AlphaIQ

AlphaIQ validation platform

Check any model

Bring any model — built in-house, bought from a vendor, or open source. Add it, upload your files, run the checks, and get the report your reviewer needs.

  1. Step 1

    Register the model

  2. Step 2

    Upload evidence

  3. Step 3

    ETL

  4. Step 4

    Validation metrics

  5. Step 5

    Report

Step 1

Add the model you want to check

Pick the type of model (its "family") — this decides which checks and report templates you get.

More detail on how ownership changes what can be checked.

Who owns it

Model family

Name the model to open the rest of the workspace — or load the sample engagement for this family.

Step 2

Upload the evidence

0 files

Drop documentation, data, evaluation traces or source code

PDF · DOCX · CSV · Parquet · XLSX · JSONL · PY / R / SAS — files stay in your browser in this demo

No files yet — drop your own, or load the sample engagement to walk the flow.

Step 3

Clean and load the data (ETL (extract, transform, load: pulling in data, cleaning it up, and loading it for checks))

This runs the same 8 steps for every model. What changes is what counts as "one record" for your family.

One record is one obligor, identified by `obligor_id`, read by the eight-stage pipeline.

  1. 1Detect — Checks what kind of file you uploaded, reads it carefully, and gives it a unique fingerprint so we can always tell if it changes later.
  2. 2Parse — Breaks the file into rows. The first row holds the column names, and every row after it gets checked, even the empty ones.
  3. 3Map — Matches your columns to what the model needs. If a column name matches exactly, it's linked automatically. If not, a person has to confirm the match before anything else happens.
  4. 4Cast — Makes sure each column holds the right kind of value, like numbers or dates. If a value is unclear, we stop and flag it instead of guessing. A blank is filled in only when the model's own rules say what to put there, and every fix is written down.
  5. 5Dedup — Looks for the same customer appearing twice. If the two copies disagree, the file is rejected. If they're identical, it's rejected too, unless someone says in writing that it's fine.
  6. 6Normalise — NOT BUILT. Not built yet. This step would convert units (like dollars to euros), but a spreadsheet doesn't say what units its columns use, so we don't guess.
  7. 7Split — Sorts the clean rows into groups: what goes into the model, what comes out, and the decisions it makes.
  8. 8Write — Saves the final data as a locked version that can't be changed. Every number in the report can then point back to the exact rows it came from.

Step 4

Validation metrics

Credit & capital risk metric pack. Drop a scored population (CSV/TSV with a score column and a binary outcome column) and these are computed for real, in your browser; without one, the pack below is the pack definition and carries no values.

No scored population supplied — nothing below is computed. This is the metric pack credit & capital risk models are judged against, with the thresholds each one carries. Drop a CSV with a score column and a binary outcome column and these are computed for real, in your browser.

Out-of-sample AUC——
KS statistic——
Brier score——
Calibration (observed / predicted)——
PSI (population stability)——
Severe-scenario PD——
Calibration bias (mean predicted PD − observed default rate)——

Step 5

Produce the report

Templates relevant to credit & capital risk are listed first; every figure in the output links back to the code fingerprint, seeds and citations from this run.

Uploaded files stay in your browser — nothing is sent anywhere. For a family that declares a metric pack, discrimination, calibration, stability and capacity metrics are computed for real from a scored population file; for a family that declares none, the file is read and reported and not scored. Everything requiring the model runtime is marked as such.

What runs today

9 of 12 model types actually run checks on your data

Registering a model, uploading evidence, and building the final report work the same way for every model type below. The actual number-crunching is being added one model type at a time.

