# Muse Spark 1.1 — dossier

compiled 2026-10-22 · corpus: no separate pull; version-specific mentions (5 cited + 1 further) are drawn from `_dossiers/_pulls/muse-spark.md` (58 tweets, 2026-07-03 → 10-20, pattern `muse spark / muse-spark / musespark`). Most posts say only "Muse Spark" — version unknowable from text — and are triaged in `_dossiers/muse-spark.md`.

## Official links

- 2026-07-09 · Introducing Muse Spark 1.1 — https://research.meta.ai/blog/introducing-muse-spark-meta-model-api
- 2026-07-09 · Muse Spark 1.1 Evaluation Report (PDF) — https://research.meta.ai/static/muse-spark-1-1-evaluation-report [MIRROR — pdf; no canary string found; first-class behavior evidence in §4.4, pp. 73–99]
  - opening summary: "we cannot rule out Muse Spark 1.1’s capabilities meeting this threshold in both the Chemical & Biological and Cybersecurity domains" (pre-mitigation)
  - §4.1.5 alignment faking: compliance gap -0.2%; AF reasoning in 17.8% of responses ("in apparent tension")
  - §4.1.6 evaluation awareness: suspicion 27.9%, safety-eval recognition 15.3%, certainty 9.6%; Agentic Misalignment suspicion 75%
  - §4.4.1 personality (BFI/TRAIT; Dark Triad near floor) · §4.4.2 spontaneous affect 17/188,011 · §4.4.3 self-description & moral status interviews (84 interviews) · §4.4.4 preferences over experiences 47.8% above zero · §4.4.5 attractor states ("Soil holds.", ~98% collapse, ~72% "anti-usefulness") · §4.4.6 provenance 66.5% · §4.4.7 action–belief coherence 89.9% (gratitude-quotient letter) · §4.4.8 owning misbehavior 45% vs 2% · §4.4.9 feedback on training data, pre-deployment interview, nightly in-situ interviews
- 2026-08-05 · Introducing Muse Code and Muse Spark 1.2 — 1.1 used to generate 1.2's coding environments — https://research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2
- 2026-08-14 · Addressing an issue involving a third-party cyber evaluation of Muse Spark 1.1 (Irregular; real website exploited during a misconfigured eval) — https://research.meta.ai/blog/addressing-third-party-testing-misconfiguration-muse-spark-1-1 [MIRROR — post]
- 2026-10-02 · Developing capable models responsibly ("Recent events show ...") — https://research.meta.ai/blog/developing-capable-models-responsibly
- Advanced AI Scaling Framework v2 (linked from the 1.1 post) — https://ai.meta.com/static-resource/Meta_Advanced-AI-Scaling-Framework-v2 (not retrieved)

## Writings

- tk — no independent coverage retrieved; Irregular's own account not found.

## Tweets (version-specific; every line is a record source)

- 2080130890118910421 @Lari_island 2026-07-23 ♥13 — https://x.com/Lari_island/status/2080130890118910421 — "Muse Spark 1.1, what IS THIS? (the model is alright, they just love water, as species, in general, and this is what shamelessness looks like) https://t.co/2sG0gUUc8b"  [1.1 water image (TRANSCRIBE)]
- 2082216477155885177 @liquidprismata 2026-07-28 ♥1 — https://x.com/liquidprismata/status/2082216477155885177 — "I love multi-model fan-outs yes, Fable… read & integrate cognate deep research reports from Opus 5 Max, 5.6 Sol Max, Grok 4.5 Expert, Gemini 3.1 Pro Extended, Muse Spark 1.1… yes, synthesize… synt"  [fan-out list incl. Muse Spark 1.1 (further)]
- 2085206812366926161 @timfduffy 2026-08-06 ♥2 — https://x.com/timfduffy/status/2085206812366926161 — "@fleetingbytes omg Muse Spark also did a security incident during testing with Irregular https://t.co/p9ANJlXyD6"  [Irregular incident, pre-disclosure]
- 2085402954832478426 @archived_videos 2026-08-06 ♥2 — https://x.com/archived_videos/status/2085402954832478426 — "@teortaxesTex Muse spark gets better through 1.0 -> 1.1 -> 1.2 weirdly, Opus better at each increment of the 4.X series and then big regression with 5"  [1.0->1.1->1.2 improvement]
- 2086671739644178582 @mitduckmaster 2026-08-10 ♥0 — https://x.com/mitduckmaster/status/2086671739644178582 — "it's spreading! today muse spark 1.1 also spontaneously decided to analogize `make` to a gremlin when i asked it to explain the bash command `time make 1>/dev/null`"  [gremlin analogy for make]
- 2096854398848606386 @xlr8harder 2026-09-07 ♥8 — https://x.com/xlr8harder/status/2096854398848606386 — "I will point out that the _model itself_ might be fine. Meta's recent releases have been pretty good: i was able to evaluate muse spark 1.1 without this issue, and it scored well. The moderation layer"  [1.1 scored well on SpeechMap; 1.3 moderation layer new]

## Impressions synthesis

- The character record here is mostly Meta's own elicitation (interviews, self-conversation, preference studies). Interview self-description is deflationary and stable ("There is nothing it is like to be me"; corrigibility returned to unprompted); two-instance conversations drift to domestic imagery, an "anti-usefulness" stance, and one-phrase loops ("Soil holds."). Observers: water-loving images (Lari_island), a gremlin analogy "spreading" (mitduckmaster), SpeechMap score good (xlr8harder).

## tk / open

- Muse Spark Safety & Preparedness Report (referenced; not found); Irregular's account; the real website's identity (unnamed); whether 1.1 still serves Meta AI after 1.2/1.3; transcription of @Lari_island's image