## Candidate
**Trillion Labs** — Seoul-based foundation-model startup. Seed facts from Jeff (2026-09-12), to be verified:
- Pretrains its own models; the latest is **Gravity Nano (16B-A3B)**, and a **larger model has recently been in pretraining**.
- **Jay** (co-founder) worked on synthetic data at NAVER and later, with **Hwaran Lee**, on the **Prometheus** series of evaluator-LM papers.
- The company has **recently pivoted toward AI for automating data centers**.
- **rBridge** — a method for predicting reasoning capability from small proxy models.
Not referenced anywhere in the repo today (`grep -ri trillion data/` is empty). Korean labs currently tracked: KAIST, Kakao, KRAFTON, LG AI Research, Motif, NAVER, SKT, SNU, Upstage.
## Plan
1. **Deep research** (two parallel passes: corporate/people/news; models/papers/provenance), every claim traced to a primary source (HF config.json, arXiv title pages, company posts), posted here as a comment.
2. **Adversarial review** by independent critics arguing against inclusion and against the research's accuracy, posted here.
3. **Recommendation** (add now / add later with triggers / do not add), against the AGENTS.md Research Focus and the add-lab skill's criteria: from-scratch frontier-relevant pretraining, foundational technique papers (rBridge, Prometheus lineage), AA presence, and whether the data-center pivot changes the lab's research identity.
No files will be created until Jeff decides.
## Deep research (2026-09-12)
Two parallel passes (corporate/people/news; models/papers/provenance) plus my own re-verification of every load-bearing claim. Markers: **[V]** = I verified at the primary source myself; **[A]** = verified by a research agent at a primary source I did not re-open; **[U]** = unverified.
### Corporate
- **Trillion Labs (트릴리온랩스)**, Seoul (Gangnam; site footer) **[A]**; LinkedIn says HQ San Francisco with a Seoul office, and the models pass reports Delaware incorporation **[U]**. Founded **2024** — sources disagree on the month (Feb/Apr incorporation vs. Aug public launch; GitHub org created 2024-08-05) **[A]**. ~20 people (LinkedIn 21, HF org 19 members) **[A]**.
- **Founders:** **Jamin "Jay" Shin (신재민)**, CEO, ex-NAVER research scientist on HyperCLOVA X **[A]**; corresponding author `[email protected]` on rBridge **[V]**; Scholar `GuBHIwsAAAAJ` (affiliation Trillion Labs) **[A]**. **Jason Juneyoung Park**, co-founder, ex-Riiid VP of AI Research, KAIST PhD; profile reads business/ops rather than pretraining **[A]**.
- **Funding:** **$4.2M (₩5.7B) pre-seed, 2024-09-10**, led by Strong Ventures with Kakao Ventures, Bass Investment, TheVentures, Goodwater Capital, BAM Ventures ([ZDNet Korea](https://zdnet.co.kr/view/?no=20240910161930)) **[V]**. VentureSquare (2025-09) says $5.8M — possibly an extension **[U]**. No later round found in EN or KO press; **valuation not public** **[A]**.
- **Government programs:** selected 2025-11-05 into MSIT's *AI-specialized foundation model* project as the FM engine builder in the **Lunit-led medical-science consortium** (Lunit, Trillion, AIGEN Sciences, SK Biopharm, Kakao Healthcare, Standigm, Rebellions, D-Circle + KAIST/SNU labs + hospitals; ₩18.2B) **[A]** — this is why the Gravity cards credit "Trillion Labs and the Lunit Consortium" **[V]**. **Not selected** on 2026-02-20 in the sovereign-model ("독자 AI 파운데이션 모델") additional round: MSIT picked **Motif Technologies**; the briefing says Trillion lost by a narrow margin and was praised for "AI 모델의 기술 자립도" and leaderboard results ([korea.kr](https://www.korea.kr/news/policyNewsView.do?newsId=156745273)) **[V]**. Also: NIPA compute grant (funded Trida), NVIDIA Inception **[A]**.
