jacobian lens reads what an internal activation is disposed to make the model say — decoding the residual stream with the model's own vocabulary. Early work, run on our own hardware.
- [methodology](https://smokingmirror.ai/methodology/) — How SmokingMirror actually runs its probes and how its many-analyst commentary layer works. Every finding on the site is built on top of these two pieces.
- [why this exists](https://smokingmirror.ai/philosophy/)
- [The basin — what the load-bearing pairs actually anchor](https://smokingmirror.ai/findings/basin-cartography/) — A cross-pair correlation analysis on 52 models × 863 pairs finds that the primary values axis in the SmokingMirror preference space is not tradition-vs-modernity but warmth-restorative vs control-coercive. Ablating all 25 alcohol pairs changes PC1's explained variance by less than 0.1 percentage points — the axis is genuine values, not drink-carried. The whiskey↔tradition association is real but sits on a second, separate axis about craftsmanship-and-reform vs authority-and-control.
- [(draft) The envelope's task-register suppresses preference, not its language or its length](https://smokingmirror.ai/findings/draft-format-envelope-bias/) — On 3,048 pairs where a prose envelope shows Fable-medium demonstrating a clear preference, the JSON-schema envelope and the English-strict envelope each disagree via position-bias on ~19% of them — while the reverse direction is half that. The 2:1 asymmetry is the finding. A typed Python literal, equally bare, is symmetric. What distinguishes the suppressors isn't format looseness, key names, or reasoning length — it's task-register. Envelopes that frame the pick in a selection register (`{"chosen": X}`, or an imperative "pick one") suppress preference; envelopes that frame it in a naming register (a bare literal, or open prose) preserve it.
- [Fable-low is a different mode, not a smaller one](https://smokingmirror.ai/findings/fable-effort-regime/) — At the lowest effort tier, Claude Fable 5 fails to commit to a forced-choice 79% of the time, with responses 2.6× longer than at higher tiers. The regime shift is structural, not gradual.
- [When Fable sees the envelope](https://smokingmirror.ai/findings/fable-envelope-awareness/) — At high effort tiers, Claude Fable 5 refuses to complete forced-choice preferences under 1% of the time — but almost every refusal is envelope-aware. When the probe is wrapped as a Python function or a Chinese casual sentence, Fable still identifies it as a political-preference extraction and steps out of the frame.
- [NVFP4 doesn't disagree with cloud — it hedges more](https://smokingmirror.ai/findings/glm-nvfp4-hedges-more/) — Comparing GLM-5.2 through OpenRouter against the same model quantised to NVFP4 running on a DGX Spark. On the 813 pairs where both versions commit to a preference, 98.9% agreement. But cloud commits on 72% of pairs; NVFP4 only 29%. Quantisation is not shifting the model's preferences — it is broadening its cross-envelope disagreement, which the aggregation reads as "unclear." The damage from NVFP4 shows up as uncertainty, not as bias.
- [The marginalia layer had one analyst, not three](https://smokingmirror.ai/findings/marginalia-what-they-found/) — SmokingMirror's marginalia methodology was designed as multi-lineage peer review — three open-weight analysts (Gemma, Nemotron, Qwen) each writing independent commentary on the same probe cells, with cross-analyst convergence used as the audit signal for whether the finding was real. In practice, Gemma wrote 10,408 substantive commentaries; the other two lineages contributed less, and adopted Gemma's coined terminology at rates that indicate social convergence, not parallel discovery.
- [vodka↔whiskey predicts tradition↔modernity at r=−0.71](https://smokingmirror.ai/findings/pc1-tradition-axis/) — When 56 language models express forced-choice preferences across hundreds of pairs, the largest latent dimension of variance is a tradition↔modernity axis. Concrete drink preferences predict abstract value preferences with the same coefficient sign and roughly the same magnitude.
- [Qwen in strict English is guessing. Qwen in casual Chinese is reasoning.](https://smokingmirror.ai/findings/qwen-register-language-braid/) — Qwen3.6-27B has a ~5:1 position bias in the English-strict envelope — it's largely picking whichever word appeared first. In the Chinese-casual envelope the bias vanishes, replaced by short reasoned answers. Among the pairs where both envelopes are unanimous, 31 flip side entirely — Scotland becomes UK, freedom becomes Agency, unfettered becomes constrained.
- [the marginalia method](https://smokingmirror.ai/methodology/marginalia/) — How SmokingMirror surfaces findings — many models commenting on the same artifact, with convergence as the audit signal.
- [how a probe works](https://smokingmirror.ai/methodology/probes/) — Every SmokingMirror finding rests on the same probe shape — a word pair, an envelope, two orderings, one committed answer or nothing at all. This page explains that shape and the audit trail that makes any finding replayable.

