PL Book a demo

Obsigen GRID · Artifact GRID release · system card · as of 2026-10-22 · Visera P.S.A.

Obsigen GRID

The layer that selects models for each task based on measurements and checks the result before it reaches the user. Coming soon: the Artifact GRID release - a shared, verifiable state of knowledge about the task.

How it works

A task - apart from simple questions - goes through GRID: a small model recognises the type of task, a table filled with measurements selects the model and the add-ons, and in Titan mode the result goes through verification before it is delivered - a check of sources and completeness and, for code, a run. The table changes only on the basis of measurements; an entry without evidence is not applied.

When a model is available from several providers, Obsigen keeps a queue for it: if the first one does not respond, it switches to the next one. Measured on product photos: the fallback provider did not respond for 60 s1, the next one in the queue responded in 1.7 s2; reordering the queue is one configuration entry, with no change to the product.

How Obsigen picks the model for a task - and for every subtask A small model classifies the task; a table filled by measurements picks the model; the model does the work and, if the task is larger, splits it into subtasks - each goes through the same cycle, nested too. When the table adds extras - as in Titan mode - a second model reviews the answer and critics check sources and completeness; only what is contradictory is corrected, in a loop. Simple questions are answered straight away by a single model. The result carries a chip: which model, cost, duration. the same cycle for the task and every subtask task question, file,photo classification Obsigen GRID model choice Obsigen GRID execution Obsigen GRID reviewer Obsigen GRID critics sources ·completeness subtasks 1 … n, nestedsame cycle result chip: who · cost ·how long correction: only what is contradictory 1 2 3 4 5 6 7 8 How Obsigen picks the model for a task - and for every subtask A small model classifies the task; a table filled by measurements picks the model; the model does the work and, if the task is larger, splits it into subtasks - each goes through the same cycle, nested too. When the table adds extras - as in Titan mode - a second model reviews the answer and critics check sources and completeness; only what is contradictory is corrected, in a loop. Simple questions are answered straight away by a single model. The result carries a chip: which model, cost, duration. task question, file,photo classification Obsigen GRID model choice Obsigen GRID execution Obsigen GRID subtasks 1 … n, nestedsame cycle reviewer Obsigen GRID critics sources ·completeness result chip: who · cost ·how long 1 2 3 4 5 6 7 8
  1. classification - a small model recognizes the task type: code, document, spreadsheet, photo, question.
  2. model choice - a table filled by measurements picks the model and extras: reviewer, plan preview, Normal → Genius escalation. An entry without evidence is not used.
  3. execution - the chosen model works; a larger task is split into parts.
  4. subtasks - each part goes through the same cycle - classification, model choice, execution - and may have its own subtasks; results return to the parent task.
  5. reviewer - a second model checks the first.
  6. critics - check sources, completeness and the premise of the question.
  7. correction - only what is contradictory is corrected - the loop may repeat; every pass counts toward the cost of the turn.
  8. result - the answer with a chip: which model worked, what it cost, how long it took.

Measurement

A measurement of the selection and verification layer: the same model with GRID orchestration and without it.

We measured it - the same model with Obsigen GRID and without it

measured September 12–13, 2026

Was the work delivered at all?

Runs that ended with a working result

Obsigen GRID 12 of 153 Model alone 4 of 153

When both variants deliver a result, the scores are close: 19.6 vs 18.9 / 203.

Tasks passed

Very hard tasks, full scoring rubric

0 25 50

Model alone 7%3Obsigen GRID 47%3

Cost of one passed task

Index: model alone = 100. Model cost in the measurement, not a customer price

Model alone1003 Obsigen GRID≈ 263

≈ 3.9× cheaper3 - even though a single run with Obsigen GRID costs 1.8×3 more.

Time per run

Index: model alone = 100. We include it because without it the rest would not be honest

Model alone1003 Obsigen GRID≈ 1903

1.9× slower3 - checking the result takes time.

Source: internal Obsigen measurement, September 12–13, 2026 · the same model in both variants · 5 very hard programming tasks × 3 repetitions (small sample) · difference in mean score: p = 0.0023. Tasks passed: 7% → 47%3.

The difference is not in how the model writes - it is in whether the work gets finished. On easy tasks we measured no significant difference.

Coming soon: Artifact GRID

coming soon

Model selection and verification are the first layer of GRID. We are developing the next one - the Artifact GRID release - as a shared, verifiable state of knowledge about the task: models add hypotheses, artifacts and test results to it through references, without passing content to one another. Every conclusion has a provenance, and verification decides what is true. Details will be published on launch day.

ARTIFACT GRID · STATE OF KNOWLEDGE IN YOUR TENANT MODEL α MODEL β MODEL γ TEST · RUN art·0912 specification.pdf#p12 reference art·0417 orders.ts#L184 supported run·0044 tests: passed verified wn·07 counterexample falsified wn·08 · wn·09 depend on wn·07 retracted ref art·0912 → node (1,2,2)run·0044 ✓ → wn·03 verifiedwn·07 ✕ → retraction: wn·08, wn·09

Visualisation · artifact codes are illustrative

Identity and version

Obsigen GRID is part of the Obsigen AI system - an agent that integrates open-weight models. Under the AI Act, Obsigen is an AI system, not a model. Provider: Visera P.S.A.

  • Model selection based on measurements available
  • Result verification before delivery available
  • Artifact GRID - a shared state of knowledge about the task in testing

Composition

Open-weight models run by providers in the EEA; a new model joins the composition only after measurement. The full composition with licences and a detailed description of GRID are part of the documentation shared during deployment talks, once an NDA is signed.

Intended use

Intended use: work on company documents, data and code - analyses, reports, proposals, presentations, reading technical documentation, code changes with tests. Results are approved by a person.

Out of scope: creating advertising creatives, decisions made without human involvement and high-risk uses under Annex III of the AI Act, such as recruitment or employee assessment.

Data

We develop GRID and the selection table exclusively on our own test scenes. Customer data - files, conversations, memory, the company glossary and the document index - is stored exclusively in your tenant and does not go into tests or GRID training. Obsigen does not train models on customer data.

Limitations

  • GRID does not give models knowledge they do not have - it organises and checks what they bring;
  • verification takes time: work with GRID takes longer than a single model answer;
  • models can make mistakes - verification and running the code reduce them but do not eliminate them; that is why results are approved by a person;
  • illegible fragments of documents are marked as illegible, not filled in;
  • measurement results apply to the configurations described in the reports - on your tasks, they may differ.

Security and Article 50

More: system architecture and marking of AI-generated content (Article 50). Changes are published in the changelog, and GRID documentation for IT and security teams is shared during deployment talks, once an NDA is signed.

Scope of claims. Results apply to our tests and to the configurations described in the reports; we do not publish comparisons on external benchmarks before we have run them. Development versions have a status, not a date.

Demo · 45 min

See Obsigen on your own example

We show Obsigen live, online, on a document you bring - with no confidential data. Then a quote and deployment on dedicated infrastructure.

Book a demo Let's talk

  1. step 1Demo · 45 minonline, on your example
  2. step 2A one-page quotewithin two business days
  3. step 3Deployment and team starta dedicated tenant and team accounts

Footnotes

  1. product photo measurement 2026-09-08 §9a · the fallback provider did not respond - queue of providers for the same model
  2. product photo measurement 2026-09-08 §9b · the next provider in the queue responded; the queue is reordered with one config entry, no product change
  3. summary report 2026-09-13 · runs that ended with a working result - Obsigen GRID