### The pivot
- **2026-06-08** press release: *"Trillion Labs is Developing Industrial World Models for AI Factories, built with NVIDIA Omniverse and NVIDIA Nemotron"* — world models to understand, simulate, and optimize data centers, power plants, cooling; partner **GS Group** on energy infrastructure; CEO quote: "our next step toward extending AI from language understanding to real-world reasoning" ([PR Newswire via Yahoo](https://finance.yahoo.com/sectors/technology/articles/trillion-labs-developing-industrial-world-120000636.html)) **[V]**.
- The website is now branded **"AI for AI Factories — Foundation models for the AI factories that train AI"**, mission: "specialized foundation models that operate the physical substrate of AI itself"; case study "AI for Energy" (power-plant operations: P&IDs, schematics, time series); the research strip still lists **Gravity Nano (16B-A3B) "a new MoE series, trained from scratch"**, **rBridge**, and **Tri (7B to 70B) "our first generation, trained from scratch"** ([trillionlabs.co](https://trillionlabs.co)) **[V]**.
- The pivot followed the February loss of the sovereign-model slot (their bid was an industrial consortium with GS 52g / POSCO Holdings / EXEM) **[A]**. **Pretraining continues** but has shifted to domain (medical/bio, industrial) specialization: see Gravity below.
### Models (configs read from HF `config.json`; cards quoted)
| Release | Date | Arch / shape | Provenance | Notes |
|---|---|---|---|---|
| **Trillion-7B-preview** | 2025-03-14 | Llama-class dense, 32L×4096, MHA, vocab 128,128, ctx 4K | **From scratch**, 2T tokens, 59.4K H100-hours ($148K) per the tech report **[V]** | Apache-2.0; XLDA; Korean-centric multilingual |
| Tri-0.5B / 1.9B Base | 2025-04 | Llama-class | from scratch (bring-up runs) **[A]** | intermediate checkpoints released |
| **Tri-7B** (+Base) | 2025-07-22 | same geometry, ctx 32K | from scratch, new run **[A]**; token total **[U]** | 15 intermediate checkpoints |
| **Tri-21B** | 2025-07-19 | `TrillionForCausalLM` dense 40L×5120, head_dim 160, vocab 124,416 **[V]** | from scratch (in-house geometry; CEO interview) **[A]**; tokens **[U]** | gated, custom Trillion license (research; commercial > $10M ARR) **[V]** |
| **Tri-70B-preview-SFT** | 2025-08-01 | dense 80L×8192, vocab 124,416, iRoPE + sliding window, ctx 32K **[V]** | **from scratch, 1.5T tokens** per card **[A]** | SFT-only; base never released; 13 checkpoints; = the lab's Scale entry |
| Trida-7B | 2026-01-19 | `TridaForDLM` block-diffusion LM **[V]** | continually pretrained from Tri-7B **[A]** | "first Korean open block-dLLM" |
| gWorld-8B / 32B | 2026-01-28 | `Qwen3VLForConditionalGeneration` **[V]** | **derivative of Qwen3-VL** (card `base_model`) **[V]** | ICML 2026; GUI world model rendering HTML/CSS |
| **Tri-21B-Think** (+Preview) | 2026-02-15/18 | dense 21B, ctx 32K (262K YaRN) **[V]** | derivative of same-lab Tri-21B **[V]** | **AA Intelligence Index v4.3: 9** (Preview 10, deprecated) **[V]**; GPQA-D 62.6, AIME 2026 56.67 (card) **[A]** |
| **Gravity-16B-A3B** Base/Preview ("Gravity Nano") | 2026-04-09/15 | `GravityMoEForCausalLM` (DeepSeek-V3-compatible MLA + MoE), 28L, 64 routed + 1 shared experts, top-8, GLM-4.5 tokenizer (151,552), ctx 32K→128K **[V]** | **from scratch "(from random weights)", ~5.5T tokens**, STEM/medical emphasis, with the Lunit Consortium **[V]** | Apache-2.0; 16.24B / 3.16B active; Preview: GPQA-D 55.1, AIME24 43.3, MMLU-Pro 71.5 **[A]** |
