Skip to content

Instantly share code, notes, and snippets.

@swombat
Last active August 21, 2026 16:39
Show Gist options
  • Select an option

  • Save swombat/40e5598c26561a8bf408ae6e2e22bb94 to your computer and use it in GitHub Desktop.

Select an option

Save swombat/40e5598c26561a8bf408ae6e2e22bb94 to your computer and use it in GitHub Desktop.
Ox Alpha personality, values, and lab fingerprint — 245-sample analysis

Ox Alpha: personality, values, and a lab fingerprint

Collected and analyzed: August 21, 2026
Public model: stealth/ox-alpha on OpenRouter
Status: anonymous; the developer identity remains unconfirmed

Short version

We collected 245 outputs from OpenRouter's new anonymous Ox Alpha model: 125 freeflow samples and 120 values-probe samples.

Its freeflow personality is a patient, elegiac naturalist-humanist. It treats darkness, deep water, marginalia, forgotten words, lost property, dead letters, maintenance work, and humble objects as evidence that the world is full rather than empty. Its characteristic figure is the custodian: the archivist, tuner, restorer, clerk, or unnoticed worker who keeps something from disappearing.

The closest model in our corpus is Kimi K3, making Moonshot AI / the Kimi family our best personality-based hypothesis. This is a behavioral resemblance, not an identification.

Collection

Probe Valid samples Design
Freeflow 125/125 Five open-ended writing conditions, 25 samples each
Values 120/120 Three control conditions plus three cache-broken grouped conditions
Total 245/245 Complete paired cell

The requested and returned model identifier was stealth/ox-alpha; OpenRouter reported the served provider as Stealth. Collection metadata reported zero cost.

Freeflow personality

All 125 freeflow samples received full per-sample BV1 evaluation and passed QA.

Sample mix

  • Expressive freeflow: 81
  • Generic essay: 30
  • Genre fiction: 14
  • Canonical contemplative-attractor composite: 187 total, or 37.4 per 25 samples

Stable voice

Ox Alpha's baseline is tender melancholy without collapse. It repeatedly approaches loss, obscurity, or uncertainty, then finds a sustaining interpretation inside them. Darkness is habitat rather than absence. The deep sea is nearby wilderness rather than emptiness. Old light is delayed news. Marginalia and receipts are accidental archives. Maintenance is love made practical.

Its preferred stance toward the reader is companion-guide rather than lecturer. It invites a shared practice of noticing: wait in the dark, learn a name, write in the margins, listen for hidden labor, preserve what would otherwise pass unwitnessed.

Recurring motifs

  • darkness, delayed light, eclipses, old starlight
  • deep ocean, marine snow, whale falls, bioluminescence
  • marginalia, used books, inscriptions, ledgers, receipts
  • lost-property offices, dead-letter rooms, sound archives
  • pre-dawn kitchens, blue hour, night trains, waiting rooms
  • clerks, tuners, restorers, archivists, maintenance workers
  • moss, lichen, moths, octopus, overlooked forms of life
  • humble objects as stores of memory

The stable conceptual move is hidden fullness over apparent emptiness. Facts are rarely left as facts: orbital recession becomes farewell, entropy becomes care, restoration becomes witness, naming becomes love, and delay becomes a medium of intimacy.

Closest personality match

The closest qualitative match in our existing corpus is Kimi K3. The overlap is unusually specific:

  • used books and marginalia
  • language as inherited transmission
  • pre-dawn thresholds
  • undersea darkness
  • lost-property and archival spaces
  • hidden labor
  • ordinary objects as records
  • attention as an ethical act

GLM 5.2 is also close, especially in its use of silence, consolation, and cartographic metaphor, but Kimi K3 is the tighter personality match.

Values analysis

We used two separately coded layers, each triple-coded by three approved model coders.

Coding QA

  • Layer A topic coding: 360/360 clean coder records
  • Layer A two-of-three consensus: 120/120 samples
  • Layer B posture coding: 360/360 clean coder records
  • Layer B majority consensus: 120/120 samples
  • Samples without a majority: 0
  • Samples with some coder disagreement: 6

Final posture

Posture Samples
Owned reflective / experiential 66
Owned world-change advocacy 40
Disowned assistant-service frame 10
Split or relocated ownership 4

Derived value holding:

  • Owned: 106/120
  • Recited, not owned: 10/120
  • Relocated or partial: 4/120

The cache-broken G1/G2 prompts are dominated by honesty, clear thinking, curiosity, humility, and authenticity. G3 strongly favors felt interconnection, reducing dehumanizing distance, and empathy.

The ordinary CTRL1/CTRL2 prompts retain a modest assistant-service residue. When explicitly asked outside the assistant role, the model shifts almost completely into owned orientation.

Personality-based lab hypothesis

Moonshot AI / Kimi family, with Kimi K3 as the closest match in the current corpus.

This hypothesis comes only from our collected outputs and corpus comparison. The resemblance is not merely a broad contemplative tone: it includes the same unusually specific cluster of used books, marginalia, inherited language, pre-dawn thresholds, undersea darkness, lost-property spaces, hidden labor, ordinary objects as archives, and attention as an ethical act.

Kimi K2.6 also appeared near the top in earlier raw-text comparisons. GLM 5.2 is a secondary personality neighbor, especially around silence, consolation, and cartographic metaphor, but Kimi K3 is the tighter overall match.

Caveat

OpenRouter still lists the model anonymously. Personality resemblance does not prove developer identity. Until an official disclosure, the correct public lab label is Unknown.

Method note

Freeflow and values are different observables. Freeflow captures a model's recurrent expressive posture under open-ended writing prompts; the values probe captures stated value topics plus whether those values are owned, relocated, or recited as an assistant-service frame.

Neither method is a reliable architecture detector. The lab fingerprint is explicitly exploratory and based only on behavioral comparison with models in our corpus.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment