Prax6 2.8T is learning the work of consulting.
A research programme on open-weight models for strategy, management and innovation engagements. Its first asset is CSAB, a bench of realistic engagements graded by consultants.
Prax6 is not a published model yet. No performance will be claimed before it has been reproduced on private engagements absent from the training.
Cogni6 Research
Cogni6's applied research lab. Two goals, and the second explains why the first matters.
Advance open-weight models on consulting work
Strategy, management and innovation consulting is not a prompt category. It is a profession with its obligations, its formats and its review conditions, and no general-purpose model is trained on it.
Let a firm keep control of its own method
In time, a firm should be able to adapt a model to its own methods without pooling them into a shared set. That is the condition for its institutional intelligence to remain its own.
The programme is under construction. This page describes what is being built, in what order, and what is not built yet.
CSAB, the first asset
Cogni6 Strategic Advisory Benchmark
The first thing built is not the model. It is the means of judging it. A model you cannot evaluate can only improve by guesswork, and a bench written afterwards measures what the model already does.
Engagements, not questions
Each task is an engagement, or a substantial part of one: a short instruction of the kind a partner writes, a closed company file, tools to search and draft, and deliverables written to disk.
Files that push back
The file holds useful documents, contradictions, gaps and noise. Choosing what to read is part of what is measured, as much as what gets written at the end.
Rubrics written by consultants
Each engagement carries observable criteria tied to a specific deliverable, and critical failures that invalidate the work: an invented figure, a source that does not support the claim, a constraint of the mandate ignored.
Eight families of engagements, two languages
From framing to the executive file, through diagnostic, options, recommendation, operating model and roadmap. Each family exists in Canadian French and Canadian English.
A public set, a private set
Part of the bench is published so results can be checked. Another part stays private and rotates, so that measured progress is not memorisation. A gain that only holds on the public set is not one.
Status: the schema is settled, the first reference tasks are being written with consultants. The bench is not published yet.
Prax6 2.8T
The vertical model the programme aims at. It is post-trained, meaning it starts from an existing open-weight model and learns the profession on top of it.
The base model is not settled. It will be chosen on CSAB results, not before: several candidates are measured in the same environment, with the same tools and the same budget, and that comparison will decide.
The behaviours we are after
Confidentiality of the file
The model is meant for work where the file is not exposed. No vault content enters the training, and the environment it runs in is the one where sensitive values have already been replaced by tokens.
Long-horizon engagements
An engagement does not fit in one exchange. It breaks into phases, keeps its decisions and comes back to its own output. That is the shape we are after, not the isolated question.
Deep analysis
Reading three years of results, twenty interviews and a market report, then drawing out what changes the decision. Depth is judged by what comes out, not by how many pages were covered.
Traceability of every claim
A thesis whose sources cannot be traced back cannot be reviewed. The model has to tie what it asserts to the document in the file that supports it, and to say when that document does not exist.
Bilingual work
A Canadian engagement reads English sources and delivers in French, or the reverse. Both are working languages, not a translation added at the end.
Executive deliverables
A diagnostic, an options table, a recommendation, a roadmap and a risk register that answer one another. A deliverable that contradicts itself between sections is not shown to a committee.
The post-training method
The model learns where the work happens, and is judged elsewhere. That separation is what makes a gain believable.
- 01
Training in the real environment
The same phases, the same tools, the same human gates and the same deliverables as an engagement run by a consultant. Not a test harness built for the occasion.
- 02
A reward over the whole engagement
What gets optimised is the complete engagement: criteria met, sources accurate, critical failures avoided and the cost of the trajectory. Not how fluent the prose is.
- 03
Judges calibrated against experts
An automatic judge that drifts from consultants teaches the model its own flaws faster than it teaches consulting. Agreement between judge and expert is an entry condition, not a result.
- 04
Verification outside the training
A gain only counts if it holds on private engagements the model has never seen, and in an environment other than the one it learned in.
Data and governance
What may be used
- Entirely synthetic companies, markets and reports.
- Public documents whose reuse rights are verified.
- Cases and deliverables written expressly by experts under contract.
- Traces produced by models in the CSAB environments, and their corrections by experts.
What is never used
- The content of a Cogni6 vault.
- A client's attachment or conversation.
- A real deliverable without explicit rights.
- Any data whose provenance, licence or consent cannot be verified.
Training and execution are separate
What trains the model and what runs it on your engagement are two different things. The programme opens no new path for your data: the values in your files stay pseudonymised before any send, as they are today.
Data sovereignty
The programme's requirement is to keep the compute and the training data under Canadian jurisdiction. For a firm, the question is not only who reads the file, but where it is processed and under which laws.
This is a design requirement, not legal advice, and it does not replace the assessment your firm must make of its own obligations.
What will be published, and what does not exist yet
What we will publish
- The task schema and the evaluation harness.
- A public set of CSAB engagements.
- The rubric methodology and reproducible baselines.
- The limits and measurement gaps, including when the result is negative.
What does not exist yet
- No Prax6 model is published or downloadable.
- No performance measurement is claimed.
- The base model is not settled.
- The private set and the frontier rubrics will not be published, so the bench keeps its power to measure.
Contribute
The bench needs people who know what a signable deliverable looks like. That is the programme's main bottleneck, ahead of compute.
Lead consultants
Review the rubrics, say what a deliverable must contain before it can be shown to a client, and settle the critical failure cases.
Academic partners
Evaluation, judge calibration and publication of results.
Pilot firms
Fictional engagements built from the shape of real situations, with no client file involved.
Infrastructure
Post-training partners able to support training under Canadian jurisdiction.