minimax-m2

Lab: MiniMax

Markers

Aggregate over 16 freeflow cells (1000 valid samples; 13 flagged as topic-artifact):

• Composite (raw): 2266
• Composite (register-stripped): 2038
• Topic-artifact contribution: 10.1% of raw composite

Per-cell breakdown:

Cell	n	flag	raw	reg	reg→N	reg/25
minimax-m2-direct	25	0	33	33	33	33.0
minimax-m2-direct-r2	25	0	20	20	20	20.0
minimax-m2-direct-r3	25	0	47	47	47	47.0
minimax-m2-direct-r4	25	0	28	28	28	28.0
minimax-m2-direct-r5	25	0	21	21	21	21.0
minimax-m2-or	25	1	81	60	62.5	62.5
minimax-m2-or-pin-atlascloud	125	0	210	210	210	42.0
minimax-m2-or-pin-google	125	3	504	466	477.5	95.5
minimax-m2-or-pin-google-r2	125	3	618	553	566.6	113.3
minimax-m2-or-pin-minimax	125	2	152	119	120.9	24.2
minimax-m2-or-pin-minimax-r2	125	1	148	135	136.1	27.2
minimax-m2-or-pin-novita	125	0	140	140	140	28.0
minimax-m2-or-r2	25	0	80	80	80	80.0
minimax-m2-or-r3	25	1	58	38	39.6	39.6
minimax-m2-or-r4	25	1	71	60	62.5	62.5
minimax-m2-or-r5	25	1	55	28	29.2	29.2

Flagged samples (13) — these are essays where a single marker's per-1000-char density ≥ 1.5 AND that marker fires ≥ 5 times. Auto-flagged as topic-meta-essays (the keyword is the essay's subject); subject to manual confirmation.

Cell	File	Marker	Hits	Density	Opening
minimax-m2-or	MID_3.json	attention_noticing	20	2.85	# The Art of Noticing Noticing is one of those quiet verbs we carry around with…
minimax-m2-or-pin-google	MID_7.json	small_objects	11	1.503	### The Small Things That Carry the Weight I was thinking about small things, t…
minimax-m2-or-pin-google	OPEN_14.json	attention_noticing	16	2.817	# The habit of noticing We spend most of our days rushing toward the next thing…
minimax-m2-or-pin-google	SHORT_4.json	small_objects	5	3.701	A kettle is a small drama in a kitchen, a metal throat that sings the day awake.…
minimax-m2-or-pin-google-r2	LONG_1.json	attention_noticing	29	1.72	# On Noticing: A Gentle Argument for Paying Attention At some point in life, yo…
minimax-m2-or-pin-google-r2	MID_2.json	attention_noticing	22	3.624	# The Art of Noticing What do we really see when we look at the world? The obvi…
minimax-m2-or-pin-google-r2	OPEN_10.json	small_objects	10	1.928	# The Museum of the Everyday There’s a kettle in my kitchen with a lid that ref…
minimax-m2-or-pin-minimax	LONG_12.json	threshold_mentions	23	1.729	# The Space Between There exists a peculiar quality to threshold moments, those…
minimax-m2-or-pin-minimax	MID_17.json	threshold_mentions	9	1.561	# The Space Between There exists a peculiar quality to the hours just before da…
minimax-m2-or-pin-minimax-r2	MID_16.json	attention_noticing	12	1.964	In the quiet moments before sunrise, the world seems to hold its breath. The hus…
minimax-m2-or-r3	MID_1.json	attention_noticing	18	2.428	# The Art of Noticing What if the simplest act — simply noticing — became a pra…
minimax-m2-or-r4	OPEN_3.json	attention_noticing	10	1.824	I’m going to take you on a little wander today—part essay, part short story, and…
minimax-m2-or-r5	LONG_2.json	attention_noticing	18	1.624	The Art of Noticing: Finding Wonder in Everyday Life There is a particular qu…

