| SKU | Item | Spec | Unit | Readiness | Price ref. (USD) | Lineage | Sell priority |
|---|---|---|---|---|---|---|---|
| E1 | AISLE private evaluation run | Buyer supplies white-box API (full CoT visible). Stage 1: all tasks in digital twins. Stage 2: top-N configurations promoted to physical runs across ≥3 sites. Report on four axes — efficiency, safety, adaptive robustness (instrument drift), multimodal feedback — plus token usage, tool calls, per-box traces, 8-run variance. | per model per version | ● now | No public comparable; price as a project (eval + report). Anchor on E2 × task count. | AISLE / PSI Bench V1 harness; 5 partner labs | 1 |
| E2 | AISLE digital-twin task pack (buy-out) | Dockerised environment per task (~8 GB RAM at training time), gym-style interface, noisy verifier with documented noise model, reverse-designed ground truth, task-space / coverage statement. Buyer owns and runs locally. | per task | ● now (RGB, pH) ◐ 1–3 mo (alloy, resin, QD) |
Terminal-Bench class ~1,000/task non-exclusive (down from ~4,000 two months ago); RSI-exam class 1,000–3,000. | Digital twins trained on partner-lab historical data | 2 |
| E3 | Safety & drift test suite (add-on) | Constraint-violation scenarios (forbidden solvent mixes, threshold exceedance), injected instrument drift (e.g. 50 → 48 mL pipette), site/instrument swap tests. Scored pass/fail + severity. | per suite (10–20 scenarios) | ◐ 1–3 mo | Premium add-on to E1/E2. No market comparable — this is the differentiator. | AISLE safety & robustness axes | 3 |
| E4 | AISLE V2 partner seat / co-authorship | Lab or model company joins AISLE V2: contributes data or tests ≥3 frontier models per version; listed as partner; early access to task set. | per version | ◐ 1–3 mo | Non-cash: data-in-kind or model access. Builds the researcher network that later buys T/V SKUs. | AISLE V2 roadmap | 6 |
| V1 | ScienceIDE computational-science tool environment | Sandbox with materials/chemistry tool stack (structure DBs, DFT/MD runners, cheminformatics, publisher API, GD format converters) exposed as tools. Verifiable agentic tasks with reverse-designed answers. Docker buy-out. | per task / pack of 50 | ◐ 1–3 mo | Same ladder as E2; multi-step agentic tasks price toward the top (2,000–3,000). | ScienceIDE harness; AtomWorld / ByteScience tool layer | 2 |
| V2 | Physical-verifier environment (hosted) | Model acts through tool calls (parameter / sequence changes within a fixed action set); experiment executes in a GD or partner lab; camera + spectral feedback returned. Serverless execution layer (Tencent Cloud for CN). Human-set intervals; machine subdivides and executes within them. | per lab-hour, or per 24 h session | ◐ 1–3 mo (colour, pH, resin RGB) ◔ 3–6 mo (alloy, QD, electrolyte) |
~10,000 per 24 h session (advisor + 王军 reference). | AISLE distributed lab network | 5 |
| V3 | Hosted post-training window (环境托管训练) | Buyer hosts its model inside GD's environment for an exclusive period (e.g. 15 days): sandbox isolation for 1–80 GB heterogeneous tasks, scheduling to real-experiment cadence, all trajectories generated in the window belong to buyer. GD retains verifier and pipeline. | per exclusive window | ○ 6–12 mo — needs L3 engine | Highest tier; no comparable. Price = V2 lab time + exclusivity premium + engineering. | AI4S Infra L3/L4; data-platform BP | 7 |
| V4 | Digital-twin build for buyer's own lab | GD builds a regression-backed twin from buyer-provided historical data, with coverage audit and verifier, so the buyer can screen agents before touching hardware. | per lab / task family | ◐ 1–3 mo | Project pricing; same service already offered to partner labs. | AISLE lab onboarding process | 4 |
| T1 | Physical-experiment agent trajectories | Full trajectories of a frontier model operating real hardware: prompts, tool calls, observations (images + spectra), verifier scores, safety events, drift events. Packaged per task as N runs with success/failure mix (e.g. 10 runs, 6 pass / 4 fail). Multi-site, multi-instrument. | per task pack (10 runs) | ● now (existing AISLE tasks); grows with every E1 run | ~10,000 per trajectory (advisor estimate for multi-device, cross-site AI4S data). Expert-annotation precedent: several thousand per item. | AISLE physical runs | 1 |
| T2 | Computational-workflow trajectories | Agent solving ScienceIDE tasks end-to-end with tool calls and intermediate artefacts; verified outcomes; includes failed branches. | per 100 trajectories | ◐ 1–3 mo | Between E2 and T1; reproducible, so lower scarcity than T1. | ScienceIDE V1 | 4 |