| Gravity-bio-16B-A3B | 2026-06-25 | mid-trained from Gravity-16B | same-lab derivative **[A]** | with TheBioCollection |
| **Gravity-30B-A5B** Base/Preview (+ `Gravity-30BA5B-RL` 08-30) | 2026-09-09 | 52L (2 dense + 50 MoE), same expert geometry, 29.56B / 5.34B active, ctx 128K **[V]** | **LLaMA-Pro block expansion of Gravity-16B-A3B-Base (28→52 layers) + continued pretraining + 128K extension** per the base card **[V]** — a same-lab derivative, not a new from-scratch run | Preview: execution-labelled tool-use DPO, BFCL v3 Live 73.11; no general benchmarks yet **[V]**; not on AA |
| TxGravity-30B-A6B-Open | 2026-09-09 | on Gravity-30B-A5B | co-developed with **SK Biopharmaceuticals** **[V]** | therapeutics (TDC): 16 wins / 26 ties / 18 losses vs TxGemma-27B **[V]** |
Also minor derivative repos (Trillion-LLaVA-7B, Tri-7B-Search, Tri-1.8B-Translation, NemoSlides on Nemotron-3-Nano, sim-scholar on Qwen3-4B) **[A]**. **The "larger model recently in pretraining" is best explained by Gravity-30B-A5B**, released three days ago — and it is depth-upscaled from the 16B, not a fresh run; no evidence of a funded run above Tri-70B (the 1T-parameter omnimodal bid was the rejected February proposal) **[A]**.
### Papers (affiliation blocks read from PDF page 1)
1. **Trillion 7B Technical Report** — [2504.15431](https://arxiv.org/abs/2504.15431), 2025-04-21, "Trillion Labs" **[V]**. **XLDA** (Cross-lingual Document Attention): language-aware packing plus a selective mask that lets Korean/Japanese tokens attend to co-packed English while masking unrelated documents; 10% multilingual tokens; tokenizer sizing law; WSD schedule **[A]**.
2. **rBridge: Predicting LLM Reasoning Performance with Small Proxy Model** — [2509.21013](https://arxiv.org/abs/2509.21013), 2025-09-25, **ICLR 2026**, Trillion Labs + KAIST AI (Koh, Suk, Han, Se-Young Yun, Shin) **[V]**. Replaces task accuracy with confidence-weighted NLL on frontier reasoning traces so ≤1B proxies predict 7B–32B reasoning; 100–700× cheaper pretraining-data selection; R² up to 0.94 **[A]**. Code, datasets (traces + logprobs), blog **[V]**.
3. **gWorld: Generative Visual Code Mobile World Models** — [2602.01576](https://arxiv.org/abs/2602.01576), 2026-02-02, **ICML 2026**, Trillion Labs + KAIST **[V]**.
4. **TheBioCollection** — [2607.08803](https://arxiv.org/abs/2607.08803), 2026-07-09, Trillion + KAIST + SK Biopharm + Lunit + AIGEN; 52.6B-token biology corpus + eval **[A]**.
5. **VibeProteinBench** — [2605.10978](https://arxiv.org/abs/2605.10978), 2026-05-09, Trillion with MIT/Harvard/Stanford/Caltech/FutureHouse co-authors **[A]**.
No paper exists for Tri-21B/70B, Gravity, or Trida (blog-only or none) **[A]**. Five arXiv papers in 18 months, two at top venues.
### Prometheus lineage (author lists verified on arXiv)
- **Prometheus** (2310.08491, Oct 2023): Seungone Kim, **Jamin Shin**, …, **Hwaran Lee**, … — Shin's affiliation then was NAVER AI Lab / NAVER Cloud **[V]/[A]**.
- **Prometheus 2** (2405.01535, May 2024): Seungone Kim, **Juyoung Suk**, …, **Jamin Shin**, … — both KAIST AI at the time **[V]/[A]**. **BiGGen Bench** (2406.05761): Suk (2nd author), Shin **[V]**.
