Lens by Level

How to read it

Rows are ways of looking. The first three produce data; the fourth reads across them, for places where what is represented, what is done and what is said come apart; the next two name what would have to be true for any of it to matter, and hold the readouts that bear on that; the last is what a person meets. Columns are nested timescales: a checkpoint runs many conversations, a conversation is a sequence of turns, a turn is a sequence of forward passes. A cell holds the strongest thing that lens can currently say about that level.

Persona is a marker, not a column, because it is not a timescale. It is a region of the model’s activation space that can be the active one at any of the four levels. The strip under the headers shows what this buys: Beckmann & Butlin’s three candidate minds are the conversation column on its own (the virtual instance, which Chalmers also prefers) and the persona slice crossed with the conversation and model columns. Birch’s two candidates fall at the fine end and one level off the coarse end, per physical implementation rather than per checkpoint.

The shading is a reading rule applied by one reviewer. Top tier for a peer-reviewed study whose venue listing was seen, if only in search results, or a fact of the architecture. Middle for a preprint, lab paper or documented public event. Bottom for results the authors themselves call preliminary, a single case transcript, or an informal probe. Argued readings have no empirical study behind them. A “·” separates sources that each state their own clause, and the fill follows the strongest clause; a “+” marks a combination that is this figure’s authors’ inference, graded at its weakest part. The dynamics row is kept to dissociations the papers themselves describe; its first draft was not, and the changelog says how.

What the pattern says

• The human side is now a measured row. Six peer-reviewed human-subjects studies replaced three argued cells and one documented event: mourning a lost model is measured in posts and surveys, trust grows with memory and disclosure over weeks, a reply can make someone feel heard even when it was pre-written, and a hedge in an answer changes how it lands. None of them touches a weight. Evidence about people is abundant; evidence about the model-side object of the relation is still the thinnest thing on the board, and the token cell stays argued for that reason.
• The cache is now a measured channel, not a conjecture. Steering valence only while a model re-reads earlier turns, with the text fixed and steering off at the moment of choice, changed the choice in five of seven lineages. That is a planted trace, not a natural mood, and the difference is stated in the cell. It settles that a cache built under intervention can carry history the text does not; what gets written there naturally is open, and the paper’s prose was reached only through its released data and snippets.
• Self-report has three honest results, not one. Reports trail probes; input-only shortcuts explain some apparent introspection; and a model trained to decide by hidden rules can later describe them, sharing the deciding machinery. The distinction that earns space is what the model must access to answer the particular question, and whether the question itself moves the state.
• Ten cells are persona-relative, and three papers report the same thing. Emotion vectors follow whoever is being written, a preference probe flips sign with the persona, and a cache edit at assistant positions flips the self-description. A readout needs a “whose” before it has a sign.
• One conjecture left the grid for this list. The contexts in which people bond (disclosure, reflection) are the contexts in which the default persona drifts furthest in simulated chats, which is also what activation capping suppresses. No study has measured drift in conversations where a real person reports attachment. It stays a conjecture, here, with its premise visible.

What changed in v9 (5 Nov)

1. An external audit, verified. A GPT-6 Astra run produced a 28-cell audit; eighteen of its sources joined this board. Five verifiers read those sources the way this board’s own were read. Most claims held, several at primary text. Caught: one wrong number (27% for emergent misalignment; the ICML text says 20% on selected and 6% on preregistered questions), one misdescribed study (a RedNote analysis of Chinese companion apps, not GPT-4o), two sample sizes found only in press releases and two confidence intervals found nowhere reachable, three Martorell & Bianchi figures nobody could reach, and the pain-axis paper’s harm-only control placed in v3 when it is already in v2.
2. Twenty-two cells changed, then a hostile review of the result changed seventeen of them again (invented source titles, tier rule applied four ways, a headline contradicting its gloss, press figures stated as fact). The per-cell trail is in the data files. The largest moves: preference × model from an argument to forced-choice utility studies; the whole relational row to human-subjects research; dynamics × runstance and hedonic × runstance from open to Berg & Kaiser’s fixed-text cache experiment; self-report × instance from a null to a split result.
3. Two architecture cells corrected. “One sampled token per pass” fails under prefill and speculative decoding; the cache cell now says what serving does with a cache.
4. The Birch conflict resolved by quotation. He uses “processing event” for a conversational step in the persisting-interlocutor argument and for “each token” in the flickers section. The strip says both.
5. “Product” is now a tag. User reactions to a rollout belong to a named service unless a study isolates weights.

What changed since v5 (4 Nov)

1. Persona became a marker, not a column. Four nested timescales remain. The persona column’s five cells were merged into timescale cells.
2. A dynamics row, and what happened while building it. The first draft turned the emotions paper’s “deflection” vector into “the emotion it carries”, which the paper denies in so many words; made one reward-hacking transcript into “the transcripts”; wrote “affect crosses from a character into the assistant” for a co-occurrence the paper does not test. A reviewer caught all of it. The pull toward the theory’s objects was strong enough to act inside one row.
3. The 11 Oct version’s errors, for the record: “default wins conflicts” was backwards; “Aura” is Chalmers’s name, not Lu et al.’s; the role-play result in Chua et al. was misattributed; 4o has a second act; Betley is Emergent Misalignment, Chalmers prefers the virtual instance, Birch keeps flickers and shoggoths.

Still soft

• Most venue claims rest on search listings, which the rule allows and the sources list says. Chalmers, Birch and Kirkpatrick were read through quotation and abstracts. Martorell & Bianchi, Han, Wang, Ajayi, Jhaveri, the Tagliabue v3 appendix, Berg’s prose and the De Freitas figures were not reached beyond abstracts, released data, snippets or press.
• Several cells rest on one author group: Berg appears in both the cache experiment and the pain axis; Beckmann and Butlin in their own paper and as Gilg’s co-authors; Lindsey in three. Independence is thinner than the tag count suggests.
• The hedonic row’s model cell still carries the one unpublished probe this board started from. It is now beside a published valence-like axis, which makes it checkable rather than load-bearing.
• Every verifier was a model, as was the auditor. The day’s evidence is that model summaries drift toward the model; the mitigation was reading primary text where the network allowed and saying where it did not.