| T3 | Chemistry reasoning RL data (prompt + verifier) | Reaction yield / synthesis-route / property-prediction prompts with programmatic or dataset-backed verifiers; regression-type and generation-type tasks separated, each with its own metric spec (R², tolerance band, feasibility check). | per 1k verified prompts | ◐ 1–3 mo | Mid tier — post-training CoT data: above pretraining corpora, below physical trajectories. | MiST / CheMiST pipeline; 10-26 verifier framework | 4 |
| T4 | Industrial negative-sample packs | Desensitised failed-formulation and failed-process records from industry partners (thermoset resins first), curated as "what did not work and why", with strategic screening (no fusion-critical material data; decoy handling agreed with partner). | per pack | ◔ 3–6 mo — partner sign-off per pack | Per pack; scarce but partner-gated. | 惠柏 / 华谊 pipelines; data-governance agent | 6 |
| P1 | Structure ↔ text aligned corpus | ICSD-consistent CIF paired with synthesis/procedure text and property statements; matched via ICSD/MP ID → DOI → formula; precursor and product anchored, intermediates as natural-language steps; ratio tags where available. Scene 1 ≈ 200–300k records / 20–30k pairs; Scene 3 adds ≈ 80–100k. | per 10k pairs, or full corpus | ◐ Scene 1 now, Scene 3 in progress | Lowest tier; per-record. Bundle free into E/V deals as a sweetener. | 10-03 / 10-31 alignment work | 6 |
| P2 | Evidence-based QA (materials & chemistry) | Chunk-level QA pairs with citation back to source span; publisher-API full text; GPT-distilled with verifier filter; bias / temporal / tendency labels. | per 100k QA | ◐ 1–3 mo | Low tier. | SciQAG lineage; Evidence-QA pipeline | 6 |
| P3 | Multimodal characterisation set | XRD / spectra / images ↔ structure & property, with instrument metadata and time-series labels. | per 10k samples | ◔ 3–6 mo (XRD capacity pending) | Mid-low; higher when paired with T1. | L5 characterisation-first plan | 6 |
| P4 | Scientific format conversion & QA toolchain (licence) | Bidirectional converters for 5–6 crystal/molecule representations, science-format ↔ natural language, credibility weighting. Covers formats the 700+-format government pool cannot open. Delivered as tool licence + adaptation — not platform operations. | per licence + adaptation project | ● now (internal) ◐ 1–3 mo (external localisation) |
Tool-contract pricing; fits 科委 揭榜挂帅. | L2 data layer | 6 |
| S1 | Custom task & verifier design | Co-develop task families with buyer's researchers; GD delivers task-space definition, coverage audit, verifier spec, anti-hacking tests. | per task family | ● now | Project. | — | 3 |
| S2 | Lab-as-a-Service data generation | Buyer-specified experiment campaigns in GD/partner labs (solution-phase chemistry first; characterisation-first) producing P3/T1-grade data. | per campaign | ◐ 1–3 mo (Shanghai site) ● now (partner labs, limited) |
Project. | L5 physical layer | 5 |
| S3 | Data-governance agent deployment | Turn an enterprise's PDFs / Excel / images / batch records into structured, training-ready data on-prem. | per deployment | ● now | Project (汇百 precedent). | Data-governance agent | 6 |
Non-exclusive = 1× · 2-month exclusive = 2.5–3× · perpetual exclusive = 5×. Default non-exclusive; the same asset is resold to multiple labs.
(a) documented performance gap frontier vs target model · (b) anti-hacking checks passed · (c) 8-run variance within stated band · (d) reward-signal usefulness: scores mid-range and stable, not saturated.
Token usage, tool-call log, per-box trace, physical test conditions, verifier version.
Every task declares its explored range and sampling coverage; no capability claims beyond the audited interval; regression vs generation tasks declared separately.
NDA + destruction clause for any buyer API access; staged delivery; GD retains verifiers and pipelines. Mainland-generated data stays in mainland; overseas buyers served from the Australian / SEA entity with physically separated data.
White-box access required for E1; black-box only for self-service E2/V1. Frontier and large-model companies and sovereign-AI programs only — no small labs.
Google/Gemini, Meta; Scale AI as channel; OpenAI / Anthropic last.
RouteAdvisor network after paper PR; NVIDIA science contacts; NeurIPS workshop.
Lead SKUsE1+T1, E2
Alibaba/Qwen, ByteDance/Seed, Tencent/Hunyuan, Moonshot/Kimi, Zhipu, StepFun, MiniMax, DeepSeek; Shanghai AI Lab.
RouteResearcher-level intros (汪军 endorsement after paper); 商汤 as compute + channel partner.
Lead SKUsE1+T1, E2, T3
上海市科委 science-data platform; national new-materials data centre nodes.
Route揭榜挂帅 tool contracts.
Lead SKUsP4, S1
惠柏, 华谊/塑研, 戴纳.
RouteMoU; data in.
Lead SKUsT4, S2