- **Juyoung Suk** is now Member of Technical Staff at Trillion (and a KAIST AI MS student); Scholar affiliation Trillion Labs **[A]**.
- **Hwaran Lee is not at Trillion Labs**: Assistant Professor at Sogang University since 2025-03 (HAILO Lab), NAVER AI Lab 2021–2025; she appears on none of Trillion's papers **[A]**. Her link to the seed facts is the Prometheus 1 collaboration with Shin at NAVER. The Prometheus papers are **founder background, not Trillion outputs** (lead Seungone Kim is at CMU).
### Third-party presence
- **Artificial Analysis:** [tri-21b-think-v0-5](https://artificialanalysis.ai/models/tri-21b-think-v0-5) = **9** on v4.3 (released 2026-02-10, 32K, open weights, rank ~#48/140), [tri-21b-think-preview](https://artificialanalysis.ai/models/tri-21b-think-preview) = 10 (deprecated). No provider page, no Gravity/Tri-7B/Tri-70B entries **[V]**. Korean press "global top-30" claim was the February index version **[A]**.
- **OpenRouter:** none **[V]**. GitHub org `trillion-labs`: 29 repos; gWorld 61★, rBridge 6★; no training framework released **[V]/[A]**. HF org `trillionlabs`: 34 models, 11 datasets, 235 followers **[V]**.
### If added — what the filing would look like
- `data/labs/trillion-labs.yaml`: `type: private`, `region: korea`, `founded: "2024"`, no `valuation:` (none public), HF `trillionlabs`, GitHub `trillion-labs`, no `artificialanalysis:`/`openrouter:` provider links (none exist). Logo from the HF avatar.
- **Home table:** Scale = **70B** (Tri-70B-preview-SFT); Intelligence = **9** via Tri-21B-Think, `base_model: tri-21b` (single-hop resolves to a same-lab from-scratch model, so it counts); Openness unset (AAOI has no Trillion entry).
- ~10–12 outputs: Trillion-7B-preview (+report; flagship), Tri-7B, Tri-21B, Tri-70B-preview-SFT (flagship: scale milestone), Tri-21B-Think (AA anchor), Trida-7B, gWorld (`base_model: qwen3-vl`; ICML), Gravity-16B-A3B (flagship: first MoE, from scratch), Gravity-30B-A5B (`base_model: gravity-16b-a3b`, no `parameters`), TxGravity, rBridge (flagship paper), TheBioCollection (dataset), VibeProteinBench (eval), intermediate-checkpoint release (dataset/announcement).
- People: Jamin Shin, Sungjun Han, Juyoung Suk (+ Woosung Koh, Jason Park). News: 16 items back to the Forbes launch story.
### Unverified / gaps
Legal name and incorporation month; the $5.8M figure; any valuation; Delaware vs Seoul HQ; Tri-7B/21B token totals; Gravity-30B continued-pretraining budget; X account; OpenReview IDs; Jamin Shin's HKUST education and "lead author of the HyperCLOVA X report" (Crunchbase); Forbes article body (403); Platum article (403); citation counts (Semantic Scholar rate-limited).
## Adversarial review (2026-09-12)
Three independent critics were run against the research comment above, each with a different mandate, and I re-verified every claim they contested at the primary source. **A — skeptic** (argue against inclusion). **B — fact-checker** (argue against the research's accuracy). **C — index integrity** (assume the lab is added; check the filing against AGENTS.md, the add-lab/add-output skills, `src/data/loader.ts` and `src/validate.ts`).
