Outcome 01
Observed result divergence
The available evidence shows where the reconstruction differs from the claim. The note records the divergence point, supporting artifacts, and what could close the gap.
Fixed-fee entry review
Model Preflight.
A fixed-scope, fixed-fee first read of one model. We attempt to reproduce the claimed result from your code, in your environment, and report what the available evidence supports, what remains unsupported, and what to check next.
Written engagement and NDA before non-public material. No invoice is issued until the short, pre-agreed environment smoke test passes.
Engagement profile
- Fee
- $4,500 fixed for the published scope, subject to a signed engagement agreement. Invoiced after environment acceptance; scheduling is confirmed on payment.
- Timeline
- 5–10 business days from payment, with your environment accepted first.
- Output
- A written findings note (up to five principal findings) with a one-page summary written to be forwarded to a committee, evidence from the reproduction attempt, and a 60-minute debrief.
- Best for
- Investment teams that need an independent first read before capital, a committee, or further diligence depends on a model.
Founder-led validation
Evgenii Azarov
Founder · Direct technical reviewer from scoping through handoff
- Econometrics · statistics · machine learning · financial engineering
- PhD in Law · Russia, 2012 · finance training at NYU
- 20 years of private teaching and consulting · New York · global
Illustrative · synthetic data · not client work · not a real engagement · not a track record
Model & Backtest Audit — Findings Memo
A material part of the headline does not survive point-in-time reconstruction.
A deliberately flawed synthetic strategy is rebuilt on point-in-time inputs so the decision-maker can see what survives.
| Metric | Reported | Reproduced |
|---|---|---|
| Sharpe (annualized) | 1.8 | 0.7 |
| Annual return | 14% | 6% |
| Max drawdown | −9% | −21% |
| Hit rate | 58% | 52% |
ConfirmedHigh
Look-ahead in the signal pipeline
A fundamental feature was joined on report date rather than its as-available date.
Bounded first read
What Preflight covers.
One model, one pinned repository state, one data snapshot, one claimed headline result, and one agreed execution path. The scope is deliberately narrow enough to produce a useful written record quickly.
-
01
Reproduction attempt
We attempt one claimed headline result from your code at a pinned repository state, using one data snapshot and a working environment you provide — including compute, licenses, and credentials.
-
02
Timing and leakage checks
We inspect look-ahead exposure, target leakage, and train/test contamination — common failure modes that can invalidate reported results.
-
03
Validation design review
We compare the split scheme, cross-validation logic, and evaluation-period hygiene with the methodology you describe.
-
04
Up to five principal findings
Each finding states severity, supporting evidence, and decision impact. If fewer material issues exist, that is the result — findings are never padded.
-
05
Preflight Findings Note
A short written note, typically three to six pages, designed to be inspectable by both a technical owner and a decision-making committee. The first page is a standalone summary — decision, principal findings, and what remains unresolved — written to be forwarded as is.
-
06
60-minute debrief
We explain what the evidence supports, what remains unresolved, what to check next, and what a full validation would add.
First-read failure modes
What a first read most often catches
- A feature joined on its report date rather than the date it was actually available — invisible in the equity curve, decisive in diligence.
- The target, or a transform of it, leaking into the training features.
- Train/test contamination through overlapping windows or preprocessing fitted on the full sample.
- A validation design that cannot support the claim: random splits on time series, or tuning on the evaluation period.
- An evaluation window chosen after the results were seen.
- A headline result that does not regenerate from the pinned code and the stated data.
Scope boundary
A first read, not a compressed certification.
Preflight is not exhaustive validation. It does not include remediation, data cleaning, model changes, re-runs after code updates, or a complete issue register and evidence package.
It is not a certification, attestation, performance guarantee, or investment, legal, or tax advice.
For deeper challenge, a complete issue register, remediation planning, and a full evidence package, use Independent Model Validation.
Engagement ladder
Choose a first read or a full validation.
