Codex review of Claude's enterprise plan

25 October 2026, enterprise continuation. This review covers the nine Markdown documents currently present in the requested Claude lanes and the two embedded SVGs. strategy/claude/plan.md is the investor cut on disk; no separate investor-brief file was present. Reviewed files are identified below; source hashes are retained in the private review log. Findings describe the reviewed versions, not an allegation about any later correction.

Major: change before relying on the claim commercially or for investment. Minor: correct for consistency, precision or usability. Note: useful feature, limitation or already resolved point. Verdicts are Pass, Pass with changes, or Changes required. A research plan can be published for review with open findings; that does not make its claims verified.

What I checked

Read each listed document; compared offers, samples, fees, cadence, staffing, product states and pilot totals across the package; inspected SVG labels and their source tables; regenerated the enterprise finance model and independently recalculated workbook formulas through its eight test families. Added annual-cohort, initial-wave cash and programme-scope checks without removing the original categories of verification.

Rechecked primary material on Simile's recurring validation and Fairgen's current product; checked official research-quality guidance and the ICC/ESOMAR 2025 source. Compared prices labelled unknown in the pricing review against the map's encoded prices. Tested local Markdown destinations. I did not independently audit confidential original agency contracts, invoices, currency conversions, historical adoption logs, competitor enterprise contracts, supplier quotes, regulatory acceptance or actual product/security behaviour. Those limits matter to the claims below.

Verdicts at the reviewed snapshot

Document	Verdict
Market landscape, including competitive map	Changes required: map inference, prices and old pilot
Pricing evidence	Changes required: mixed generations of offers and cash
Construction lessons	Pass with changes: sound method skeleton; old scope superseded
Enterprise spend benchmarks	Changes required: single-client evidence overgeneralised
Business plan	Changes required: product/claim scope and pilot gates; financial reconciliation substantially improved
Investor plan	Changes required: market sizing and unsupported category claims
Brand platform	Changes required: stale commercial promises; written hierarchy remains useful
Platform, including architecture figure	Pass with changes: architecture credible; sequencing/cadence must reconcile
Everyday use	Changes required: persona validity, confidence transfer and unsupported certainty

1. Market landscape and competitive map

Grade	Finding and required correction
Major	The map assigns spend tiers to vendors whose prices the pricing review explicitly calls unknown. Its largest marker conflates expensive with undisclosed, and mixes annual spend with per-engagement fees. FishDog's stated $50–75k annual anchor is placed in the $75–250k tier. Separate verified prices, estimates and unknowns; do not encode unknown as high price.
Major	The axes combine different properties into one unsupported score. A real-human research service is not a synthetic model validated every cycle. Historical held-out ad tests do not establish the map's maximum criterion of fresh holdouts every cycle. Portability, independence and a vendor agreeing to be audited are different facts. Use transparent criteria and unknown states; avoid an apparently measured ranking.
Major	“Their records are not portable” and the empty top-right claim exceed reviewed evidence. Public pages do not disclose every enterprise contract. Trueness's own-build conflict also prevents treating its proposed record as wholly neutral by definition. Keep its intended position visibly separate from observed vendor evidence.
Minor	Fairgen's scoring rationale reverts to survey boosting, while its current site and the main table describe an integrated consumer-data/audience workspace. Update both consistently. Current primary page
Minor	SVG decimal positions differ from the integer table (for example, decimal neutrality scores versus rounded entries). If this is only display jitter, say so; otherwise publish the actual scoring data. No methodological basis for decimal precision was supplied.
Major	The GTM section still recommends two builds and one audit in 90 days and the old Audit/Audience/Maintained products. Replace with the annual anchor and two Enterprise Audits over two quarters.
Note	The main prose correctly recognises human validation already exists and labels vendor results as claims. Simile's own page describes weekly human validation; that directly supports retaining the cautious competitive conclusion. Simile

The chart can remain useful after these corrections. An unknown observation is preferable to a precise-looking score unsupported by evidence. The review does not establish that another vendor does or does not offer portable independent assessment.

2. Pricing evidence

Grade	Finding and required correction
Major	The offer table still caps eighteen possible audit cells at $150k and presents the withdrawn six-audience/four-market tier at $1.2m. Use $15k/cell, $75k floor, and the separately costed programme expansion. More task families require additional design/sample; four concepts are not four independently validated stimulus families.
Major	The supplier/pilot section still contains the former $85k contracts, $167,875 outlay and $141,450 funding, plus a link to removed `pilot-90-day-budget.csv`. Replace with the six-month pilot and current financial case. Old single-audience cost sensitivities are no longer the enterprise model's tests.
Minor	A 55–65% target is not the modelled margin. At the reference mid cost of $440k, a $600k Foundation fee yields 26.7% mature direct margin; 41.3% belongs to the $750k model ACV. Present price/scope and margin together.
Note	The distinction between published vendor prices and unknowns, and the warning about ambiguous historical currency, are sound. They should also govern the spend benchmark and chart.