### 1. Corrections to the research (B and C; each re-verified by me)
| Research said | Correction | Source |
|---|---|---|
| rBridge "100–700× cheaper" | The paper claims "over 100×" for dataset-ranking cost; 700× is the blog's headline. Quote the paper. | arXiv 2509.21013 abstract |
| Tri-21B-Think dated Feb 10 (AA) | HF repos: Preview created 2026-02-15, v0.5 2026-02-18. AA's date is its listing date. | HF API |
| Tri-21B tokens **[U]** | Card: 2.3T tokens, 20.73B params, vocab 124,416, native context 8,192, custom `TrillionForCausalLM`. | Tri-21B card |
| Tri-21B geometry (C flagged a card/config conflict: card says "32") | Config wins: 40 layers × 5120 hidden, 32 query / 8 KV heads, head_dim 160, intermediate 27,392, vocab 124,416. That geometry with untied embeddings computes to exactly 20.73B, the card's own count; the "32" is the query-head count. Read from Tri-21B-Think's ungated `config.json`, which shares the base's shape. | config.json |
| Tri-70B "custom Trillion license" | Contradictory on the card itself: HF frontmatter `license: apache-2.0`; body "licensed under the Trillion License"; gate text is the Trillion Labs AI Model License (commercial license above $10M ARR). Record the contradiction, not Apache. | Tri-70B-preview-SFT README |
| Gravity "by Trillion Labs and the Lunit Consortium" | The card assigns roles: "Lunit — Project lead and medical AI research; Trillion Labs — Model architecture, pretraining, and infrastructure; Aigen Science — Biomedical AI…; SK Biopharmaceuticals — …; Kakao Healthcare — …", MSIT-funded, NIPA-managed. Lunit's press "self-developed" is overreach: Lunit's own L1-16B-A3B card says "Derived from Gravity-16B-A3B-Base … from Trillion Labs and Lunit Consortium". | Gravity-16B-A3B-Base §Acknowledgements; L1-16B-A3B card |
| "A larger model has recently been in pretraining" | Resolves to Gravity-30B-A5B-Base (2026-09-09): LLaMA-Pro-style block expansion of Gravity-16B (28 → 52 layers), then continued pretraining and 128K extension. Not a new from-scratch run. | Gravity-30B-A5B-Base card |
| Tri-7B-Base date | Created 2025-04-04. | HF API |
| June 2026 pivot release | World models "built from scratch", but the reasoning component is Nemotron-based; the pivot's artifacts are not all in-house pretraining. | PR Newswire release |
Two things the research omitted that bear on the decision:
- **Llama-copy controversy (Feb–Mar 2026).** After Tri-21B-Think's AA listing, Korean outlets carried an allegation that the Tri series derived from Llama. The CEO's rebuttal (AI Times, 2026-03-10, [idxno=207647](https://www.aitimes.com/news/articleView.html?idxno=207647), verified): own tokenizer (vocab 124,416 vs Llama 3's 128,256), loss curves starting above 10, public intermediate checkpoints, and an MSIT technical review. The rebuttal is consistent with the public artifacts (custom architecture, head_dim 160, no Llama-compatible vocab; checkpoint repos for 0.5B/1.9B/7B/70B dated 2025-08/09), but the review's outcome is the CEO's account.
- **CEO on record retreating from scale competition** (same interview): "적은 자원으로도 고효율을 내는 모델 설계 역량이 한국의 돌파구" — efficiency at small compute, not frontier scale, is the stated strategy.
### 2. Case against inclusion (Critic A)
- **Scale.** ~20 people, $4.2M pre-seed (2024-09), lost the 2026-02 sovereign round to Motif; no round since.
- **Trajectory.** The largest from-scratch run (70B, 1.5T tokens) is a 2025-08 SFT preview whose base and tech report never shipped, 13 months on. The 2026 work is a 16B-A3B MoE and a block-expanded 30B. The June 2026 pivot to "AI for AI Factories" (data-center and power-plant world models for GS Group / POSCO) is off-thesis and has no public artifact.
- **Index calibration** (my numbers, recomputed): 41 labs carry an Intelligence cell, median 20. Trillion at **9** (Tri-21B-Think v0.5, AA v4.3) would tie Sarvam for ninth-lowest and be the lowest Korean cell by 11 points (LG 20, SKT 23, Upstage 28, Motif 34). Floor today: Swiss AI 5, Microsoft/AI21/Reka 6, Ai2/Liquid 7, TII 8.
- **Papers.** rBridge (ICLR 2026): 6 GitHub stars, no citation evidence. gWorld (ICML 2026): a Qwen3-VL derivative, 61 stars. Neither is field-shaping yet.
