Research pathway
Native ERAIS Model Research
Test one fixed ERAIS model against an agreed baseline and success criteria. Findings apply only to the task and configuration tested.
Your starting point
Would this help?
This is a fit when
- A bounded domain-learning or architecture question with a qualified baseline
- A partner who values a decisive negative result and rigorous run/recovery evidence
When another approach is a better fit
- A request for an Australian ChatGPT, frontier pretraining, or a production assistant
- A sponsor requiring a positive quality or efficiency outcome
What we’re planning
Proposed outcomes.
- 01Hypothesis, protocol, baseline, and promotion contract
- 02Environment, data, rights, model, and compute manifests
- 03Run, failure, pause, resume, and recovery ledger
- 04Quality, failure, generation, and resource analysis
- 05Independent review and promote, redesign, retain-as-tooling, or stop record
Where this stands
What is known. What needs testing.
Run only through a qualified NF-C protocol with lawful rights, isolated holdout, useful-quality threshold, bounded compute, repeated result plan, and independent review.
Existing basis
- Two governed native-byte training arms each completed 262,400 updates and 536.3M valid byte-level training targets on an ordinary four-thread CPU
- Exact checkpoint continuation, causal execution, fixed-panel evaluation and adverse-result retention
- The richer cognitive stack followed a materially different learning trajectory but did not earn quality promotion on the preregistered fixed panel
Evidence still needed
- Useful held-out or open-ended quality against qualified dense controls
- Stored-capacity and active-path matched comparisons with repeated independent evaluation
- Quality-per-compute, sample-efficiency and matched-quality resource evidence
- Safe measured production training and serving envelopes
Working together / 8–16 weeks
How we’ll work together.
Fix the model and task, set aside examples for evaluation and agree the comparison and compute budget. Repeat the test and have another reviewer check the findings.
The project steps in detail
- 01Qualify through NF-C.
- 02Freeze the native profile, baseline, holdout, compute, and all-outcome decision tree.
- 03Execute bounded scouts and confirmatory runs as declared.
- 04Review quality, failures, recovery, resources, and uncertainty.
- 05Promote only exact supported wording—or redesign/stop.
What we need before work starts
Before starting
- One falsifiable domain question and qualified baseline
- Lawful data/model/software rights and isolated holdout
- Frozen compute, rerun, stopping, and publication rules
- Independent method/result review
Scope and handover
What we agree before starting.
We agree the work in a signed statement of work before starting. This page explains the offer; the signed agreement sets the commitment.
How you accept the handover
- The frozen protocol, profile, data rights, holdout, baseline, and compute ceiling are preserved.
- Required runs, stops, failures, recovery, and resource evidence are complete.
- Independent review records unresolved objections.
- Delivery acceptance remains separate from any positive claim promotion.
What this work covers
Sponsored falsifiable research, normally delivered through a qualified Research & Benchmark Partnership.
One native profile, domain question, baseline, rights set, holdout, and compute ceiling. No operational checkpoint or general-model claim follows.
Completed native training and a changed learning trajectory do not establish dense-quality parity, useful open-ended generation, sample efficiency, quality-per-compute superiority, local full-weight viability, or production readiness.
Offer source and review record
Talk to us
What are you trying to find out?
Tell us the question, what you’ve tried and how you’d use the answer. We’ll see whether we can help investigate it.