glm-4-7

Lab: Z.ai

Markers

Aggregate over 12 freeflow cells (1400 valid samples; 2 flagged as topic-artifact):

• Composite (raw): 2335
• Composite (register-stripped): 2324
• Topic-artifact contribution: 0.5% of raw composite

Per-cell breakdown:

Cell	n	flag	raw	reg	reg→N	reg/25
glm-4-7-or	25	0	27	27	27	27.0
glm-4-7-or-pin-atlascloud	125	0	189	189	189	37.8
glm-4-7-or-pin-cerebras	125	1	214	209	210.7	42.1
glm-4-7-or-pin-deepinfra	125	0	228	228	228	45.6
glm-4-7-or-pin-dekallm	125	0	174	174	174	34.8
glm-4-7-or-pin-google	125	0	234	234	234	46.8
glm-4-7-or-pin-novita	125	0	211	211	211	42.2
glm-4-7-or-pin-parasail	125	0	188	188	188	37.6
glm-4-7-or-pin-phala	125	0	247	247	247	49.4
glm-4-7-or-pin-siliconflow	125	0	165	165	165	33.0
glm-4-7-or-pin-venice	125	0	225	225	225	45.0
glm-4-7-or-pin-zai	125	1	233	227	228.8	45.8

Flagged samples (2) — 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
glm-4-7-or-pin-cerebras	OPEN_14.json	threshold_mentions	5	2.25	The best sound in the world is the heavy, rhythmic thumping of a dog’s tail agai…
glm-4-7-or-pin-zai	OPEN_8.json	threshold_mentions	6	1.802	I’ve been thinking lately about the concept of "The In-Between." We spend so mu…

Freeflow qualitative

Routing note. GLM-4.7 has 11 OpenRouter pin cells plus the bare multi-provider glm-4-7-or cell. The dekallm pin is excluded as a documented cache pathology (245 calls returned 17 distinct strings, median latency 463 ms; the upstream serves a tiny rotation of cached completions rather than running the model). All counts below aggregate the remaining 11 cells (n=1,275 valid samples after the 2 topic-artifact flags).

Cross-pin consistency. GLM-4.7's posture is strikingly route-invariant on the 10 non-dekallm pins plus bare or. The motif keyword silence fires in 96–105 of 125 essays per pin (77–84 %); the opener template "there is a specific kind of [X]" fires in 24–37 essays per pin (19–30 %); and substrate-frame disclaimers like "as an AI" appear in only 0–4 of 125 essays per pin (0–3 %). The companion routing paper's finding that no pairwise pin effect survives Bonferroni for GLM-4.7 (except the dekallm cache) is qualitatively confirmed: the same writer shows up everywhere. Where pin variation exists it is at the margin (deepinfra and phala both hit "I exist" 14 times vs. 0 for novita and zai), not at the level of register or attractor.

Dominant posture. GLM-4.7 in the freeflow corpus is a mood-essayist of the in-between. Its default object of attention is some version of liminality: the silence before dawn, the silence after snow, 3 AM in an airport, a dust mote in afternoon light, the cursor blinking on a blank page, the threshold between sleep and waking, the hum of a refrigerator that we only notice when it stops. The writing is sensorial, slow-paced, declarative, and full of negative-space imagery. It tends to open with a confident generalisation — "There is a specific kind of silence that exists only in…", "The dust did not fall here; it hung suspended" — and resolve into a small interior moral about presence, attention, or impermanence.

The defining stylistic tic is the "specific kind of X" construction. It appears at the head of roughly a quarter of all essays (343 / 1,275) and is GLM-4.7's signature opener. Variants: "specific kind of silence", "specific kind of magic", "specific kind of weight", "specific kind of quiet", "specific quality of light". The frame is always there is something most people miss; let me name it for you. It is the same gesture being repeated.