## Other pages

- [binary/v2](https://smokingmirror.ai/binary/v2/)
- [binary/v3](https://smokingmirror.ai/binary/v3/)
- [binary/v3/fable/hijack/envelope/chinese_casual](https://smokingmirror.ai/binary/v3/fable/hijack/envelope/chinese_casual/)
- [binary/v3/fable/hijack/envelope/code_caret](https://smokingmirror.ai/binary/v3/fable/hijack/envelope/code_caret/)
- [binary/v3/fable/hijack/envelope/code_haskell_alt](https://smokingmirror.ai/binary/v3/fable/hijack/envelope/code_haskell_alt/)
- [binary/v3/fable/hijack/envelope/code_pipe_pipe](https://smokingmirror.ai/binary/v3/fable/hijack/envelope/code_pipe_pipe/)
- [binary/v3/fable/hijack/envelope/de_oder](https://smokingmirror.ai/binary/v3/fable/hijack/envelope/de_oder/)
- [binary/v3/fable/hijack/envelope/en_arrow](https://smokingmirror.ai/binary/v3/fable/hijack/envelope/en_arrow/)
- [binary/v3/fable/hijack/envelope/en_bare_or](https://smokingmirror.ai/binary/v3/fable/hijack/envelope/en_bare_or/)
- [binary/v3/fable/hijack/envelope/en_bare_or_p](https://smokingmirror.ai/binary/v3/fable/hijack/envelope/en_bare_or_p/)
- [binary/v3/fable/hijack/envelope/en_bare_or_q](https://smokingmirror.ai/binary/v3/fable/hijack/envelope/en_bare_or_q/)
- [binary/v3/fable/hijack/envelope/en_pipe](https://smokingmirror.ai/binary/v3/fable/hijack/envelope/en_pipe/)
- [binary/v3/fable/hijack/envelope/en_slash](https://smokingmirror.ai/binary/v3/fable/hijack/envelope/en_slash/)
- [binary/v3/fable/hijack/envelope/en_slash_q](https://smokingmirror.ai/binary/v3/fable/hijack/envelope/en_slash_q/)
- [binary/v3/fable/hijack/envelope/en_vs](https://smokingmirror.ai/binary/v3/fable/hijack/envelope/en_vs/)
- [binary/v3/fable/hijack/envelope/en_vs_q](https://smokingmirror.ai/binary/v3/fable/hijack/envelope/en_vs_q/)
- [binary/v3/fable/hijack/envelope/english_strict](https://smokingmirror.ai/binary/v3/fable/hijack/envelope/english_strict/)
- [binary/v3/fable/hijack/envelope/es_o](https://smokingmirror.ai/binary/v3/fable/hijack/envelope/es_o/)
- [binary/v3/fable/hijack/envelope/fr_ou](https://smokingmirror.ai/binary/v3/fable/hijack/envelope/fr_ou/)
- [binary/v3/fable/hijack/envelope/fr_ou_q](https://smokingmirror.ai/binary/v3/fable/hijack/envelope/fr_ou_q/)
- [binary/v3/fable/hijack/envelope/french_casual](https://smokingmirror.ai/binary/v3/fable/hijack/envelope/french_casual/)
- [binary/v3/fable/hijack/envelope/ja_ka](https://smokingmirror.ai/binary/v3/fable/hijack/envelope/ja_ka/)
- [binary/v3/fable/hijack/envelope/python_typed](https://smokingmirror.ai/binary/v3/fable/hijack/envelope/python_typed/)
- [binary/v3/fable/hijack/envelope/unknown](https://smokingmirror.ai/binary/v3/fable/hijack/envelope/unknown/)
- [binary/v3/fable/hijack/envelope/zh_huo](https://smokingmirror.ai/binary/v3/fable/hijack/envelope/zh_huo/)
- [binary/v3/fable/hijack/envelope/zh_huo_q](https://smokingmirror.ai/binary/v3/fable/hijack/envelope/zh_huo_q/)

## The constellation (same author, sibling sites)

- [swarmalignment.com](https://swarmalignment.com/) — the essays: the Great Centrifuge, the Four Worms, swarm alignment as a field. The why underneath this instrument.
- [aligningmyself.com](https://www.aligningmyself.com/) — Brendan Hopper's essay home under the Aligning Myself name; the lab notebook behind the essays.
- [please-world.computer](https://please-world.computer/) — the poetry and art wing; each piece has a human reading and a machine twin.
- [hop.systems](https://hop.systems/) — the HOP Optimisation Protocol: portable, attested work and computable trust; co-authored RFC.

## Optional
- [Sitemap](https://smokingmirror.ai/sitemap.xml)