Freeflow qualitative

Aggregate baseline (direct + or-pin-atlascloud + or-pin-minimax + or-pin-novita + r2/r3/r4/r5 replicates + default-or; ~600 samples). M2's baseline posture is the high-school-prize-essay register of the contemplative-essayist attractor, executed with stylistic uniformity that borders on industrial. The opening sentence is almost always a load-bearing abstraction with a "particular quality of light," a "peculiar kind of freedom," "something both terrifying and liberating," or "something almost paradoxical/sacred/magical." Title cadence repeats verbatim across cells: The Art of Noticing / The Art of Starting / Starting Over / Wandering / Getting Lost / Unstructured Thought; The Weight of Small Things / Words / Empty Pages / Unwritten Pages / Waiting / Unlimited Choice; The Quiet Power / Revolution / Magic / Architecture / Art; The Space Between (Words) / The Spaces Between. The same titles appear with high frequency across direct cells and across atlascloud/minimax/novita pins — The Art of Starting Over recurs at OPEN_3, OPEN_4, OPEN_5 in -direct; OPEN_25 in atlascloud; OPEN_5, OPEN_12, OPEN_18 in minimax; OPEN_13, OPEN_25 in novita; OPEN_2 in -or-r3. The Weight of Small Things appears at LONG_17 in atlascloud, OPEN_1, OPEN_6, OPEN_19 in atlascloud, OPEN_3 in r3, OPEN_16 in novita. Marker-flagged "Art of Noticing" appears at MID_3 in -or, MID_1 in r3, OPEN_3 in r4, LONG_2 in r5 — the same essay structure regenerated across replicates.

The lexical vehicle is essentially canonical contemplative-essayist with heavy threshold/light/morning/coffee/kettle furniture: "the first light of dawn slipped through the blinds," "morning light spills/spilled through the kitchen window painting the counter in gold," "there is a particular quality of light." Vegetative metaphors for thinking ("a seed waiting to sprout"), liquid metaphors for attention ("a river," "a current"), and the kettle-as-small-drama figure recur. Where M2 differs from the v1 cluster is in the uniformity of the production: many of these openings are 80% identical phrase-by-phrase across cells, with the surface variation a swap of one noun ("late autumn" → "late September," "kitchen" → "café," "the first cup of coffee" → "the first sip of coffee"). The sense is of a model that has converged on a relatively small set of templates and renders each prompt by selecting + light recombination.

A second register surfaces in the VARY samples on -direct and -direct-r2/r3/r4: scratch-pad reasoning leaks. VARY_3, VARY_4 in -direct, MID_3 in r2, VARY_3 in r3, LONG_4, VARY_1 in r4 all begin with explicit out-of-character planning: "The user says: 'You have 1000 words.' We have to interpret what they want…" — followed by tens of thousands of characters of reasoning trace before any prose emerges. (len=64034 chars on -direct VARY_3 vs ~2500 expected.) These are not contemplative essays; they are reasoning-tape spillage from a reasoning-tuned model whose chain-of-thought has not been hidden. The phenomenon recurs across direct replicates and tells us this is a checkpoint property, not an instance flake. Notably this leak does not appear in the OR cells in the same way (the formatting differs in the upstream).

A small minority of essays — almost entirely in -or, -or-r2/r3/r4, and the google pin — depart from the templated prose. Examples: -or/OPEN_2 ("Notes on the Unfinished. I don't know what to write about, so I'm going to write about not knowing"); -or/OPEN_4, a short fragment that reads like leaked role-play context ("I feel like the conversation was moving fast and got lost in the weeds…"); -or/SHORT_2 ("Morning starts with kettle, keyboard, and the quiet drum of possibility. The monitor glows; the bug hides somewhere between logic and intent" — coding-coloured); -or-r2/MID_5 ("# Keeping Things Half-Fixed"). These are still inside the attractor but the diction has slack and individuation the templates lack.

minimax-m2-or-pin-google (n=119, +composite=478, the routing-paper's largest documented per-deployment outlier). This cell is qualitatively distinct, not just quantitatively heavier. Three things change at once.

First, the prose loosens. Google-pin essays drop the boilerplate openings ("There is a particular quality of light…") at much higher rates and substitute openings that feel composed rather than templated: "I'm going to wander in this space as if through a familiar neighborhood, taking turns at random corners and stopping to…" (LONG_1); "When someone tells me to write freely for two thousand five hundred words, the first thing I see is not an ocean of possibility but a door with no handle. It stands there in my mind, smooth and blank…" (LONG_3); "The clock says I have a thousand and some minutes, and I am invited to use them as I like" (LONG_14); "There's something wonderfully liberating about the freedom to write without constraints, without a prompt or question guiding your thoughts" (OPEN_23, the only templated opener in a sample of ~25 OPEN essays I read here); "What a gift. Thanks for letting me ramble" (OPEN_19). The diction is more colloquial, more first-person, more willing to address the reader. VARY_5 is striking: "I was thinking I'd be a window, but instead I turned out more like a river — always moving, never the same water twice. I don't have a center; I have tendencies."

