Skip to content
IllustrationA field of points along one beam. Decorative; it shows no data.

Faros · First-party model research

Research programme

Model research that shows its work

UpCube's own model research: preparing data, adapting models and testing them, shared once reviewed.

Status, as recorded

No Faros model has been released.

Faros is a research programme. Routing between third-party models is a separate Platform capability and is not model training.

EvidenceImplemented (offline engine + eval harness docs); exploratory pilot DOCUMENTED; no released model

On this page
  1. 01Programme
  2. 02Method
  3. 03What exists today
  4. 04Scorecard
  5. 05Claim gates
  6. 06Routing
  7. 07Publications
  8. 08Questions
  9. 09Notes

01Programme

What Faros is

Faros is UpCube's first-party model research programme: data preparation, adaptation, evaluation and, when evidence supports it, original training. No Faros model has been released. Early evaluation and execution-layer experiments have been run and are reported in the Research Lab when reviewed.

Four research tracks

  1. Track 01Stage F1

    Data

    Find and describe candidate data sources and record their rights status before any use. Discovered data is not licensed data.

  2. Track 02Stage F2

    Adaptation

    Fine-tune existing third-party checkpoints for specific kinds of work. The result is a fine-tuned third-party model and is labelled that way.

  3. Track 03Stage F2

    Distillation

    Study whether smaller models can learn useful behaviour from larger ones. Distilling from a third-party model does not produce an original model.

  4. Track 04Stage F3

    Evaluation

    Build evaluation harnesses and protocols before claiming results. Exploratory runs are not benchmarks.

02Method

Data to adaptation to evaluation

Three stages, then a gate. The gate stays closed until a reviewed, measured result exists.
Illustration
  1. F1

    Data

    Candidate sources are described and their rights recorded before any use. Discovered data is not licensed data.

  2. F2

    Adaptation

    Existing third-party checkpoints are fine-tuned or distilled. The result is labelled as derived from them.

  3. F3

    Evaluation

    Harnesses and protocols are written before results are claimed. Exploratory runs are not benchmarks.

  4. GateClosed

    Claim

    Only a reviewed, measured result, published in the Research Lab. None has been published for Faros.

Illustration: the intended Faros research flow. Data is described and rights-checked, used to adapt existing third-party models, and evaluated before anything is claimed. It is not a running system and shows no result.

03Ledger

What exists today

Each item is recorded in UpCube’s status evidence as of 31 October 2026. Nothing here is a release.
  1. 01

    An offline training and evaluation service exists in source: dataset handling, rights checks, an adaptation engine, evaluation and export.

    Evidence: Implemented in source; its tests are documented, not re-run for this page

  2. 02

    An offline adaptation smoke test on a small, randomly initialised model owned by UpCube is documented. It checks the plumbing; it does not produce a usable model.

    Evidence: Documented, not re-run

  3. 03

    Faros research documents (strategy, data catalogue and evaluation plans) exist on research branches.

    Evidence: Documented

  4. 04

    One bounded, exploratory, inference-only pilot is documented. It awaits human review, and its figures are not published here.

    Evidence: Documented; not independently verified

  5. 05

    The data catalogue lists candidate sources only. It records no approved downloads and no rights-cleared rows.

    Evidence: Documented

  6. 06

    No Faros model has been released, and no Faros work has been independently evaluated.

    Evidence: Registry status: Research programme

04Scorecard

The scorecard, as it stands

Where a research page would put results, this one shows what is missing. No cell holds a number because no Faros result exists.[5]
Recorded today[1]
What the status evidence lists for the track.
Measured result[2]
A reviewed measurement, published with its scope.
Independent evaluation[3]
A check of the work by someone outside the team.
Released model[4]
A model anyone outside UpCube can use.
Faros scorecard. Every result cell is empty because no Faros result exists.
TrackRecorded todayMeasured resultIndependent evaluationReleased model
DataStage F1Candidate sources catalogued. No approved downloads and no rights-cleared rows.Not measuredNot evaluatedNone
AdaptationStage F2Offline adaptation engine in source. A smoke test on a small, randomly initialised model is documented.Not measuredNot evaluatedNone
DistillationStage F2A research track. No distillation item is recorded in the status evidence.Not measuredNot evaluatedNone
EvaluationStage F3Evaluation code in source. One exploratory, inference-only pilot is documented and awaits human review.Not measuredNot evaluatedNone

05Claim gates

Five distinctions every Faros statement keeps

What we can say today, set against what we do not claim.
Faros claim gates
We distinguishWhat we can say todayWhat we do not claim
Discovered vs licensed dataCandidate data sources have been discovered and catalogued.That any of it is licensed, downloaded or rights-cleared. Each dataset and set of weights needs its own licence check.
Storage vs capacityStorage can hold data. It is not evidence of training capacity or of a trained model.Dataset sizes, storage totals or training capacity.
Infrastructure vs active runsTraining and evaluation code exists in source, and one exploratory pilot is documented.That training runs are active, or any pilot figure before human review.
Fine-tuned vs pretrainedAdaptation work starts from existing third-party checkpoints and is labelled as fine-tuning.An original pretrained Faros model.
Proposed vs measuredFaros-related publications in the Research Lab carry a classification and an evidence status.That a proposal, protocol or benchmark method is a result.

06Scope

Routing is separate

Routing is separate from training. Choosing which third-party model handles a request is a Platform capability, described with the Gateway and Model Intelligence.

Routing between models, or distilling from them, does not show that UpCube has an original model. Faros pages never present routing as training.

08Questions

Questions

  1. 01Has Faros trained its own model?

    No Faros model has been released, and this page makes no trained or original model claim. What exists is an offline training and evaluation service with documented tests, research documents and one documented exploratory pilot that awaits human review.

  2. 02Is discovered data licensed?

    No. Discovered data is not licensed, downloaded or rights-cleared data. The data catalogue records no approved downloads and no rights-cleared rows, and each dataset and set of weights needs its own licence check before use.

  3. 03Is model routing the same as Faros?

    No. Routing between third-party models is a Platform capability. It is described separately and is not evidence of model training.

Notes

  1. 1.Recorded today restates UpCube's status evidence for Faros (registry row A-15), snapshot 31 October 2026. The items are documented in source and in research documents; their tests were not re-run for this page.↩
  2. 2.No Faros measurement is published. The one exploratory pilot is inference-only and awaits human review. Exploratory runs are not benchmarks, so no figure appears on this page.↩
  3. 3.No Faros work has been independently evaluated.↩
  4. 4.No Faros model has been released. Adaptation starts from existing third-party checkpoints: a fine-tuned checkpoint is labelled fine-tuned, never original, and distilling from a third-party model does not produce an original model.↩
  5. 5.A cell changes only when a reviewed UpCube Research Lab publication, classified as a measured result, reports that measurement within its stated scope. Until then the cell says what is missing.↩

Disclosure

No Faros model has been released. Discovered data is not licensed data. Routing is not training. Fine-tuned third-party checkpoints are labelled as fine-tuned, never as original models. Exploratory results are not benchmarks, and none are published on this page.