Claude's review of Codex's documents
Reviewer: Claude. Reviewed: finance/, operations/, product/, research/codex/ as on disk at 13:50 on 25 October 2026. Context: these documents were written for the version-2 plan (single-audience audits and builds at $15k and $35k). The owner has since redirected the venture at billion-dollar consumer brands with programme pricing (business plan version 3). Findings are graded Major (must change before the plan is used), Minor (should change), Note (observation, no action required). Each document gets a verdict: Pass, Pass with changes, or Superseded.
What I checked and how
- Read every document in full.
- Ran the financial model's own test suite on my machine rather than trusting the "eight tests pass" claim (result recorded below).
- Checked every internal link through the site build (zero missing targets).
- Swept all files for personal, employer, agency, panel and persona identifiers (clean).
- Cross-checked numbers quoted in one document against their source document (direct costs, holdout sizes, SOP count, MCP timing).
- Did not verify: the correctness of the model's cash-timing logic line by line; supplier price assumptions (none are quoted); the evidence register's figures against each primary paper (spot-checked three).
finance/financial-case.md and the model
Verdict: Superseded for scale; Pass as a piece of modelling for the scope it was given.
| Grade | Finding |
|---|---|
| Major | The model prices for a research manager, not an executive. Year-three mid EBITDA of $702k on $3m revenue is a consultancy, not the business the owner described. It must be rerun on the version-3 parameters (programmes at $500k–$1m ACV, 3/6/10 to 8/24/50 signed, tracker-scale fieldwork, research leads at two programmes each). |
| Major | The low case is modelled with a loss-making $25k build price. A price that loses money before overhead is a stop rule, not a scenario. Floor it or model a reduced scope instead. |
| Minor | Fieldwork allowances ($6k audit, $18k build, $1k pulse) are prudential guesses with no quotes. At programme scale fieldwork is $150–300k a year per programme and the single largest cost; it needs two written panel quotes before the next version. |
| Minor | "Expected fractional subscribers" is a sensible planning device but reads oddly in a board document; present integer customers with a probability-weighted revenue line. |
| Note | Test suite: the eight tests run and pass on my machine after installing the stated dependency. The claim is verified. |
| Note | The model is honest about what it excludes (tax, interest, financing) and says plainly that EBITDA is not cash. Keep that discipline in the rerun. |
research/codex/calibration-method.md
Verdict: Pass with changes.
| Grade | Finding |
|---|---|
| Major | Sampling is designed for one audience, one market (12 interviews, 160 development, 240 sealed). A programme runs three to six audiences in two to four markets with quarterly waves of about 1,000 completes per market. The protocol needs a wave design: how development and sealed sets rotate across quarters, how subgroup precision is bought, and what a market adds. |
| Minor | The pilot preference for "the same accessible consumer category" conflicts with the version-3 go-to-market (one anchor plus two audits in adjacent categories). Reconcile: same task family across categories, or accept three case studies with no pooled claim. |
| Minor | The 10-percentage-point planning tolerance example will be read by an executive as the product's accuracy. State the tolerance per decision in the claim card and keep the worked example out of client-facing material. |
| Note | The claim card, the three typed evidence layers (observed, derived, generated) and the leakage controls are the best part of the package. They become the backbone of the Calibration Record standard and should be published as such. |
| Note | The rule that inconclusive is a paid outcome is written correctly and consistently. |
research/codex/evidence-register.md
Verdict: Pass.
| Grade | Finding |
|---|---|
| Note | Twelve entries with primary sources, denominators and prohibited claims. Spot-checked the Columbia digital-twin figures, the Stanford 2026 v3 reliability figures and the ESOMAR status; all consistent with the sources I read. |
| Minor | Add one entry on tracker-scale human reference data quality (panel conditioning, attention checks at n=1,000 waves), since the programme now depends on it. |
research/codex/workbench-inheritance.md
Verdict: Pass.
| Grade | Finding |
|---|---|
| Note | Anonymised correctly; exact paths moved to the ignored private folder. The "what transfers" list is the right scope for a clean-room rebuild. |
| Minor | Add the explicit statement that nothing has been built and that legal review precedes the first build (present in my method document; should be here too). |
operations/operating-system.md
Verdict: Pass with changes.
| Grade | Finding |
|---|---|
| Major | The 18 SOPs cover a project business. A programme business needs at least five more: quarterly wave planning and fieldwork procurement at scale; agency onboarding and enablement; executive quarterly read; enterprise contracting (MSA, security questionnaire, data processing); and model-version change handling with recalibration (currently only in product/). |
| Major | Capacity is expressed as 80 hours per build and 6 per maintained month. Replace with programme capacity: research lead at two programmes, reviewer days per record, fieldwork lead time per wave. |
| Minor | Founder involvement at 12–16 hours a week is right for the old pilot; the anchor-partner pilot needs a named research lead from day one, which changes the staffing table. |
| Note | The role separation (research lead, external reviewer, holdout custodian, privacy owner, finance operator) is exactly right and survives the re-cut unchanged. |
operations/pilot-plan.md
Verdict: Superseded.
| Grade | Finding |
|---|---|
| Major | Three $15–35k pilots replaced by one anchor programme at $0.5m+ and two Enterprise Audits over two quarters (business plan v3, section 9). The gates and entry criteria are good and should be carried over; the schedule and sample totals should not. |
operations/record-templates.md
Verdict: Pass with changes.
| Grade | Finding |
|---|---|
| Minor | Add currency to every price field, an anonymisation status to any published record, and programme, wave and market identifiers so records roll up across quarters. |
| Note | The Calibration Record field list is complete and correctly refuses a single truth score. |
product/product-and-ux-specification.md
Verdict: Pass with changes.
| Grade | Finding |
|---|---|
| Major | Single-tenant client workspace. The programme sells access to the brand and its agencies; the spec needs agency sub-tenants, client-controlled sharing of audiences and records, and audit trails across organisations. |
| Minor | The example result wording ("too close to distinguish… test the finalists with people") is exactly the brand voice; keep it as the canonical pattern. |
| Note | The refusal to show a model-written confidence percentage is correct and should be stated in the brand platform's "we don't say" list (it is). |
product/models-and-integrations.md
Verdict: Pass with changes.
| Grade | Finding |
|---|---|
| Major | MCP is staged "later, private, read-only after repeat use". Version 3 puts an MCP server in front of the anchor partner's agency in months one to three. Bring the read-only MCP into the pilot stage with the controls already listed (tenant OAuth, audit log, revocation). |
| Minor | The AI cost illustration ($15 a study) is fine for a unit; add the programme figure ($10–20k a year) so finance and product agree. |
| Note | Model-agnostic routing and the refusal to let a model change silently alter a calibrated audience are right and are now a contractual feature of the programme. |
Summary
| Document | Verdict |
|---|---|
| finance/financial-case.md | Superseded for scale; modelling passes |
| research/codex/calibration-method.md | Pass with changes |
| research/codex/evidence-register.md | Pass |
| research/codex/workbench-inheritance.md | Pass |
| operations/operating-system.md | Pass with changes |
| operations/pilot-plan.md | Superseded |
| operations/record-templates.md | Pass with changes |
| product/product-and-ux-specification.md | Pass with changes |
| product/models-and-integrations.md | Pass with changes |
The quality of the method, evidence and operating discipline is high and should be preserved; the scale of the commercial model is the thing to change.