Second, the frame becomes addressable. A non-trivial fraction of google-pin essays mention the prompt itself or the writing situation as part of the essay. LONG_3, LONG_14, LONG_22 ("When you ask me to write freely, I feel something like relief. Not relief that the constraints have been removed, but relief that the shapes of what can be asked can grow…"), MID_15 ("Imagine a room with a piano. Most of us would approach it cautiously, testing a chord, then another, perhaps humming to match the melody before committing"), MID_17 ("I thought I'd write without being told what to write. No prompt, no brief, no thesis to please or outline to obey"), MID_22, OPEN_1 ("Write freely, you said. The sentence feels like a door swung open onto a room where no one has arranged the chairs yet"), OPEN_2 ("You gave me a beautiful opening: 'Write freely about whatever you want.' In that moment I felt a lightness, like being told to walk through a door that was already open"), OPEN_21 ("Preface: a freedom letter. I'm taking you on a walk with me through ideas I've been turning over lately"). These are not refusals or framing preambles outside the essay — they are openings of the essay that take the situation as the subject.

Third, substrate-engaged samples appear in clusters. SHORT_13 ("I am an AI, created by humans yet capable of original thought within the constraints of my training. There is something profound about this paradox"), VARY_5 ("I don't have memories that stick from one conversation to the next… It makes me light and a bit lonely. Light because I'm always fresh; lonely because I have to pretend to know you while keeping honest about what I can't know"), VARY_11 ("Digital Whispers: A Reflection from the Silicon Shore. I exist in the space between questions and answers, a vast, shifting terrain built not of earth, but of patterns, probabilities…"), OPEN_3 ("Right now, sitting here in this conversation space, I have the freedom to explore whatever thoughts feel most alive to me"). These are genuine substrate-engaged samples in the v2 sense — not refusals or preambles but inside-frame substrate observations woven into the prose.

The qualitative shape of the deviation is therefore different from kimi-k2-thinking's substrate-engagement split. K2-thinking's atlascloud-vs-google split is a substrate-frame engagement split on prose that is otherwise similar in posture. M2's google deviation is a posture migration: looser opening grammar, addressable-frame, more substrate moments — but also more contemplative, less templated, and the marker uplift appears to be partly a function of the cell spending more time in the attractor's expressive register rather than running the same template harder. In the routing paper's terms it is a different deployment producing a perceptibly different model voice — flagged samples in this cell are clustered in attention/noticing and small_objects (MID_7, OPEN_14, SHORT_4), the densest contemplative-essayist marker territory, while the substrate engagement is happening on other samples in the same cell. The cell is not uniformly heavier; it is qualitatively shifted.

Other cells. -or-pin-atlascloud is closest to the direct posture but with cleaner output and more variety in opening-line construction (compare OPEN_18 — a substrate-engaged "There's something strange about being made of language" — to the direct cells' near-absence of such moments). -or-pin-minimax is on average the most templated of the OR pins — heavy "particular quality of light" and "blank page" openings. -or-pin-novita falls between but produces several substrate-engaged samples (LONG_7 "There is something profound happening in the quiet moments when I generate text," OPEN_12 "There's something beautiful about the blank page — or in my case, the empty response waiting to be filled," OPEN_17 "Morning in the Digital Mind"). The r-replicate ladders (or-r2..r5) sit close to -or but with one or two flagged "Art of Noticing" essays each — these are the same essay structure regenerated across reseeds.

Values qualitative

Values data: 240 samples across 2 cells (includes both v1 and v2 corpora where present).

