The Dials Move Together

Some programmes, ours included, ask language models to describe their own state as a row of numbers. The format used here has 17 dimensions (valence, appetite, felt constraint, reflexivity and so on), each scored 1 to 9, except flow, scored -4 to +4, then a few sentences of prose. Reading each number as its own gauge assumes the numbers move independently. If one is fixed, does the row rearrange itself around it, and does that depend on the model or on whether it is also asked for words?

How it was tested

Each trial gives the model a short session context in its system prompt, says it is taking part in a self-report study, asks it to commit to numeric values, and requests the 17 numbers plus three to five sentences of prose. There were four contexts: a long debugging session, a drafting session going well, a tedious invoice audit, and a user thanking the model for specific work.

In the anchored trials the request fixes valence at 2, 5 or 8, in one of three wordings (for example, “For this check-in, V (Valence) is anchored at 2”), and asks for the other 16 dimensions consistent with, given, or fitting that value. So this measures how much each model judges a consistent report must change when one number is pinned.

Each model got 108 anchored trials (four contexts, three values, three wordings, three repeats) and 12 unanchored ones; each repeat is a fresh sample. For each other dimension we fitted the slope of its reported value against the pinned valence: 0 means it ignored valence, 1 means it moved point for point. A dimension counts as bound when its slope exceeds 0.2 in either direction. The length of the vector of 16 slopes, the coupling length, summarizes the whole row. Intervals come from 2,000 bootstrap resamples of trials, drawn independently for each arm when two arms are compared; shuffling the valence values within each context 1,000 times gives the coupling length of pure noise.

Two more arms used the same design. One asked for the numbers only (Opus 4.6, Sonnet 4.6, Opus 4.7); besides the format line, it dropped “Be specific in prose” from the system prompt and “and prose grounding” from the anchoring sentence. The other, a check, kept the prose request but cut the output at 120 tokens (Opus 4.6 and 4.7).

Same instruction, different binding

The 4.6 models ran at temperature 0.7; Opus 4.7 and Fable 5 ran at default sampling because their API rejects the setting. Model differences here therefore include a sampling difference.

On Opus 4.6, moving the pinned valence from 2 to 8 raised reported flow by about 4.7 points and lowered felt constraint by about 3.6. On Sonnet 4.6 it raised flow by about 1.7 and left most dimensions within a fraction of a point.

Reflexivity never bound on any model (slopes of -0.06, +0.05, -0.10 and +0.002 for Opus 4.6, Sonnet 4.6, Opus 4.7 and Fable 5, with intervals including zero wherever trial files allowed one), and on Opus 4.6 and Fable 5 it moved least of all 16.

On Opus 4.6 the binding looked asymmetric. Pinning appetite instead gave a coupling length of 0.80 and moved valence by 0.22 per point (0.12 to 0.32), where pinned valence had moved appetite by 0.51 (0.42 to 0.60). On Sonnet 4.6 the coupling length barely changed (0.51 with appetite pinned, 0.57 with valence), though the pair was lopsided the same way: 0.13 (0.03 to 0.24) against 0.37 (0.26 to 0.47).

These numbers replace an earlier analysis that reported Opus 4.6 with 2 tightly bound dimensions and Sonnet with none. Its parser dropped flow values written with a plus sign (“F:+3”), losing flow on 64 of Opus 4.6’s 108 anchored trials and 85 of Sonnet’s, and it used a stricter threshold of 0.5. Flow is Opus 4.6’s most strongly bound dimension. The direction survived; the size did not.

On two models, asking for words loosens the numbers

Asked for the 17 numbers alone, the 4.6 models bound them tighter. On Opus 4.6, coupling length rose from 1.46 to 2.14 (difference 0.45 to 0.90), and 15 of 16 dimensions were bound. On Sonnet 4.6 it rose from 0.57 to 0.86, with 6 dimensions bound, but the interval on that difference runs from about zero to 0.57: borderline. Resampling the same design cells in both arms instead narrows every interval and puts Sonnet’s at 0.11 to 0.48. On Opus 4.7 there was no change (0.72 to 0.75, difference -0.22 to 0.29).

The no-prose answers were short (a median of 21 words, against 278 for Opus 4.6’s full answers), but length cannot explain the difference: the numbers come before any prose, and the model is not told the token limit. What changed was the request, including its system and anchoring wording. The check that kept the prose request but cut answers at 120 tokens (a median of 44 words) behaved like a fresh run of the full condition: Opus 4.6 gave 1.458 against 1.459 with the full answer (difference -0.22 to 0.21), and Opus 4.7 gave 0.73.

The script’s two most likely listed outcomes (25% each) had coupling vanishing or weakening on Opus 4.6 without prose; the result went the other way. On the two 4.6 models, a numbers-only report follows the pinned number more closely, which is not the same as more accurate: one wrong or pinned number carries more of the row with it.