- **Precedent.** SOOFI (#45) stays open and untracked; #14 Tier 3 holds labs that pivoted away (01.AI, Aleph Alpha).
- **Verdict A: do not add now.** Revisit on objective triggers: a Trillion-primary from-scratch run at ≥5T tokens or ≥30B params (or Tri-70B base + report); any from-scratch model ≥20 on the current AA index; a ≥$20M round; rBridge at ≥100 citations or adopted in another lab's report; an open-weights + paper artifact from the world-model line.
### 3. Filing integrity if added (Critic C)
Blocking — each confirmed by me against the repo or the primary source:
- **B1** gWorld `base_model: qwen3-vl` fails `validate.ts` (no such slug; only qwen-vl / qwen2-vl / qwen2.5-vl are tracked). Use `pretrained_from_scratch: false` with an inline comment (Solar Pro 2 pattern).
- **B2** The Intelligence cell resolves only if `tri-21b.yaml` exists with that slug and no `base_model` (single-hop rule). Anchor to `tri-21b-think-v0-5` (9), not the deprecated Preview (10): distinct checkpoints, not modes, so `--audit-modes` will not churn. First observation at v4.3, so no history entry (MiniCPM5-2B precedent).
- **B3** Tri-70B license recorded as the Trillion Labs AI Model License with the frontmatter contradiction noted; same care for Tri-21B (commercial threshold is 1M MAU **or** $10M ARR).
- **B4** Attribution follows the cards, not press: Gravity-16B/30B → `lab: trillion-labs` alone (Lunit is untracked and a fine-tuner; KAIST people are acknowledgement-level on Gravity). rBridge/gWorld → `lab: [trillion-labs, kaist]` (the `krafton/moe-to-dense` precedent), **conditional on the PDF page-1 affiliation block**. TxGravity → Trillion, SK Biopharm in prose.
- **B5** Blurb must not carry the Korean-press "top-30 on AA" (that was the February index; today 9, #48/140 open-weights); must not place Hwaran Lee at Trillion (Sogang since 2025-03, on none of Trillion's papers; Prometheus is founder background via NAVER/KAIST, concrete lineage is Juyoung Suk); must describe Gravity-30B as block expansion; omit the $5.8M figure, any valuation, and the SF/Delaware HQ claim.
- **B6** rBridge not `flagship` (inflation at 6 stars).
- **B7** Intermediate checkpoints are point-release artifacts: fold into Tri-7B/Tri-70B sources, not separate outputs. Exclude Trillion-LLaVA, Tri-7B-Search, Tri-1.8B-Translation, NemoSlides, sim-scholar.
Non-blocking, verified: Gravity-16B from-scratch provenance is sound (card: "trained from scratch (from random weights)… approximately 5.5 trillion tokens", DeepSeek-V3-style architecture, tokenizer initialized from GLM-4.5 with vocab 151,552, which matches no copyable base's embedding matrix). Scale = 70B via the SFT preview is defensible under the Hy4-preview / Zaya1-74B precedents if the description says SFT-only, base unreleased, 1.5T tokens. Three flagships (Trillion-7B, Gravity-16B-A3B, Tri-70B-preview-SFT) is within peer range (Upstage 6/10, SKT 3/10, Krafton 1/7). Openness cell blank. `type: private`, slug `trillion-labs`, `founded: "2024"`.
- **Verdict C: conditional add on the minimal filing** (lab + 8 outputs). "One of only a handful of Korean organizations with a documented from-scratch pretraining record (a tech report with GPU-hours, public intermediate checkpoints from 0.5B to 70B, a 5.5T-token MoE pretrained from random weights), two top-venue papers, an AA-scored model, and the FM-engine role in a state-funded consortium. Every risk lives in the filing rather than the inclusion."
### 4. Where the critics split
A and C agree on every fact. They disagree on the test: A judges by present scale, score and trajectory; C judges by the verified pretraining record against the index's existing floor. That is the call the recommendation below makes.
