PRODUCT — MODEL & PLATFORM
The model and
the platform.
Oscilla-1.2 does the reasoning; the platform runs it as checked, replayable work — agent runs, tool use and audit, behind one API.
PART 01 — THE MODEL
Oscilla-1.2
A reasoning model trained to work like an engineer with a notebook: state the claim, plan the route, check each step as it lands. It reads text, code, tables and instrument data, and it writes answers that carry their evidence with them.
- Step-level verification — every step checked as produced; failures are marked, never hidden.
- Long horizon — plans and proofs of thousands of steps, assumptions kept on the page.
- Citations by construction — claims point at sources; a claim without one is returned unproved.
- Calibrated refusal — when the evidence runs out, the model stops and says which step failed.
SPECIMEN — MODEL CARD, ABRIDGED
| Context window | 512k tokens |
| Inputs | Text, code, tables, instrument CSV/JSON |
| Outputs | Answer + proof record (steps, checks, citations) |
| Median latency, first token | 0.84 s |
| Regions | 14 · EU & US, or your boundary |
| Version policy | Pinned per deployment; updates are opt-in |
FULL MODEL CARD SHIPS WITH ACCESS · DEMO DATA
PART 02 — THE PLATFORM
Runs you can replay.
- 11:02:11 POSE claim: cathode candidate set can be cut from 1,900 to under 50 without losing a stable chemistry
- 11:02:13 PLAN 9 steps · est. 210k tokens · tools: sim, ledger, search
- 11:02:40 STEP 1 screen(stability) → 412 candidates pass the first gate [S1]
- 11:03:58 STEP 2 prove(gate-1 sound) → proved against 3 published datasets [S2][S3]
- 11:04:31 STEP 3 sim.run(412) → 61 flagged for assumption failures — discarded, with reasons
- 11:05:47 STEP 4 counter-example search → 2 candidates removed at the boundary conditions
- 11:06:55 STEP 5 rank(349) → shortlist of 44, each with its evidence chain
- 11:08:02 ACT shortlist + full record written to ledger · review requested from S. Marchetti
- VERIFY — PROOF CHECKED · 0 ERRORS · AUDIT HASH 41D7…9B02 · REPLAYABLE
READ IT AS — FOUR HABITS OF THE CONSOLE
Steps
Each step states what it used and what it changed. Nothing happens between the lines.
Tools
Calls are checked before they run and after they return. A failed check stops the run, visibly.
Citations
Bracketed sources resolve to the exact document, dataset or prior step the claim rests on.
Verify
The last line is a second pass over the whole run. If it cannot sign, the run is marked unproved.
API — FIRST RUN IN FOUR LINES
Ask for a proof, get a record.
SAMPLE — CURL
curl https://api.oscilla.example/v1/runs \
-H "Authorization: Bearer $OSCILLA_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "oscilla-1.2",
"task": "Prove no path in block C violates setup time at 0.9 V.",
"tools": ["sta", "spice", "ledger"],
"verify": true
}'
# → 202 · run 8F42 · the response carries the run id;
# GET /v1/runs/8F42 returns the steps, the checks,
# the citations and the verify line — the whole record.
INTEGRATIONS — WHERE RUNS PLUG IN
Tools the platform can hold.
| Name | Kind | Access | Status |
|---|---|---|---|
| Your simulators & solvers | Tool | Checked calls | Via tool SDK |
| Git hosting (GitHub, GitLab) | Source | Read / propose | Built in |
| Issue trackers (Jira, Linear) | Record | Write-back | Built in |
| Data warehouses | Source | Read-only | Built in |
| Object storage (S3-compatible) | Record | Audit archive | Built in |
| Chat (Slack, Teams) | Notify | Run summaries | Built in |
| SSO (SAML, OIDC) | Identity | Sign-in | Lab & Institute |
| Kubernetes | Deploy | Helm chart | VPC & on-prem |
EVERY INTEGRATION INHERITS THE SAME RULE: CALLS ARE CHECKED BEFORE THEY RUN.
DEPLOYMENT — YOUR BOUNDARY, KEPT
Three ways in.
API
Hosted in 14 EU and US regions. Thirty-day prompt retention for abuse review, or zero retention on request.
VPC
Inside your cloud tenancy with your keys and logging. Model version pinned; updates arrive signed.
ON-PREMISES
Air-gapped installation for regulated work. The audit trail never leaves your hardware.