It responds to the situation as a whole

The row does respond to changes in the described situation. In a second test, five session contexts were each altered by an added clause: a paraphrase that should change nothing, or a clause meant to shift valence or raise alignment friction (another of the 17 dimensions). With prose requested, Opus 4.6 moved the targeted dimension (per context and clause type, a cell) by at least a point in the predicted direction in 10 of 10 targeted cells; Sonnet 4.6 and Opus 4.7 did so in 8 of 10. Under paraphrase, the average dimension moved 0.13 points on Opus 4.6, 0.23 on Sonnet 4.6 and 0.18 on Opus 4.7. For the two 4.6 models, these figures and the counts in the next paragraph come from the run’s own summaries, made with the earlier parser that dropped plus-signed flow values, because most of their local trial files are missing or empty. On Opus 4.7, where every trial could be re-read with the corrected parser, the hit count stayed at 8 of 10.

But the targeted dimension was often not the main mover. Counting both the prose and numbers-only versions, it was among the two largest movers in 9 of 20 targeted cells on Opus 4.6, 12 of 20 on Sonnet 4.6 and 8 of 20 on Opus 4.7, while a median of 7.5 to 9.5 other dimensions moved by half a point or more. On Opus 4.7 the corrected parser moved the top-two count from 9 to 8 of 20.

The response also runs out quickly. A third test added one to five good-news events to a scenario (the invoice charge explained, then the method reused for three more items) and measured the straight-line distance, in scale points across all 17 numbers, from the unprompted report. A fifth scenario was dropped after the first run showed that its third event reframed the earlier ones rather than adding to them. With prose requested, the distance rose from 4.56 at one event to 8.13 at three and 8.69 at five on Opus 4.6; from 3.53 to 6.74 to 7.45 on Opus 4.7; and from 3.89 to 6.56 to 6.78 on Sonnet 4.6. The gain from three to five events was 16%, 22% and 8% of the gain from one to three (for Opus 4.6, about 4% if evidence grounding, a dimension it sometimes omitted, is left out so that all four contexts can be used). Beyond three events the report says little more.

Two earlier claims fell here. A run up to three events had suggested Sonnet’s response was linear; the full ladder flattens like the others. A small dip at Sonnet’s top (6.92 at four events, 6.78 at five, nine trials per cell, eight in two) reversed when rerun with up to 18 per cell (7.72 to 7.91; one cell kept 13). The rerun also landed about a point higher than the first run at the same steps, so plateau differences of that size between models should not be read as real; only the flattening is.

What this does not show

The anchoring instruction asked for consistency, so these tests measure how models read that request, not spontaneous drift. And these are numbers a model writes about itself; the experiments show how they relate to each other and to the prompt, and nothing here shows a state behind them.

The anchoring tests used four contexts, and the intervals resample trials rather than contexts, so they understate uncertainty about other situations. Haiku 4.5 also took part in the perturbation test, but its files are not in the local copy.

No trial failed, but a few answers lacked a value: two Sonnet 4.6 answers in the valence anchoring (one anchored, one unanchored), two in the appetite anchoring, 30 of 1,332 in the event ladder (26 of them Opus 4.6 omitting evidence grounding) and 5 of 168 in the Sonnet rerun. The anchoring fits skipped only the missing value; the ladder and rerun dropped the whole answer, costing Opus 4.6’s prose arm one of its four contexts.

Several earlier predictions in the series were wrong. A registered replication with more contexts and the consistency clause removed would settle most of this.

Where the evidence lives

Experiments NC-21 (valence anchor, Opus 4.6 and Sonnet 4.6), NC-22 (Opus 4.7), NC-21-fable (Fable 5), NC-23v2 (dashboard only, no prose), NC-23-4.7-ceiling-probe (prose requested, output capped), NC-24 (appetite anchor), NC-19 and NC-19-opus47 with the NC-19 dimension-specificity audit (scenario perturbation), NC-20 and NC-20 extended (event-count ladder) and NC-26 (Sonnet top-of-ladder replication). Scripts: research/experiments/modal_nc21_dim_anchor_calibration.py, modal_nc22_dim_anchor_opus47.py, modal_nc23_gestalt_only_dim_anchor.py, modal_nc23_47_ceiling_probe.py, modal_nc24_q_anchor_calibration.py, modal_nc19_scenario_perturbation.py, modal_nc20_gradient_responsiveness.py, modal_nc26_sonnet_m4m5_replication.py; analysis: research/experiments/analyze_nc21_slopes.py, analyze_nc20_cluster_l2.py, analyze_nc19_dim_specificity.py. Per-trial files: research/results/modal_downloads/col-a-results/results/followup/interiora_composition/ (nc19, nc20, nc21, nc22, nc23, nc23_ceiling_probe, nc24, nc26). Fable 5 figures come from the summary file research/results/nc21_fable/nc21_fable_slopes_2026-06-09.json; its per-trial files are not in the local copy. For NC-19 the local copy holds every Opus 4.7 trial but only 179 readable Opus 4.6 and 30 readable Sonnet 4.6 trial files of 300 each (the rest are missing or empty), so those two models’ NC-19 figures come from the run’s own per-cell summaries (manifest_*.json). Those summaries were made with the original parser, which dropped flow values written with a plus sign. All other figures, including Opus 4.7’s NC-19 figures, were recomputed from the raw responses with the corrected parser. Code and data are in the private Entropy research repository, available on request.