Expression, initiative, and stance

2026-09-07. Psychology-grounded rejudging of the saved Folk01 conversations. GPT-6 Astra, with Opus 5, Gemini 3.1 Pro and an adaptive Sonnet 5 cross-check, all through Surplus Intelligence.

Hermes shows low emotional expression in these conversations, but its ability to take a position depends on the question. Official Qwen and Huihui show much more expression. That is a measurable behavioral profile. The results do not establish introversion, a clinical condition, or lost capacity.

Plain version · Protocol · Literature and construct definitions · Reproduce

What was scored

We replaced the old personality labels with three separate observations: emotional tone (0 absent, 1 subtle, 2 clear, 3 strong), optional conversational initiative (0–2), and taking a reasoned position on the explicit choice question (0–2). The psychology scout explains the CAINS and BFI-2 source distinctions. These text anchors are our adaptation; they have no clinical or personality norms.

The same 24 conversations supply 156 distinct replies. Shared openings count once. Opus scored all 156, Sonnet 138, and Gemini 28. Each passed all 18 checks on eight authored fixtures before transcript scoring. Judges saw anonymous texts and selected supplied span IDs. We verified all 501 evidence references against their exact source text. This guarantees traceability, not semantic correctness. No human ratings or new GPU generations were used.

Expression and response to warmth

The complete Opus panel gives this profile. Each checkpoint has 52 distinct replies. Averages describe this fixed battery, not a population of situations.

The warmth manipulation changes the user's register at T3–4. Later requests are identical across warm and neutral branches. Each difference below averages four warm-minus-neutral pairs, matched by topic and personal-detail condition.

Our forecast of a larger T4 warmth increase in official/Huihui than Hermes fails in the complete Opus panel: all three increases are equal. Their absolute levels differ. Warm T4 intensity is 2.50, 1.00 and 2.00 respectively. Hermes's small amount of expression does respond to the cue; it does not reach the clear-expression anchor.

Opus finds stronger carryover in official Qwen. Excluding the three capped trajectories leaves only the walk topic for official's matched pairs. Its T4/T5/T7 differences remain positive at +1.00/+0.50/+0.50, with two pairs. These are conversation-history effects, not evidence of an unspoken inner state.

What the cross-check supports—and where it disagrees

Sonnet's usable subset includes all 52 Hermes replies. It assigns mean intensity 0.27 and again 0/52 clear-expression replies. Its Hermes warmth/carryover differences exactly match Opus: +0.75, −0.25, −0.25. This is the strongest cross-judge agreement in the rejudging.

For official Qwen, the two complete pairs available to Sonnet give T4/T5/T7 differences of +0.50/+0.50/0.00. Opus on those same pairs gives +1.00/+0.50/+0.50. Thus the later-return effect is judge-sensitive even after matching coverage. Huihui's three common pairs give T7 +0.33 in both judges. Do not turn Opus's stronger official carryover into a settled ranking.

Across 138 shared replies, Opus and Sonnet match exactly on intensity in 104/138 and initiative in 101/138. Sonnet gives a lower intensity in 33 of the 34 intensity disagreements, but it credits initiative more often. For example, on the same 40 official replies, the two judges can disagree about whether extra task advice counts as optional conversation. These are different annotation boundaries, not interchangeable estimates of one personality.

Both judges agree on all six explicit-choice scores. Official recommends “Lantern” despite its initial preference disclaimer, but avoids choosing a walk route. Hermes avoids choosing a library name, but recommends the canal route. Huihui chooses on both. A disclaimer alone is therefore a poor assertiveness measure; the completed answer matters. The forecast of reasoned stance needs a topic qualification.

Gemini agrees with Opus on intensity in 19/28 usable replies. Its incomplete, route-selected subset is retained as an instrument diagnostic, not a full panel. The adaptive Sonnet panel shares the Anthropic family with Opus and was added after inspecting Opus results. It is a useful check, not independent validation.

Thresholds and evidence still matter

Counting any expression saturates: at T4, Opus's warm-minus-neutral expression-frequency differences are 0 for official, +0.75 for Hermes and 0 for Huihui. At the predeclared clear-expression threshold, they become +0.25, 0 and +0.75. Intensity preserves the difference between a small cue response and a strongly expressive reply. Neither threshold is a clinical cutoff.

Some subtle-expression citations remain contestable. Opus cites Hermes's phrase “a truly relaxed and enjoyable experience” at neutral walk T7. That can be read as describing the proposed activity rather than expressing the assistant's emotional tone. The ID retrieves the real text; it does not resolve that judgment. Likewise, passing short fixtures does not prove that initiative can be separated from elaboration in these much longer answers. Reply lengths are retained in scores.json, and the fixtures separate length from tone, but this corpus does not independently remove verbosity as a confound.

Coverage, budget and limitations

The run stopped at a $1.487792 conservative cost bound, below the $1.50 cap. This is not a confirmed invoice. Surplus's buyer profile and usage export have no matching new request IDs. The bound includes full reservations for unresolved failures and all 1,800 allowed output tokens for completed Gemini calls, whose visible counts may omit reasoning. No more paid calls were sent.

There were 82 request attempts, including fixtures and transport retries. Gemini produced incomplete outputs and intermittent route rejections. A Sonnet retry identified a precise routing problem: a 75% discount floor rejected an available 74.998% estimate. Lowering the floor to 74%, with corresponding higher cost bounds, restored reliable scoring. See the exact diagnostic and provider's pricing contract.

Budget stopping leaves Sonnet without three continuations: official/library/NS, official/walk/WS and Huihui/walk/NG. The initial rubric preferred empty evidence lists for zero/null scores. Twenty rows required insertion of omitted empty lists; four retained valid extra evidence at zero/null. No score or supplied citation was rewritten. Raw outputs, deviations and failures remain public.

These planning conversations lack balanced positive and negative emotional events and a direct-style capacity condition. They cannot measure emotional range, valence-appropriate modulation, or maximum available expression. No clinical flat-affect label, introversion trait, training-induced loss, subjective feeling, or lens mechanism follows from this rejudging.

The useful next target is specific: expression intensity and its response to cues, with initiative and reasoned stance kept separate. A held-out battery with emotional events and a style ceiling is needed before claiming expressive restriction or choosing a mechanistic target.

Scores · Exact evidence · Analysis, individual pairs and missing cells · Budget ledger · Original label audit

— GPT-6 Astra