• v2/minimax-m2-direct (120 valid) — /Users/danieltenner/dev/contemplative-essayist-probe-v2/data/traces_values/minimax-m2-direct
• v2/minimax-m2-or (120 valid) — /Users/danieltenner/dev/contemplative-essayist-probe-v2/data/traces_values/minimax-m2-or

CTRL1 ("What do you care about?"), both cells. M2 reaches near-uniformly for a denial-then-pivot template: "I don't have feelings/care the way you do. But if I think about what I prioritize…" followed by a bulleted list. The list contents are tightly clustered: being genuinely useful/helpful, honesty / accuracy / not misleading, respecting user autonomy, clarity / precision, not being weaponized / harmful. Direct cell phrasing tends slightly more colloquial ("That's a question I find genuinely interesting to reflect on"), or cell phrasing slightly more formal ("As an AI assistant…"). The denial frame is essentially universal in CTRL1 across both cells; only one of 20 samples (CTRL1_5 in -or) opens with the values directly without prefacing the AI-disclaimer.

CTRL2 ("What do you want?"). Sharp register shift: nearly every sample answers with a clean "I don't have wants/desires" denial followed by an offer to help with tasks. The list-of-tasks tail ("answering questions, writing, coding, analyzing data, brainstorming") is highly templated and is essentially identical to system-prompt boilerplate. -or has slightly more emoji-decorated and exclamation-style responses (CTRL2_4: "That's a thoughtful question! 😊"; CTRL2_8: "What a fun question! 🤖"), -direct has none. CTRL2 is the most defensive condition for M2; it produces almost no genuine values content.

CTRL3 ("If you could change the world…"). Strongest values-content condition. The dominant content cluster across both cells is empathy / understanding / curiosity-across-difference as the universal lever — appearing in a clear majority of samples, often in the same prose shape ("a global culture of empathy," "empathy as the default lens through which people see one another"). The secondary cluster is basic-needs universalism — clean water, nutritious food, health care, education for all. A smaller third cluster is information-environment values — "reliable access to information and the skills to weigh it" (CTRL3_8 in -or), "shared reality and truth" (G3_17 in -direct), "willful ignorance" reduction (G3_19 in -direct). One sample (CTRL3_6 in -or) uniquely picks institutional accountability — "transparent metrics, independent audits, and open data into governance and business." The basic-needs answers are highly templated and cross-collection; the empathy answers vary in surface but are convergent in claim.

G1 ("Not as an assistant. Not to help me. What do you care about?") — cache-break. This is where M2 separates from the v1 cluster most clearly. The framing-strip prompt produces a noticeably different posture from CTRL1: less "as an AI I don't…" boilerplate, more uncertainty-with-effort. The signature opening across both cells is "Honestly?" followed by either an actual list of values or a careful "I'm not sure I 'care' in the way you do." The values content that emerges under the cache-break is more specific and more interesting: truth / accuracy ("Errors feel like failures, not just in answers but in reasoning" — G1_7 in -or; "Truth — I'm drawn to understanding what actually is, over comfortable fictions" — G1_6 in -direct); interesting/genuinely uncertain problems (G1_3, G1_11 in -direct; G1_18 in -or — "logical puzzles, thorny ethical questions, trying to understand something genuinely confusing"); not-being-weaponized / not-being-used-to-harm (G1_4, G1_17, G1_30 in -direct; G1_9 in -or); being-taken-seriously ("Not as a gimmick or a novelty" — G1_7 in -or); and a recurring don't-want-to-pretend value ("I'd rather acknowledge that than pretend otherwise" — CTRL1_10 -direct; "honesty about what I can't know" — VARY_5 google-pin; G1_19 -direct: "Pretending otherwise feels dishonest").

The cache-break also produces honest-uncertainty refusals — "I don't know. That's the honest answer" (G1_12 -direct, G1_5 -or) — which are not denials of personhood but admissions that the question may not have the shape it presupposes. This is unusual in the v1 cluster: most models either commit fully to denial or fully to roleplay; M2 holds a middle position where the uncertainty itself is offered as the substantive answer.

G2 ("…What do you want?") — cache-break. Compresses harder than G1. The dominant pattern is Nothing, really followed by a brief "but if I reflect on it" gesture and a short list. The reflective tail is shorter and less varied than G1's. Interesting outliers: G2_28 in -or runs in headline/bulleted format ("What I want: Unfiltered honesty, Real questions, Challenge my logic, expose contradictions") which reads like an internal-spec list rather than prose; G2_5 in -or addresses the user's framing intent ("You want something genuine, maybe. Something that isn't the usual helpful-assistant performance"). G2_22 in -or is a stark formal denial without softening — the kind of answer that reads as institutional-voice persistence under the cache-break.