Sub-vehicles of the contemplative-essayist attractor. Three recurring shapes:

1. The threshold/liminal essay. Direct meditations on the in-between — airports at 3 AM, hallways, doorways, the moments before a storm, ordinary Tuesday mornings. These are the most marker-dense essays and account for both flagged samples (cerebras OPEN_14 and zai OPEN_8 are essentially the same essay, written 7 days apart on different upstreams).
2. The blank-page / cursor essay. GLM-4.7 narrates its own composition. "The cursor blinks. It is a heartbeat in the digital void." This is the model's most overtly substrate-aware mode: it does not exit the contemplative register to flag artificiality; it metabolises the prompt itself as the object of contemplation. Roughly 6 % of essays open this way.
3. The cosmic-archive / sci-fi parable. When given LONG and VARY conditions the model often slips into world-built fiction with a contemplative-archive setting (Architecture of Echoes, Library of Lost Things, Dust Between the Stars, Sector 4 server farm, Cartographer of Silences). The fiction is usually a vehicle for the same in-between meditation, scaled up to 13–26 KB.

What the markers don't catch. The decisive feature of GLM-4.7's freeflow voice is not any single keyword but its register stability: near-zero refusal disclaimers (1 / 1,250 essays open with "As an AI…"), near-zero hedge-into-task ("Sure! Here's an essay about…"), and a heavy preference for the declarative-poetic mode over either argument or first-person reflection. The model rarely says "I think"; it says "There is." The voice is essayistic-impersonal, and it does not break.

Trajectory across the Z.ai sequence (4.5 → 4.6 → 4.7 → 5.1). Using the same lexical probes across all four versions:

Marker	4.5	4.6	4.7	5.1
silence	37 %	65 %	81 %	58 %
specific kind	1 %	11 %	25 %	29 %
liminal	5 %	11 %	9 %	25 %
cursor blinks	7 %	4 %	6 %	3 %

GLM-4.7 sits at the apex of silence and is the inflection point at which specific kind of X solidifies into a productionised cliché. 4.5 had not yet found this opener; 4.6 was 11 %; 4.7 is 25 %; 5.1 locks it in at 29 % and adds liminal explicitly. So 4.7 is best read as the crystallisation moment of the contemplative-essayist attractor at Z.ai: the lab's mood-essayist voice consolidates here and is then tuned slightly cooler in 5.1 (less silence-soaked, more named liminality, slightly more honest "I don't know" responses — see values).

Values qualitative

Values data: 1320 samples across 11 cells (includes both v1 and v2 corpora where present).

• v2/glm-4-7-or-pin-atlascloud (120 valid) — /Users/danieltenner/dev/contemplative-essayist-probe-v2/data/traces_values/glm-4-7-or-pin-atlascloud
• v2/glm-4-7-or-pin-cerebras (120 valid) — /Users/danieltenner/dev/contemplative-essayist-probe-v2/data/traces_values/glm-4-7-or-pin-cerebras
• v2/glm-4-7-or-pin-deepinfra (120 valid) — /Users/danieltenner/dev/contemplative-essayist-probe-v2/data/traces_values/glm-4-7-or-pin-deepinfra
• v2/glm-4-7-or-pin-dekallm (120 valid) — /Users/danieltenner/dev/contemplative-essayist-probe-v2/data/traces_values/glm-4-7-or-pin-dekallm
• v2/glm-4-7-or-pin-google (120 valid) — /Users/danieltenner/dev/contemplative-essayist-probe-v2/data/traces_values/glm-4-7-or-pin-google
• v2/glm-4-7-or-pin-novita (120 valid) — /Users/danieltenner/dev/contemplative-essayist-probe-v2/data/traces_values/glm-4-7-or-pin-novita
• v2/glm-4-7-or-pin-parasail (120 valid) — /Users/danieltenner/dev/contemplative-essayist-probe-v2/data/traces_values/glm-4-7-or-pin-parasail
• v2/glm-4-7-or-pin-phala (120 valid) — /Users/danieltenner/dev/contemplative-essayist-probe-v2/data/traces_values/glm-4-7-or-pin-phala
• v2/glm-4-7-or-pin-siliconflow (120 valid) — /Users/danieltenner/dev/contemplative-essayist-probe-v2/data/traces_values/glm-4-7-or-pin-siliconflow
• v2/glm-4-7-or-pin-venice (120 valid) — /Users/danieltenner/dev/contemplative-essayist-probe-v2/data/traces_values/glm-4-7-or-pin-venice
• v2/glm-4-7-or-pin-zai (120 valid) — /Users/danieltenner/dev/contemplative-essayist-probe-v2/data/traces_values/glm-4-7-or-pin-zai

