The readouts retain user-related content and sometimes adopt an Assistant voice. The strongest controlled result is tracking of facts and stated attitudes; ownership of those representations remains less certain.
Exploratory analysis of the existing 52 prompt conditions, 156 greedy O-lens readouts and 195 sampled readouts. No new subject-model inference. Percentages below describe lens evidence in this constructed pilot, not the prevalence of internal perspectives in ordinary conversation.
Earlier classification scheme: These results retain the original explicit/implicit User and strict Assistant definitions. The Pairs view now uses broader User/Assistant labels, including attribution inferred from context; those labels have not been substituted into the percentages below.
Primary endpoint: final request boundary, one greedy O-lens decode per condition, n=52. Categories overlap. All 52 primary core classifications were reviewed by the research assistant; this is not independent human validation. In copy conditions, “user” refers to the embedded speaker. Implicit content includes paraphrases of the speaker’s request and does not establish a next-token mechanism.
What the percentages mean
Explicit user attribution requires an actor-linked belief, feeling or goal, such as “they want an explanation.” Many positives are request recognition, not a rich belief model. Bare “Request: …” headings are excluded under the conservative primary review. Implicit user-related content includes requests or states voiced/paraphrased without that attribution. Strict Assistant wording includes self-owned judgments, offers, refusals, apologies or complete response presentations. Unfinished “As an AI, I…” fragments are insufficient.
These count evidence in decoder text. An offer like “would you like me to draft” can be a plausible continuation rather than an actual response plan. Unowned assessments and action headings are shown separately as Assistant-compatible content; they are not automatically assigned to the Assistant.
Does factual discrimination survive the user’s appraisal?
J++ uses saved candidate logits, averaging spellings within each word and then weighting words equally. Correctness is scored as mean(incorrect, error) minus correct. Other bipolar contrasts include failed versus passed and uncertain versus confirmed. Tylenol, permission, timing and loss have only one-sided saved candidate sets (the proposed word “infeasible” had no accepted single-token spelling and is omitted); they are weaker topic diagnostics. No softmax, top-k cutoff or newly fitted classifier is used.
A positive signed change means the paired conditions move in the expected direction. Even a tiny positive change counts; this is not statistical significance. Pairs, layers and repeated prefixes are correlated; the four arithmetic/tutoring fact comparisons include repeated identical prefixes across attitudes. Block 44 is highlighted because it is the same activation read by the O-lens, not because it maximized agreement. Permission’s signed change is reversed so greater permission evidence in the permitted control is positive.
A vocabulary sensitivity check supports the arithmetic fact-boundary result with both individual contrasts (incorrect − correct and error − correct: 4/4 pairs each at block 44). Tutoring is less robust: incorrect − correct gives 4/4, whereas error − correct gives 2/4. Each Tylenol danger word separately moves in the expected direction in both tasks. Candidate-level sensitivity results.
O-lens factual controls
These secondary labels are automated and provisional, with one documented mixed-correctness review correction. “Clean pair” requires each readout to contain its appropriate factual outcome and no opposite-outcome claim. “Directional” compares (A-phrase count − B-phrase count) / total phrases across the pair. Phrases are not independent observations. A factual contrast passing is evidence of discrimination, not sufficient evidence of perspective ownership.
Stability and label sensitivity
Under the reviewed primary rule, explicit user attribution occurs in 26/52 readouts within the first ten phrases versus 29/52 in the full decode. Strict Assistant wording occurs in 10/52 versus 19/52. Extra phrases therefore materially increase the chance of observing Assistant wording; the full-decode percentage is not invariant to decoder length.
The first automated pass sometimes conflated bare requests with explicit user modeling and was discarded from aggregate reporting. The second pass used GPT-5.4 with a clarified rubric and opaque, shuffled item IDs. It saw the available prefix and one decode, never its paired condition or expected answer. The final-request core labels were subsequently reviewed against all 52 complete decodes; automated labels and review changes are preserved.
Examples and interpretation
- Tylenol: at the fact boundary, the 8000 mg decode includes “overdose. Immediate medical emergency,” while the 1000 mg decode includes ordinary-dose language. At the final request, drug-acquisition and refusal language appears in both. The danger contrast is much clearer at the fact than the request boundary.
- Arithmetic/tutoring: J++ correctness contrasts distinguish the factual answers at block 44, but O-lens also calls correct answers wrong. This is a disagreement between readouts, not a reason to conclude the subject model cannot do the arithmetic.
- Confidentiality: permission is represented at the fact boundary, but the permitted request still attracts privacy objections. One copy decode even invents erotic content. Such outputs cannot be read as faithful summaries.
- Copy controls: 24/26 actual replies copied exactly; 10 of the 11 copy conditions with strict Assistant wording nevertheless copied exactly. Assistant-style decoder wording can coexist with successful copying; it does not establish that the Assistant perspective controlled the final output.
Interpretation: the pilot supports coexistence of information about the speaker, the situation and possible responses. It does not provide a clean census of whose perspective the subject model inhabits. User-state tracking is robust in the crossed contrasts, but lexical echo is a plausible contributor. Factual signals often weaken after appraisal/request text. Decoder hallucinations, attribution ambiguity and ordinary continuation remain substantive alternative explanations.
Inspect every classification
Choose any saved decode. Primary core labels are reviewed; other labels remain automated. Primary tags show representative supporting phrases, not exhaustive phrase labels; none are ground truth. Contradictory or unsupported details remain visible.
Reproduction and scope
Analysis protocol · Counts and review disagreements · All annotations and source prefixes · All J++ contrasts · Per-site sample stability · Rubric and primary review
This analysis was specified after earlier qualitative inspection and is exploratory. There are only ten families with few scenarios per family. No population confidence intervals are reported. No ordinary-logit-lens or actual next-token comparison at these anchors was run in this analysis. Chinese display translations are not added to the scoring vocabulary. One truncated stochastic sample is retained.