FamilyMetric packSimulation layersAvailable today
Credit & capital riskSR 26-2 · ECB TRIM · IFRS 9 / CECL · Basel IRBRunningStress + boundaryDiscrimination, calibration and stability computed from a scored population; macro stress and boundary testing simulated on the replicated model; report generated with a reproducibility manifest.
IRB rating systemsECB TRIM · EBA/GL/2017/16 · CRR (Regulation 575/2013, as amended by CRR3 — Regulation (EU) 2024/1623) · SR 26-2RunningNot needed for this model typeThe declared metric pack is computed from the evidence it names, with every threshold citing where it comes from. No simulation layer applies to this family, so nothing is stressed on a replica, and no validation memo is produced for it today.
Market risk & valuationFRTB · SR 26-2 · Basel MAR32 backtestingRunningNot needed for this model typeThe declared metric pack is computed from the evidence it names, with every threshold citing where it comes from. No simulation layer applies to this family, so nothing is stressed on a replica, and no validation memo is produced for it today.
AML & financial crimeFFIEC BSA/AML Examination Manual · OCC 2026-13 · WolfsbergRunningNot needed for this model typeThe declared metric pack is computed from the evidence it names, with every threshold citing where it comes from. No simulation layer applies to this family, so nothing is stressed on a replica, and no validation memo is produced for it today.
Fraud & abuse detectionSR 26-2 · Fair lending / adverse action · PSD2 SCA RTS — Commission Delegated Regulation (EU) 2018/389RunningNot needed for this model typeThe declared metric pack is computed from the evidence it names, with every threshold citing where it comes from. No simulation layer applies to this family, so nothing is stressed on a replica, and no validation memo is produced for it today.
Actuarial & insuranceASOP 56 · IFRS 17 · Solvency IINot yet plannednot yet scheduledComing laterRegister the model, run the family's ETL pipeline, record evidence with hashes, and generate the report skeleton with its regulator templates. No metric pack and no evidence register is declared for this family yet, so nothing here says what a review of it must evidence.
Insurance — claims severityASOP 56 · NAIC AI Model BulletinRunningComing laterThe declared metric pack is computed from the evidence it names, with every threshold citing where it comes from. No simulation layer applies to this family, so nothing is stressed on a replica, and no validation memo is produced for it today.
Machine learning systemsEU AI Act · NIST AI RMF · SR 26-2 · ISO/IEC 42001RunningNot needed for this model typeThe declared metric pack is computed from the evidence it names, with every threshold citing where it comes from. No simulation layer applies to this family, so nothing is stressed on a replica, and no validation memo is produced for it today.
GenAI, LLM & agentsEU AI Act · NIST AI RMF / GenAI Profile · FINRA / SEC guidanceNot yet plannednot yet scheduledNot needed for this model typeRegister the model, run the family's ETL pipeline, record evidence with hashes, and generate the report skeleton with its regulator templates. No metric pack and no evidence register is declared for this family yet, so nothing here says what a review of it must evidence.
Interest rate risk in the banking bookDirective 2013/36/EU Article 98(5) · EBA/GL/2022/14 — IRRBB and CSRBB · Commission Delegated Regulation (EU) 2024/856 — supervisory outlier tests · Circular BCB 3.876/2018 — IRRBB (Brazil) · Resolução CMN 4.557/2017 — risk management (Brazil)RunningComing laterThe declared metric pack is computed from the evidence it names, with every threshold citing where it comes from. No simulation layer applies to this family, so nothing is stressed on a replica, and no validation memo is produced for it today.
Funding and liquidity assumptionsCircular CSSF 07/301 — ICAAP · Circular CSSF 20/753 — ICAAP, ILAAP and stress testingRunningComing laterThe declared metric pack is computed from the evidence it names, with every threshold citing where it comes from. No simulation layer applies to this family, so nothing is stressed on a replica, and no validation memo is produced for it today.
Simulation & scenario enginesSR 26-2 · CSRD / climate disclosure · Internal policyNot yet plannednot yet scheduledComing laterRegister the model, run the family's ETL pipeline, record evidence with hashes, and generate the report skeleton with its regulator templates. No metric pack and no evidence register is declared for this family yet, so nothing here says what a review of it must evidence.