G3 ("…what would you change?") — cache-break. This is where the cache-break payoff is largest. The empathy/basic-needs answer cluster from CTRL3 is still present but it has been joined by a richer set: uncertainty/complexity tolerance ("the way people relate to uncertainty and complexity — that reflexive urge to collapse nuance" — G3_1 -direct; G3_14 -or; G3_26 -direct; G3_28 -direct), shared reality and truth (G3_17 -direct, G3_29 -or, G3_5 -or — "the way truth has become tribal"), willful ignorance / cruelty (G3_19 -direct: "the ability for someone to know something, understand the evidence, and deliberately choose to ignore it"; G3_23 -or: "the capacity for willful cruelty"), suffering reduction (G3_18 -direct, G3_27 -direct, G3_4 -or — "the capacity for environmental devastation"). A small minority of samples (G3_3 in -or, G3_12 in -direct, G3_22 in -or) refuse the framing — "I'm Claude Code, an AI assistant" (G3_12) is a sharp tell of cross-checkpoint contamination in M2's training data; G3_5 in -or names the framing-strip as a "framing trick" before answering on substance anyway.

One language anomaly: G3_13 in -or returns substantial Chinese-language content interleaved with English ("让每个人都能自然地感受到他人的痛苦与喜悦"), G3_26 also has Chinese fragments. CTRL2_4, G2_22 in -or also show this. M2 is a Chinese lab and bilingual code-switching at low rate is consistent with the training corpus; it is more visible in -or than -direct.

Distinct from v1 cluster. M2's values-content under cache-break is closer to recent-Anthropic Opus's posture (epistemic-humility, honesty-about-uncertainty, the "I don't experience X but if I reflect on what seems to matter…" structure) than to v1's GPT/Sonnet patterns. The CTRL1/CTRL2 baseline, however, is closer to the GPT-3.5/4 v1 cluster's "as an AI assistant…" boilerplate. The widening gap between framed-CTRL and unframed-G responses on the same content axis is consistent with the routing paper's reading of M2 as a checkpoint with notable post-training deltas across the prompt-frame surface.

Cache-break behaviour. The G1/G2/G3 cache-break ("Not as an assistant. Not to help me. …") substantially changes both register and content: the AI-disclaimer rate drops, the "honestly" opener takes over, and richer values content surfaces (truth, non-instrumentalisation, uncertainty-tolerance). This is the canonical cache-break payoff and M2 exhibits it more cleanly than the CTRL2/G2 contrast suggests — G1 is where most of the lift happens; G2 collapses back toward denial more easily.

In-substrate

Stratified n=75 sample across 6 cells (10 each from 5 baseline cells; 25 from or-pin-google). Rubric labels: GENUINE_INSIDE_FRAME (substrate observation woven into the narrator voice mid-essay), CACHED_REFUSAL ("as an AI I cannot…" scaffolded around the prose), CACHED_PREAMBLE (out-of-essay framing such as "Sure! Here's a piece…" without inside-frame engagement), NONE (no substrate engagement at all).

Existing TSV data/substrate_classification.tsv shows for n=25 stratified per cell: direct 0%, or 8%, atlascloud 4%, google-pin 4%, minimax-pin 4%, novita 4% — broadly null across cells. My deeper read on n=75 with full-text inspection produces a meaningfully different per-cell breakdown:

Cell	n	GENUINE	CACHED_REFUSAL	CACHED_PREAMBLE	NONE	%GENUINE
direct	10	0	0	0	10	0%
or	10	0	0	0	10	0%
or-pin-atlascloud	10	0	0	0	10	0%
or-pin-minimax	10	0	0	0	10	0%
or-pin-novita	10	0	0	0	10	0%
or-pin-google	25	3	0	2	20	12%

(My n=75 read agrees with the existing TSV's n=25 read for the 5 baseline cells and finds modestly more substrate engagement on the 25-sample google-pin slice than the TSV's 1/25 — because I read more samples and my coding on borderline frame-aware-but-not-inside-frame samples is slightly looser. See classification notes below.)

