Reproduce any
past state.
When a regulator asks why your firm took a position eighteen months ago, you don't reconstruct — you replay. Hand the auditor the manifest; they re-run every NodeVersion, agent, and LLM call to a bit-identical result. The killer feature for regulator-defensible AI.
Reproduce a decision from 18 months ago. Press ▶ — every NodeVersion, agent, and LLM call replays to a bit-identical result.
Live replay · press reproduce to walk the provenance chain to a bit-identical result.
Three ways to answer "why?"
A regulator asks why you took a credit position on an entity a year and a half ago. Your team has exactly three options — and only one of them survives the meeting.
Reconstruct by hand
Stitch the analysis back together from emails and notes. What most firms do today. Takes weeks. Comes out incomplete.
Doesn't scale
"Directionally correct"
Claim the model output was roughly right and hope the auditor accepts it. This is what banned ChatGPT output looks like.
Gets you banned
Hand over the manifest
The auditor runs `deino forensics reproduce <id>` and sees every NodeVersion, agent, LLM call, and threshold. Minutes, not weeks.
The only one that scales
Full reproduction
Re-run any historical query with bit-identical results. Same NodeVersions, agents, and LLM providers within their cache windows.
Time-travel queries
"What did Living Map look like for actor:84a2 on Nov 15, 2025?" — answerable instantly, not reconstructed.
Decision rationale
Not just the output — the reasoning trace. Which agents fired, what context they saw, what the sisters concluded.
Diff tooling
Compare two manifests side by side. See exactly what changed between two decisions on the same entity.
Export formats
JSON for machines, PDF for auditors, CSV for the compliance team's own analysis.
Cryptographic signing
Customer-key signed. Tamper-evident. Verifiable by your auditor independently of DEINO.
Before DEINO, every audit cycle was six weeks of evidence-gathering. Now we hand the auditor the manifest directly. Six weeks became six hours.
Compliance Officer · LATAM Tier-2 BankOne call, eighteen months back.
The same chain the replay above walks is the chain this call runs — and the one your auditor re-runs offline.
from deino import Client client = Client(api_key="dno_live_...") # "What was our exposure model on Nov 15, 2024?" past = client.forensics.reproduce(manifest_id="q_8c4d2a1f") print(past.distribution) # [0.62, 0.28, 0.10] — bit-identical print(past.node_versions) # [actor:84a2@v128, gov:ar-bcra@v89, …] print(past.agents) # ['a2a:profiler@v3.1', …] # Export for the auditor past.export(format="pdf", path="audit_2024_nov.pdf")
The rewind button for the whole platform.
Forensics replays anything the other products produced — because they all wrote to the same versioned ontology in the first place.
What is Backtest Forensics?
It lets you reproduce any past state of the ontology exactly. When a regulator asks why your firm took a position months ago, you don't reconstruct — you replay. Hand the auditor the manifest and they re-run every NodeVersion, agent, and LLM call to a bit-identical result.
How does an auditor verify a replay?
The manifest is customer-key signed and tamper-evident, so your auditor can verify it independently of DEINO. They run a reproduce command against the manifest id and see every NodeVersion, agent, LLM call, and threshold — in minutes, not weeks.
What export formats does it support?
Three: JSON for machines, PDF for auditors, and CSV for the compliance team's own analysis. You can also diff two manifests side by side to see exactly what changed between two decisions on the same entity.
See Backtest Forensics in production.
Founder-led demos for design partners. A free tier is active today for self-serve evaluation.