Hugging Face
Public datasetInvestment banking
BankerToolBench
Complete investment-banking workflows: task framing, data gathering, modelling, and review-ready output, scored against practitioner-written rubrics.
The machine behind decision-grade agents.
Category
A specialised machine for your firm.
Claude and Codex provide the platform. ogram sits on top as a tailored machine for the work your firm actually does.
It is shaped around your documents, your standards, and your review process. Long-running, auditable, decision-grade.
Not another chat window. Not a generic assistant.
External benchmarking
Measured against real professional work.
ogram edge is calibrated against public benchmarks built by independent practitioners: one for investment banking, one for law. They test complete work, from task framing to reviewable deliverable, under expert rubrics and institutional-grade standards.
BenchmarksBankerToolBenchLegal Agent Benchmark
Harvey
Open benchmarkLegal work
Legal Agent Benchmark
Full legal matters scored under an all-pass standard. A task passes only when every rubric criterion passes: facts, conclusions, citations, structure, and analysis.
What LAB reveals
Harvey LAB update · May 28, 2026 ↗10.4%
current top LAB all-pass score for end-to-end legal tasks under Harvey's strict scoring standard.
Even the newest frontier model completes roughly one task in ten. Professional work is far from saturated.
LAB leaderboardRelative all-pass scale
Claude Opus 4.8
10.4%
Claude Opus 4.7
7.1%
Claude Sonnet 4.6
5.4%
Claude Opus 4.6
4.2%
GPT-5.5
2.1%
Gemini 3.5 Flash
0.8%
Harvey's initial LAB results reported frontier models on a holdout set reflecting LAB's task distribution. The May 28 update adds Opus 4.8 at 10.4%, the first score above 10% under the all-pass standard.
The standard ogram edge targets
ogram edge places reliability infrastructure between frontier models and the deliverable: bounded search, verified execution, controlled review, and traceability back to the source paragraph. Against these benchmarks, performance is measured. It is not declared.
Hugging Face and Harvey are marks of their respective holders, cited here as references to public benchmarks. LAB ranking data: Harvey, Initial Results on Legal Agent Benchmark, May 26, 2026, and Opus 4.8 update, May 28, 2026.
How it runs
From the working platform to the ogram machine.
The partner keeps working in Claude or Codex. When the question needs specialised depth, ogram runs the dedicated machine and returns the answer.
- Step 01Partner
Work stays in place
The partner works inside Claude or Codex, in the language and rhythm they already use.
- Claude
- Codex
- Partner seat
- Step 02Call
Specialised request
When the task needs depth, the platform sends a bounded call to the ogram machine built for that firm and workflow.
- Scoped task
- Source perimeter
- Answer contract
- Step 03ogram VM
Long-running execution
The dedicated machine runs the specialised work: reads the authorised sources, checks the reasoning, and keeps the task recoverable.
- Verified
- Traceable
- Checkpointed
- Step 04Answer
Precise return
The answer comes back into the partner's working surface with the reasoning, sources, and limits intact.
- Precise answer
- Source lineage
- Ready to use
The seven failure modes
Agentic AI breaks where the stakes are highest.
Hallucination, memory loss, context rot, agent drift. These are not edge cases. They are structural failure modes that make current harnesses unsuitable for investment-grade work. ogram addresses each one architecturally.
01
Hallucination
Structural verification and source-grounding at every inferential step. Every number, every citation, traceable to its origin.
02
Memory loss
Persistent state management across extended agent sessions. Nothing is forgotten between the start of the diligence and the final memo.
03
Context rot
Active monitoring and remediation of context degradation over long horizons. The agent that finishes is as sharp as the agent that began.
04
Compaction loss
Preservation of critical information when context windows compress. The facts that matter survive. The noise does not.
05
Agent drift
Continuous alignment between agent behaviour and the original objective. No wandering. No scope creep. No polite deviation.
06
Interruption
Checkpoint, recovery, and resumption of multi-hour workflows. A crashed runtime does not mean a lost afternoon.
07
Orchestration
Coherent coordination of specialised sub-agents working parallel workstreams. One plan, many hands, one output.
Evidence on paper
Every claim leaves an artefact.
Source citations. Long-context traces. Checkpoint logs. The scaffolding produces documents the way a reliable analyst does — because investment-grade work must survive a review.
Source GroundingClaim anchored
The Acquisition Committee notes that consolidated net revenue for the fiscal year ending 31 December 2023 reached CHF 2,341.8 million, representing a year-on-year increase of 7.4% against the prior period. Adjustments under Schedule 4.7(b) reduce normalised EBITDA to CHF 412.6 million.
SHA_Project_Helvetia_v14.pdf
Schedule 4.7(b) · p. 184 · ¶ 3
Exhibit
Extracted passage
“…consolidated net revenue for the fiscal year ending 31 December 2023 totalled CHF 2,341.8 million, as independently verified pursuant to Clause 9.3 of the Share Purchase Agreement and confirmed by the statutory auditors' report…”
01 · Source grounding — claim & anchor
Long-Context FidelityRunning · 5h 33m
Diligence corpus · Project Helvetia · 487 documents · started 09:14 CET
312 / 487
Drift
0.00
Compaction
14
events, recovered
Sub-agents
7
aligned
Started 09:14:02 CET · 2024-03-0714:47:19 CET · elapsed 5h 33m 17s
02 · Long-context fidelity — coherence over 487 documents
Fault Recovery2024-03-07 · agent.log
14:31:08
Checkpoint 13/22 written
state 4.2 GB · SHA-256 a3f8c2d1
14:31:47
Runtime interrupted
cause: upstream timeout · exit code 124
14:32:14
State restored from checkpoint 13/22
integrity verified · delta 0 bytes
14:32:14
Execution resumed · agent state verified
interruption window: 27s · no re-derivation required
A crashed runtime is not a lost afternoon.
03 · Fault recovery — checkpoint & resume
Built for work that continues after the first answer.
Investment-grade AI is not a single generation problem. It is a control, memory, and verification problem carried across the full life of the case.
Discuss your workflow→