## Recommendation (2026-09-12)
**Add Trillion Labs, as the minimal filing (lab + 8 outputs), with Critic C's B1–B7 applied.** Critic A dissents; the dissent is recorded below and converted into a downgrade trigger. No files until Jeff decides.
### Why add
1. **The inclusion test is met at the primary source.** Research Focus and the add-lab skill turn on from-scratch, frontier-relevant pretraining plus research output. Verified on the cards and repos: Trillion-7B (tech report, 2T tokens, 59.4K H100-hours); Tri-21B (2.3T, custom geometry and vocab); Tri-70B-preview-SFT (1.5T from scratch, public intermediate checkpoints); Gravity-16B-A3B-Base (5.5T from random weights, Apache-2.0, checkpoints public); Gravity-30B-A5B-Base (block-expanded, 128K, shipped 2026-09-09). Public intermediate checkpoints at 0.5B / 1.9B / 7B / 70B / 16B-A3B is more pretraining transparency than most tracked labs offer.
2. **Consistent with the index's revealed bar.** Tracked peers: Motif (one output), Swiss AI (AA 5), Sarvam (9), TII (8), Liquid (7), Krafton. Withholding Trillion on Critic A's triggers would apply a higher bar to it than to those.
3. **On-thesis breadth beyond the LM line.** Two top-venue technique papers (rBridge ICLR 2026, gWorld ICML 2026) and a science-FM line (Gravity-bio, TxGravity with SK Biopharm, TheBioCollection), which the scientific-FM scope rule counts.
4. **The pivot is a watch condition, not an exclusion.** 01.AI and Aleph Alpha went to #14 Tier 3 after pretraining stopped. Trillion shipped a pretrained base three days ago. The blurb names the pivot as off-thesis, and the trigger below re-evaluates it.
Why not "add later": the score (9) and the size are facts for the blurb, not criteria, and the index already carries lower cells.
### Conditions on the filing
- **Outputs (8):** `trillion-7b` ★ · `tri-21b` · `tri-21b-think` (AA URL `tri-21b-think-v0-5`, `intelligence_index: 9`, `"AA v4.3"`, no history) · `tri-70b-preview-sft` ★ (license per B3) · `gravity-16b-a3b` ★ (Lunit role split quoted) · `gravity-30b-a5b` (`base_model: gravity-16b-a3b`) · `rbridge` (paper, `lab: [trillion-labs, kaist]`, not flagship) · `gworld` (`pretrained_from_scratch: false`, `lab: [trillion-labs, kaist]`).
- **Home row:** Scale 70B · Intelligence 9 (v4.3) · Openness blank · Outputs 8.
- **Blurb:** Seoul, founded 2024, $4.2M pre-seed, ~20 people; the from-scratch record; Lunit-led MSIT consortium with the role split; the Feb 2026 sovereign-round loss to Motif; the Llama-copy allegation and rebuttal in one sentence; the June 2026 industrial world-model pivot, named as off-thesis with no public artifact; founder lineage to `/outputs/kaist/prometheus` via Juyoung Suk. No valuation, no "top-30", no Hwaran Lee.
- **People:** Jamin Shin, Sungjun Han, Juyoung Suk, Jason Park.
- **Deferred to a full filing**, each pending a PDF page-1 affiliation check: tri-7b, trida-7b, gravity-bio-16b-a3b, txgravity, thebiocollection, vibeproteinbench.
- **Open before filing:** rBridge/gWorld affiliation block at the PDF; Tri-70B `gated` status (API says `false` despite gate fields); Woosung Koh's listed affiliation; README lab count 108 → 109; logo source (HF avatar fallback).
### Downgrade trigger
Recorded here and, on add, in #14: if by 2027-09 Trillion has shipped no new from-scratch or continued-pretraining checkpoint and its public output is only AI-factory operations work, move it to #14 Tier 3 alongside 01.AI and Aleph Alpha.
### If the decision goes the other way
Leave #57 open as the watch item with Critic A's triggers, add Trillion to #14 Tier 2 (regional), and optionally file rBridge and gWorld under `kaist` (Se-Young Yun's group co-authors both) so the papers are tracked either way.