CASE 01
Chip verification
A semiconductor team points Oscilla at timing closure. Counter-example search runs overnight; engineers arrive to a short list of real violations, each with its proof.
SIGN-OFF TIME
SIGN-OFF TIME, 6 MONTHS
OSCILLA-1 · REASONING MODEL & AGENT PLATFORM
Reasoning,
measured like
a signal.
Oscilla builds a reasoning model and an agent platform for engineering and science teams — available by API or private deployment.
SPECIMEN — LIVE READINGS, PUBLIC API
MODEL
Oscilla-1.2
CONTEXT WINDOW
512k tokens
MEDIAN LATENCY, FIRST TOKEN
0.84 s
REGIONS
14 · EU & US
OBSERVATIONS — CAPABILITIES
Four capabilities,
drawn to scale.
Long-horizon reasoning
Proofs and plans that run for thousands of steps without losing the thread — the model keeps its assumptions on the page, where you can audit them.
Verified tool use
Every tool call is checked against the plan before it runs and against its result after — a call that fails its check does not reach your systems.
Agent runs, with an audit trail
Agents that work for hours leave a record a regulator would accept: every step hashed, every source cited, every decision replayable.
Private deployment
The same model inside your VPC or on your own hardware. Your data crosses the boundary as queries — it never leaves as training material.
METHOD — THE PROTOCOL
Pose. Plan. Prove. Act.
-
01
Pose
The problem is stated as a claim with its assumptions listed — nothing enters the run unstated.
-
02
Plan
The model drafts a route and prices it in steps and tokens before spending either.
-
03
Prove
Each step is checked as it is produced. A step that cannot be proved is marked, not smoothed over.
-
04
Act
Only proved steps reach your tools. The whole run is written to the audit trail as it happens.
CONSOLE — THE PLATFORM AT WORK
One run, as recorded.
RUN 8F42TASK — CHIP VERIFICATION, BLOCK CWALL TIME 41.2 S
- 09:14:02 POSE claim: no path in block C violates setup time at 0.9 V
- 09:14:03 PLAN 6 steps · est. 38k tokens · tools: sta, spice, ledger
- 09:14:09 STEP 1 sta.run(netlist_c) → 212 paths under threshold [S1]
- 09:14:17 STEP 2 prove(path_881) → proved, margin 41 ps [S1][S2]
- 09:14:28 STEP 3 spice.sim(cor=slow, 125 °C) → margin holds, 38 ps
- 09:14:36 STEP 4 counter-example search → none found in 1.2 M cases
- 09:14:41 ACT report written to ledger · sign-off requested from M. Okafor
- VERIFY — PROOF CHECKED · 0 ERRORS · AUDIT HASH 9BE2…C41A
- TOKENS
- 36,204
- TOOL CALLS
- 14 · ALL CHECKED
- COST
- $0.31
- OUTCOME
- CLAIM PROVED
PLATE 01 — INTERLUDE
Every step checked. Nothing smoothed over.
MEASUREMENTS — AGAINST A BASELINE
Counted, not claimed.
| Measurement | Oscilla-1 | Baseline model | Method |
|---|---|---|---|
| Verified proofs completed | 96.4% | 71.2% | 4,208 problems, engineering & physics |
| Failed proofs shipped as correct | 3.1× fewer | — | Independent re-check of every “proved” result |
| Median time per verified proof | 41 s | 96 s | Same tasks, same tool access |
| Median cost per verified proof | $0.31 | $0.58 | List prices, November 2026 |
| Steps with a citation or a check | 100% | 63% | Audit-trail inspection, 200 runs |
DEMO DATA
3.1×
THE KEY NUMBER — FAILED PROOFS
Fewer failed proofs shipped as correct than the baseline model, across 4,208 engineering and physics problems — every “proved” result independently re-checked.
SEE A RUN, CHECKED LIVE →CASE NOTES — WHERE IT WORKS
Four benches, four results.
CASE 02
Materials discovery
A lab screening battery cathodes uses agent runs to plan and check 1,900 simulations, discarding the ones whose assumptions fail before they consume compute.
CANDIDATES TO SYNTHESIS
CANDIDATES REACHING SYNTHESIS, PER QUARTER
CASE 03
Grid planning
A transmission operator tests connection offers against stability constraints. Every accepted plan carries the proof a regulator will ask for.
STUDIES PER WEEK
STUDIES PER WEEK, 12 MONTHS
CASE 04
Drug-safety review
A pharmacovigilance group checks interaction claims against the literature. Claims without a citation are returned unmarked — visibly unproved, never quietly dropped.
CLAIMS CHECKED
CLAIMS CHECKED PER QUARTER
CONTAINMENT — SECURITY & DEPLOYMENT
Your boundary, kept.
API
Hosted in 14 EU and US regions. Data in transit and at rest is encrypted; prompts are retained for 30 days for abuse review, then deleted — or not retained at all, on request.
VPC
A dedicated deployment inside your cloud tenancy, with your keys, your logging and your network rules. Model updates arrive signed, and you pin the version.
ON-PREMISES
Air-gapped installations for regulated work. The audit trail stays on your hardware; support is by signed bundle, never by remote access.
- SOC 2 TYPE II — UNDER NDA
- ISO/IEC 27001
- GDPR · UK GDPR
- NO TRAINING ON CUSTOMER DATA
- ANNUAL PENETRATION TEST, PUBLISHED SUMMARY
PEER REVIEW — REPORTS ON FILE
Reviewed like a paper.
REVIEWER 1 · VERDICT — ACCEPT
“The audit trail is the first I have seen that a verification team can actually re-run. We checked eleven runs line by line. Eleven held.”
REVIEWER 2 · VERDICT — ACCEPT
“It marks the steps it cannot prove instead of hiding them. That single habit changed how much of our screening we let it run unsupervised.”
REVIEWER 3 · VERDICT — ACCEPT WITH NOTES
“Note: the model is slower than a chat model and says so. For planning studies, the proof attached to each answer is worth more than the speed.”
PAPERS — FROM THE LAB NOTEBOOK
Read the research.
PAPER · OCT 2026
Proof-carrying inference: checking each step as it is produced
How Oscilla-1 attaches a check to every reasoning step, and what that costs in time and tokens.
FIG. 07 — STEPS CARRYING A CHECK, % OF ALL STEPS
READ THE PAPER →NOTE · SEP 2026
The audit trail as an interface
Agent runs are long. We describe the record format that lets a second team replay any run exactly.
FIG. 08 — REPLAY TIME PER RUN, AS RECORDS GREW
READ THE NOTE →PAPER · AUG 2026
Counter-example search at 1.2 million cases
Trying to break a proved claim is part of proving it. Methods and failure modes from chip verification.
FIG. 09 — SEARCH DEPTH BEFORE THE FIRST COUNTER-EXAMPLE
READ THE PAPER →QUESTIONS — ASKED AT THE BENCH
Before you ask.
Each reasoning step is checked as it is produced — arithmetic re-computed, claims matched to their citations, tool results compared against the plan. A step that fails is marked in the output and in the audit trail, never silently repaired.
Yes. Lab and Institute plans include VPC deployment; Institute adds air-gapped on-premises installation with signed updates and a pinned model version.
No. Customer prompts, documents and run data are never used for training, on any plan. API retention is 30 days for abuse review and can be set to zero.
Oscilla is slower and says so. It is built for work where an answer must carry its evidence: proofs, plans and agent runs that another team — or a regulator — can check afterwards.
Three tiers — Explorer, Lab and Institute — plus per-token usage. The full table, including the comparison and per-million-token prices, is on the pricing page.