Every answer, traced.
- plan: Revenue against guidance, then consensus
- retrieve: search_documents · 3 queries · 18 passages
- graph: graph_lookup · NVIDIA · 12 sourced links
- read: read_document · guidance table, p.2
- write: Verdict, then three cited insights
- check: 9 figures matched · 0 unverified
- The answer: NVIDIA printed above its guided range again in Q2.
- Done: model qwen3.8-27b, 7,180 tokens in, 412 out, no unverified figures, recorded in the answer log.
Every figure, checked.
What counts as the same figure
Figures are copied exactly. A large amount may be scaled and rounded, never changed. Single digits are too noisy to check.
- 1,754,877,990원17.5억
- 1,234,000,000$1.2B
- 60.4550060.46
Pick your model.
- FP8 · DGX Spark
Qwen3.8 27B
The open model for a desk, on the machine in your office.
- 0.17 s to the first word
- 62 tokens a second across eight people
- A 7,000-token read in 1.3 s
- Apache 2.0 licence
- Nothing leaves the machine
Measured on a Core in use
- 4-bit · DGX Spark
Qwen3.8 Flash-Next
A faster open model for quick questions, on the same machine.
- 27.6 tokens a second for one person
- 43 tokens a second across four
- A 4-bit build, light on memory
- Nothing leaves the machine
Measured on a Core in use
- Your key · cloud
Claude, GPT or Gemini
A frontier model under your firm’s own agreement.
- Your contract, your bill
- Questions and passages go to the provider
- The same citations and figure checks
- The same answer log and review
Admins choose who may use it
What compliance downloads.
A model of your own, with your partnership.
- 01
Start
An open Qwen model on your machine, used as published.
- 02
Collect
The answers your analysts approve, and the citations they correct.
- 03
Tune
A new version trained on your machine. Your documents and examples stay on it.
- 04
Examine
Questions from your own work, held back from training. It must beat the version in use.
- 05
Promote
Compliance switches it on. Earlier versions are kept, so you can roll back.
Treasurer’s own would be named Treasurer-32B v2. We publish no scores for a tuned model until it has passed that exam on a firm’s own questions.
What tuning teaches.
- Your house style
- The rating box, the decision line and the order your reports take.
- Your definitions
- Guidance against consensus, adjusted against reported, the way your desk uses them.
- When to say no
- “Not in the sources” instead of a plausible number to fill a gap.
Questions.
Is the model trained on our data?
Not to start with: Core runs the model as published, and the model card says so. With your partnership, we tune a version on your machine from the answers your analysts approve. It passes an exam on your own questions before Compliance switches it on. Your data never trains a model for anyone else.
How do we know an answer’s figures are right?
Every figure in an answer is matched against the passages the model read, allowing for rounding and units. A figure with no source is flagged before anyone relies on it, and every citation opens the sentence it came from.
What does an answer cost?
On an open model on your machine, nothing per question. With a cloud model, the provider bills its tokens under your agreement. Either way, Core counts every answer’s tokens by person and by model.
Do you publish scores for a tuned model?
Not yet. A score means something only on your own questions, so a version is examined on work held back from training before it goes live. We publish no figures until a model has passed that exam.
Can we bring our own model server?
Yes. Core works with any server that speaks the OpenAI API, such as vLLM, on your network.