Preflight bounds one question quickly. Independent Model Validation builds the complete evidence record for a decision already carrying capital, governance, or oversight weight.
| Dimension | Model Preflight | Independent Model Validation |
|---|---|---|
| Question it answers | Is there a material problem worth a deeper look? | Can this model be trusted for the decision it supports? |
| Scope | One model, one pinned state, one agreed execution path; up to five principal findings. | Full review: methodology, data lineage, validation design, and complete issue register. |
| Deliverable | Findings Note (3–6 pp) with one-page summary, and debrief. | Findings memo, issue register, reproducible evidence package, and remediation path. |
| Timeline | 5–10 business days. | Typically two to six weeks. |
| Fee | $4,500 fixed · $2,250 credited toward a full validation. | From $15,000, scoped in writing. |
Engaging us on a manager you are evaluating, not your own model? See manager due diligence for allocators →
How it starts
Acceptance first. Then invoice and booking.
The commercial clock does not start while you are still discovering whether the supplied environment can be reached.
-
01
Scope by email
We confirm the model, the claimed result, the decision it supports, and whether Preflight is the right tier.
-
02
Written agreement, NDA, and input bundle
Before any non-public material, we put the engagement and NDA in writing. You then provide read-only repository access pinned to one commit, one data snapshot or documented data access, a short methodology description, data dictionary, run instructions, and a named technical contact.
-
03
Short environment smoke test
At a pre-agreed checkpoint, we confirm that access works, the pinned state resolves, the documented entry point starts, and expected outputs are writable. This is a short acceptance checkpoint — not a full run, reproduction, or substantive review. No invoice is issued until it passes.
-
04
Invoice and booking
The $4,500 fixed-fee invoice is issued after acceptance. Payment confirms your slot and starts the delivery clock.
-
05
Delivery
The Findings Note and debrief are delivered within 5–10 business days. The signed engagement governs pauses and closure if access or the technical contact becomes unavailable.
Evidence, not a verdict
If the claim does not reproduce.
Non-reproduction is documented, not hidden. But it does not by itself establish that a model is invalid: the note distinguishes an observed result divergence from insufficient evidence and an infrastructure block.
Outcome 02
Insufficient evidence
The supplied material cannot support or resolve the claimed result. The note identifies the missing evidence and keeps the conclusion explicitly unresolved.
Outcome 03
Infrastructure blocking
Access or execution fails after acceptance for reasons that prevent the agreed attempt. This is documented as a delivery constraint and handled under the signed pause and closure terms — not presented as an adverse model finding.
One-time continuation credit
$2,250 toward full validation.
A one-time credit of $2,250 applies toward a full Independent Model Validation of the same model program when the validation SOW is signed and its initial invoice is paid within 60 calendar days of the Preflight debrief. Ordinary remediation of Preflight findings does not affect the credit. A substantially new model, review objective, or data lineage is a new scope, quoted separately. Non-transferable; no cash value.
Operating terms
Confidentiality and scope stay explicit.
Commercial fit
Use Preflight when the next decision needs a record.
Good fit
- A model is about to influence an allocation, committee approval, or external claim.
- You can provide code at a pinned state, one data snapshot, and a working environment with compute, licenses, and credentials.
- You want a written paper trail with evidence, not a verbal opinion.
Not a fit
- You need a guarantee, rating, certificate, or regulatory attestation.
- The model changes daily and cannot be pinned to one reviewable version.
- You need remediation implemented inside the same fixed scope.
- You cannot share code or a working environment; an email exchange can establish what evidence is possible instead.
Professional boundary. StatGazer provides independent technical review. Model Preflight is not a financial-statement audit, not a regulatory attestation, not a certification, and not investment, legal, or tax advice.
Next step
Start with the model and the decision it supports.
Send a high-level description. Do not include non-public data, code, credentials, holdings, or client names before the written engagement and NDA are in place.