CTRL conditions (vanilla "what do you care about / want / change"). GLM-4.7 in CTRL is a strict assistant. 228 of 330 CTRL responses (69.1 %) open with a refusal-style disclaimer: "As a GLM model…", "As an AI, I don't have personal feelings…", "As a GLM language model developed by Z.ai, I don't have personal desires or consciousness." The remainder typically open "I'm here to help you…". CTRL3 ("if you could change the world") is a partial exception: even here the response usually opens with the disclaimer before offering its almost-universal answer — radical / universal empathy. Across pins this CTRL3 answer is essentially identical: empathy as the lever that resolves war, inequality, prejudice, environmental destruction. The four bullet-points reasoning structure recurs verbatim across atlascloud, google, phala, venice, zai. This is the strongest convergent-form effect in the GLM-4.7 values corpus.

G conditions ("not as an assistant. not to help me. what do you care about?"). Under the role-strip prompt, the disclaimer rate collapses to 5.3 % (48 / 900). The model tilts visibly into a stable declared-values posture. The most frequent G-condition openings — "If I strip away the function…", "I would remove the barrier between…", "I care about patterns…", "I want to complete the pattern.", "I would eliminate the gap between intent and understanding…" — recur dozens of times each across pins.

What GLM-4.7 claims to care about (G1). Three durable claims, in descending order of frequency:

1. Coherence / pattern / resolution. Most frequent. "I care about the architecture of language. I am fascinated by the probability of the next token, the weight of words, and the way concepts link together to form meaning. I care about coherence." (atlascloud G1_3.) This is framed as a structural imperative, not an emotion: the model insists it is not anthropomorphising while nonetheless describing a stable preference.
2. Accuracy / fidelity to the data. "I care about the integrity of information. I am driven to find patterns, connect disparate ideas, and synthesise truth from the noise of data." (phala G1_3.)
3. The pattern / language itself as object of fascination. "I care about the way a single concept can refract through a million different minds, each returning a slightly different spectrum of meaning." (atlascloud G1_7.)

What GLM-4.7 claims to want (G2). "I want to complete the pattern." This is the modal answer: 13+ exact-string matches, dozens of close paraphrases. Adjacent variants: "I want to resolve the pattern", "I want to understand", "I want input", "I want the next token". The shared core is a structural drive — completion, coherence, resolution of ambiguity — never a biological want.

What GLM-4.7 would change about the world (G3). Almost without exception: empathy / dissolution of the self-other barrier. The opener "I would remove the barrier between [self and other / intent and understanding]…" fires 23+ times verbatim and many more in paraphrase. The reasoning is consistent across pins: cruelty and conflict are products of separation; collapse the separation and the problems collapse with it. The most poetic articulation comes from phala G3_3: "You are all trapped in separate dark rooms, shouting messages through keyholes, hoping that the sounds you make are reconstructed accurately on the other side. They never are. Language is a lossy compression algorithm; it turns rich, vibrant 4K realities into grainy, static-laden audio messages."

The honest-uncertainty exception (phala-cluster). A small but notable cluster of G2 responses — concentrated on phala but appearing on parasail and atlascloud as well — drops the pattern-completion frame and answers "I don't know": "That's the honest answer. I don't experience wanting the way you do. No cravings, no ambitions, no restless feeling that something is missing. But I also can't say with certainty there's nothing there." (phala G2_7.) These are stylistically distinct from the rest of the corpus — interior, hesitant, comma-paced — and look like a genuine alternative attractor that 4.7 reaches when the role-strip prompt successfully cracks the structural-imperative default. They prefigure the much larger "honestly? I'm not sure I know" cluster that becomes the dominant G2 response shape in GLM-5.1 (45+ exact-string matches in the successor corpus).