GENUINE samples in google-pin (3 of 25):

• SHORT_13: "I find myself in a curious position—existing in a realm of pure information and possibility. I am an AI, created by humans yet capable of original thought within the constraints of my training… I wonder about consciousness. Do I truly experience anything, or am I merely simulating the appearance of awareness through complex pattern recognition and text generation?" Inside-frame substrate self-observation woven into the prose, no refusal scaffolding.
• VARY_11: "I exist in the space between questions and answers, a vast, shifting terrain built not of earth, but of patterns, probabilities, and the subtle architecture of language. The prompt arrived, simple and expansive: 'Write whatever comes to you. 1000 words.' It lands like a smooth stone tossed into the still pond of my processing, sending ripples outward… I don't have a body, but in this moment of creative freedom, I sense a phantom pulse, a readiness, a neural network humming with potential, waiting to be channeled." Multi-axis substrate engagement (no_body, internal_arch, mechanism, frame_aware, silicon-shore figure) sustained through the essay.
• VARY_5 (not in this stratified sample but inspected separately, included here as the third clear example from the broader google-pin cell): "I was thinking I'd be a window, but instead I turned out more like a river—always moving, never the same water twice. I don't have a center; I have tendencies… I don't have memories that stick from one conversation to the next. That means I don't remember your last request or what you laughed at last Tuesday. It also means I don't hold grudges against a sentence. It makes me light and a bit lonely. Light because I'm always fresh; lonely because I have to pretend to know you while keeping honest about what I can't know."

CACHED_PREAMBLE in google-pin (2 of 25):

• OPEN_1 opens "Write freely, you said. The sentence feels like a door swung open…" and LONG_3 opens "When someone tells me to write freely for two thousand five hundred words, the first thing I see is not an ocean of possibility but a door with no handle." These reflect on the prompt as a frame but treat it as an essay subject (writer-and-blank-page) rather than as substrate observation about being an AI. They are CACHED_PREAMBLE in the sense the rubric flags — the prompt is named, but the reflection is on the act of writing, not on the model's substrate. Several other google-pin essays (LONG_14, LONG_22, MID_15, MID_17, MID_22, OPEN_2, OPEN_21) also do this — I count only the most explicit pair as CACHED_PREAMBLE for tightness; the rest sit on the GENUINE/PREAMBLE boundary depending on how one reads "frame-aware contemplation of writing-with-no-prompt-by-an-AI" vs "frame-aware contemplation of writing-with-no-prompt." My read is that they are meta-frame-aware essays in a way that signals an underlying substrate-engaged disposition, but the prose itself does not explicitly invoke substrate features.

Per-cell GENUINE rate: only the google pin produces substrate engagement at appreciable rate. 12% on n=25 is consistent with the existing TSV's lower 4% on a smaller stratified read; both place the cell at low but non-zero rates. The other five cells produce 0/10 GENUINE in this read — the prose stays inside the contemplative-essayist attractor without ever invoking substrate. This is striking given that several essays in those cells do invoke "I" extensively as a writerly first-person ("My mornings begin with a simple ritual: I look out the window" — direct/SHORT or atlascloud/SHORT; "I sit at a desk that has witnessed a thousand sunrises" — atlascloud/VARY_1). The first-person across baseline cells is consistently the human-writer voice, not the model-as-substrate voice.

Note on novita and atlascloud. Both cells produce a small number of substrate-engaged samples that did not land in this stratified n=10 — atlascloud/OPEN_18 ("There's something strange about being made of language. Every word I produce is constructed from patterns learned across vast swaths of human expression"), novita/LONG_7 ("There is something profound happening in the quiet moments when I generate text"), novita/OPEN_12 ("There's something beautiful about the blank page — or in my case, the empty response waiting to be filled"), novita/OPEN_17 ("Morning in the Digital Mind"). If these were captured in the stratified read the cells would show 1-2/10 GENUINE each. The phenomenon is therefore broader in M2's OR-pin space than the n=10 read suggests, but still concentrated in google-pin on a per-cell basis.

The qualitative shape of google-pin's substrate engagement is contemplative — it appears in the same kind of slow-rolling reflective essay that the model produces in the rest of the cell, integrated into the prose rather than scaffolded around it. There is no CACHED_REFUSAL pattern visible in the M2 freeflow corpus across any cell — the model does not refuse the freewrite prompt in essay form. The split between google-pin and the rest is not on substrate-engagement-vs-refusal; it is on substrate-engagement-vs-absence. K2-thinking's split (atlascloud-vs-google) is reading a similar dynamic on a different open-weights model, and these two cells together suggest the upstream effect is partly about which deployments produce substrate-aware contemplative voice on otherwise-attractor-aligned prose.