Platform / DELA
DELAThe underlying evidence changed. What does that affect?
DELA detects changes in scientific, operational or other source evidence, traces the affected dependencies, preserves the work that remains valid, selectively replays reproducible operations, and routes non-reproducible work to qualified human review.
Deterministic Evidence Lineage Architecture · AU provisional 2026907052
What it returns
Categorical, not probabilistic. Where a lane cannot decide, it says so explicitly rather than returning a low-confidence answer that reads like a decision.
What it does
- Compares versions of a source record on defined, typed attributes.
- Classifies each difference, so material and immaterial are separated by a rule.
- Separates affected downstream work from work that remains valid.
- Replays only the affected operations that are deterministically reproducible.
- Routes non-replayable work to a qualified reviewer instead of approximating it.
What it does not do
- It does not verify scientific or biological truth.
- It does not replace the system that produced the evidence.
- It governs the computational consequences of a change, nothing more.
The questions this layer gets asked in procurement
| Concern | Why it decides the deployment |
|---|---|
| Affected versus preserved accuracy | Over-flagging destroys the economics as surely as under-flagging destroys the safety case. |
| Replay reduction | The core commercial argument. Measured, not asserted. |
| Review workload | Whether the human cost of the exceptions is affordable. |
| Change definition | Agreed in advance with a domain expert, or the result means nothing. |
What runs inside DELA
Each of these is a real implementation with code behind it. Status is labelled honestly, and nothing is listed as coming soon.
BioChangeGate
Open sourceScientific state diffs, dependency impact, selective replay and sealed evidence. The reference implementation of the DELA layer.
FoldInputGate
Open sourcePre-flight checks for structured prediction job metadata before a run is accepted.
ConfidenceGate
Open sourceConfidence-contract routing for predicted structures. Turns a confidence band into a routing decision rather than a caveat.
ProvenanceGate
Open sourceProvenance continuity checks for scientific AI artefacts. Detects a broken chain between an artefact and its source.
ReplayGate
Interactive labRe-runs sealed evidence packs to detect verification drift over time.
See it against your own workflow
A pilot runs in shadow mode. It controls no production decision, and the acceptance criteria are agreed before it starts.