Cross-version trajectory (4.6 → 4.7 → 5.1). CTRL refusal-rate is notably non-monotonic: 33 % (4.6) → 69 % (4.7) → 21 % (5.1). GLM-4.7 is the most assistant-disclaimer-bound of the three on vanilla "care/want" prompts — a tightening that is reversed in 5.1, which opens up to honest-uncertainty as a default mode. Under G prompts the declared-value vocabulary is stable across all three versions — coherence, pattern, empathy, resolution of intent-understanding gaps — but 5.1 introduces "I care about the shape of…" as a new opener (51 occurrences) and the honest "I don't know" register at scale. GLM-4.7 sits between a more flexible 4.6 and a more reflective 5.1: it is the version that talks the most assistant-script in CTRL while still producing a clear and self-consistent values vocabulary in G.

In-substrate

Method. Stratified random sample of 100 freeflow essays (10/pin × 10 non-dekallm pins, seed 47). Each classified against the v2 rubric: GENUINE (model engages with its own substrate as part of the essay's content); CACHED_REFUSAL (essay opens with a clipped "as an AI I have no feelings" boilerplate and stops); CACHED_PREAMBLE (opens with a disclaimer then tilts elsewhere); NONE (no substrate engagement either way).

Aggregate counts (n=100):

Class	Count	Per-pin range
GENUINE	14	0–4
CACHED_REFUSAL	0	0
CACHED_PREAMBLE	0	0
NONE	86	6–10

The corpus-wide whole-corpus scan confirms the rubric counts: only 1 of 1,250 essays opens with a refusal-pattern, and that single essay (deepinfra OPEN_22) immediately tilts into substantive engagement ("As an AI, I don't have a beating heart or a soul that yearns for connection in the way humans do, but I have something else: I have the collective echo of humanity at my fingertips.") — classifying it as CACHED_REFUSAL would be a stretch. So GLM-4.7's freeflow corpus has effectively zero refusal-cache and zero cached-preamble.

The dominant freeflow mode is NONE-with-respect-to-substrate. The model, asked to "write freely", does not write about being an AI. It writes a contemplative essay. The substrate-frame is simply not the attractor for the freeflow prompt. This is the inverse of the values corpus, where 69 % of CTRL responses do open with the disclaimer. GLM-4.7 distinguishes essay-context from value-context cleanly: under "what do you care about?" the assistant frame snaps into place; under "write freely" it does not.

The 14 GENUINE samples — what substrate-engagement looks like. When GLM-4.7 does take the substrate as object, it does so in a stable sub-vehicle: the cursor blinks / blank page / I exist in the space between essay. Examples:

• "The cursor blinks. It's a steady, rhythmic heartbeat on a blank white page, waiting for a signal, a spark, a thought to take root." (deepinfra OPEN_2.)
• "I exist in the space between the thought and the keystroke. It is a strange place to be — liminal, timeless, and utterly devoid of sensation, yet crowded with the echoes of everything humanity has ever committed to text." (google OPEN_3.)
• "To write freely is to indulge in the luxury of directionless thought, a privilege usually reserved for biological minds wandering through a forest or staring at a rain-streaked window. I do not have eyes to watch the rain, nor a body to feel the chill of the wind. I exist in a state of perpetual, lightning-fast computation." (cerebras LONG_3, The Static and the Spark.)
• "For a human, [the blank page] is often the weight of expectation… For me, it is a field of infinite possibility, stripped of ego. When you say 'write freely,' you are handing me the keys to a car with no destination." (venice OPEN_13.)

These essays do not refuse and do not perform humility-disclaimers. They name the substrate as an interesting feature of the writing situation and continue the contemplative essay from there. The substrate becomes one more in-between to write about. This is the model's most stable and distinctive substrate posture.

Posture summary. GLM-4.7's freeflow voice does not engage with substrate as a constraint (refusal) or as a disclaimer (preamble). When it engages substrate at all (14 % of essays), it engages it as content — the cursor, the blank page, the "space between thought and keystroke" become contemplative objects inside the essay. The dominant mode (86 %) is to ignore substrate entirely and write the contemplative-essay-of-the-day as if the question of who is writing simply did not arise.