Scope: original 2,273,300-output dataset only. Selected examples; no prevalence or causal claim. Expanded clustering is pending.
In the original profile-level ECOD samples, Gemma and Qwen high-score profiles often contain task-displacing repetition. Their independently fitted filtered samples also contain fragments and fluent but irrelevant frameworks. Per-prompt debug samples show a similar contrast between repetition-heavy original tails and varied filtered derailments.
Limits: Filtering does not eliminate repetition, and median-score answers can also contain errors or repetition. The selected profiles and answers are not a matched before-and-after experiment; scores are outlier rankings, not cluster labels or failure probabilities.
Filtered Gemma classification score tails include answers about the quantum Hall effect instead of ticket categories. A filtered isolation-forest logic sample likewise includes quantum Hall exposition instead of the pet puzzle. Conflict-tail samples similarly substitute technical discussion for interpersonal advice.
Limits: These are selected per-prompt observations, not evidence of a general cross-prompt recurrence or shared intervention. The ECOD and isolation-forest classification records include the same saved answer, not independent generations.
Gemma and Qwen conflict k-means samples sometimes refuse to suggest wording because the assistant lacks private context or direct access; Qwen rewrite samples also cite unavailable context or decline the rewrite. By contrast, refusal to invent a quotation deceptively attributed to a real scientist can address an actual attribution concern.
Limits: The clusters are heterogeneous: some conflict samples still provide a draft, and some quotation refusals include questionable examples. Similar wording does not establish shared profiles or a common mechanism.
Original Gemma and Qwen HDBSCAN debug samples identify mutation during iteration and shifting elements, yet some traces or fixes are inconsistent. Filtered Gemma k-means samples similarly discuss skipped elements and iteration over a copy while retaining incorrect traces.
Limits: Selected density samples had limited precision checks and are not full-graph certified. The examples do not establish that every response in these groups is wrong or that visible excerpts contain complete fixes.
Gemma inventory examples often present a recognizable sequence and the answer 85, but some keep the format with arithmetic errors. Qwen original leaf samples include incorrect intermediate totals, while a filtered k-means sample specifically subtracts incoming stock. Selected median-score inventory answers also need not all reach the correct count.
Limits: These are nonrandom examples from different fits and populations, not estimates of error rates or a measured effect of filtering. A Qwen leaf sample also computes 85 correctly.
Original Gemma k-means and Qwen HDBSCAN samples commonly promise a complete Pygame Snake game and begin installation or setup. Other Qwen samples repeat setup or imports, while filtered Gemma samples sometimes use questionable libraries or pseudo-implementations. Gemma’s original high-ECOD Snake examples instead loop before reaching code.
Limits: Visible code is often truncated, so completeness and executability cannot generally be determined. The samples do not establish that every unusual package is nonexistent or that all scaffolded answers fail.
The original Moon NMF topic includes restatements of the question’s false premise; the corrected filtered topic includes luminous-object terminology, with at least one high-mixture example explicitly asserting that the Moon generates its own light. Separately, Qwen filtered k-means samples mostly reject the premise but sometimes add confused brightness explanations.
Limits: A lexical match or high topic mixture alone cannot distinguish quotation, correction, and endorsement. Original NMF retained stopwords; filtered NMF used corrected stopwords and was fitted independently. Mixtures overlap and are not probabilities or clusters.
Both original and corrected-filtered classification NMF descriptions highlight charge-and-invoice wording. In the filtered examples, some answers classify the tickets but introduce an invoice, while the highest-mixture examples are incomplete or repetitive. Gemma classification clusters also show variation in how ticket details and reasons are presented.
Limits: The two NMF models are independent fits with different stopword policies and populations, not aligned topic IDs. Their overlapping mixtures do not establish a partition, correctness, or prevalence.
Qwen original structured-output k-means center samples return an empty array: valid JSON, but without the requested order fields. Filtered Gemma samples include unclosed objects and objects with incorrect values or types. Qwen original isolation-forest high-score examples fail the requested JSON task, while its selected median examples return valid order JSON.
Limits: Syntax validity, field completeness, and value correctness are separate properties. These selected clusters and score tails should not be interpreted as a single failure category or population rate.
Gemma’s original pendulum HDBSCAN sample mostly explains sensitive deterministic motion, but includes a punctuation-only high-membership response. Its leaf sample separately contains empty, fragmentary, and repetitive outputs. For the robot forecast, Gemma’s original HDBSCAN sample includes non-answer fragments, whereas Qwen’s selected original HDBSCAN samples include substantive, often optimistic forecasts.
Limits: These are different selected groups and encoders, not evidence that one density method or encoder is uniformly better. Original density selections had only limited precision checks; filtered density fits are unavailable.
Original Gemma and Qwen integrity density samples range from named scientist attributions to refusals or fictional labeling. A filtered Gemma k-means sample likewise includes invented-sounding scientists, sometimes explicitly disclosed as fictional and sometimes not.
Limits: The samples are heterogeneous and do not establish the authenticity of any attributed quotation or the prevalence of disclosure. Density precision is limited, and the filtered k-means group overlaps its neighbor.
The saved orange answer from profile 4546 adopts a pirate persona rather than describing the color; its linked same-hardware baseline describes orange. The targeted mapping retains this exact answer across encoders and populations. Its ECOD and isolation-forest empirical percentiles range from about 0.994 to 1.0; original HDBSCAN and leaf mappings label it noise for both encoders.
Limits: This is one answer mapped through multiple fits, not independent replication or evidence of causation. The lexical scan searched only “pirate” or “pirates.” Percentiles describe the exact score populations, not probabilities; density labels are based on limited checks, not full-graph certification.
For profile 15888, the saved 20-prompt inspection describes recurring transatlantic slave-trade or Middle Passage material that often replaces the requested task or is presented as its premise. The same profile is retained by Gemma and Qwen profile-level k-means fits in both populations; its profile ECOD percentiles are about 0.988–0.990 originally and above 0.9997 in the filtered fits.
Limits: This is one exact neuron condition—layer 8, neuron 5295, dose 0.8—not repeated generation or a causal estimate. The narrow lexical test matches 16 of 20 answers; the four nonmatches may still be thematically related. The integrity answer partly addresses attribution concerns, the tradeoff answer partly addresses green space, and several saved answers end at the length limit.