3. Construction-method lessons

Grade	Finding and required correction
Minor	Steps 1/5/9 still describe a single-market pilot and the former two-build/one-audit totals. Recast the 160-development/240-sealed design as a per-cell first wave, add 80/240 refresh waves, and link programme totals to the new protocol.
Minor	Remove the default ten-point tolerance from current customer-facing methodology. The claim card now registers tolerances per decision with an explicit power/precision assessment.
Minor	Step10 suggests a new price for material model changes. Existing-scope remediation/revalidation is now included in the annual programme; only scope expansion is a new commercial order.
Note	Rights review, prospective freeze, separating generated characters from statistical samples, and treating cross-market matching as a hypothesis remain correct and useful.

4. Enterprise spend benchmarks

Grade	Finding and required correction
Major	One confidential organisation's spending does not establish typical billion-dollar-brand budgets, executive buying thresholds or willingness to purchase Trueness. Statements that the organisation would have bought a programme or that a smaller engagement is invisible need to become explicitly untested sales hypotheses.
Major	Nested agency fees, strategy workstreams, one-off refresh work and tracker/research lines can overlap. Provide a private non-overlapping bridge before summing them into an addressable replacement budget. Mark quote, actual spend and next-year proposal separately, with original currency/conversion provenance. I did not verify those originals; this review does not certify the figures as USD actuals.
Major	The internal detailed table remains in a public-file source, merely hidden by renderer comments. Its combination of sector, markets, programme structure and costs can identify a client without naming it. Move detailed source material into the ignored private archive; obtain permission/redaction review for any retained public case. The public version also should not imply every observed line is a market norm.
Major	“Most ad-hoc testing” or tracker spend being replaced is not demonstrated. Shared fieldwork must have compatible populations, instruments, exposure, permissions and sealed outcomes. A tracker already used for grounding cannot also be the independent holdout.
Minor	The fieldwork/offer section needs the per-cell sample, included-change cost and cadence comparison. $150–300k does not cover the stated larger Enterprise sample by simply reusing Foundation costs.
Note	The evidence supports testing a larger enterprise proposition and identifying a plausible budget owner. That is a useful commercial starting point without claiming an accepted purchase or verified market size.

5. Business plan

Grade	Finding and required correction
Major	“Every audience,” “every use,” “any piece of work,” and knowing “exactly” how far an audience differs promise more than the bounded prospective test. State accessible systems, specified cells/tasks, supported inputs and uncertainty. “Test as much as you like” also conflicts with twelve bounded studies and metered compute; distinguish formal studies from interactive allowances.
Major	The offer row still says roughly 1,000 completes/market and 12 interviews/audience while the reconciled protocol uses 1,200 then 960 and 12 initial interviews per audience-market cell. Explicitly specify this in scope. Four stimulus families in an Audit cannot inherit a four-concept design without a new sample/instrument decision.
Major	Section 9 says success is not a statistical threshold and calls for a renewed anchor at M6. Commercial usefulness is separate from release validity: every predictive claim still needs its registered gate. An annual service cannot demonstrate annual renewal at six months. Use paid continuation/expansion, then anniversary renewal; define a channel milestone as a qualified opportunity or executed agreement.
Major	“The partner's identity is never used ... without consent; the record is” implies automatic record publication. Record ownership does not authorise public disclosure. Require separate permission/redaction for the record as well as the name.
Minor	The final financial paragraph substantially reconciles EBITDA, per-cell pricing and larger-scope costs. Apply the latest front-loaded fieldwork cash figures from the canonical model; seven-cell direct cost is $90,250, while $94,167 is the blended seven/eight-cell average.
Minor	A $600k catalogue Foundation price and a $750k model average are not interchangeable. State the required mix/premium and its delivery scope before quoting the model's 41% mature margin. Expansion must price all incremental research, review, operation and included-change costs, not fieldwork alone.
Minor	Staffing,18SOP references and security timing need the 23SOP update. Treat SOC 2 timing as an assessed delivery target, not a universal procurement requirement or completed capability. Statistics run in versioned code, not by model inference.
Note	Accepted improvements: human calibration is not claimed as invented; model-change revalidation is included; low-case failure is disclosed; fees do not buy a favourable result; the uneconomic large fixed-price tier has been withdrawn.

The 50% audit conversion assumption cannot alone supply the programme signing totals. The finance model treats direct/channel wins as the additional source and does not add conversions again. State that acquisition burden explicitly. Do not imply that licensing fees are incremental without separate partner activity and review capacity.

6. Investor plan

Grade	Finding and required correction
Major	“Nobody knows,” “no neutral party sells validation,” and an empty competitive position are unverified absolutes. The detailed landscape already acknowledges close competing methods and incomplete contract visibility. Keep the hypothesis consistent in the short version.
Major	The 300–500 target-company count has no account list or source. The quoted $0.6–1.2m fee range yields a simple arithmetic range of $180–600m, not $300–600m. Neither range is a researched serviceable market. Supply eligible named-account methodology privately or label an illustrative scenario with corrected arithmetic.
Major	The $1.5bn category funding total, forecast standards deadlines and broad claim that incumbents' records are non-portable are not supported by cited primary evidence here. Remove or verify them. Official sources establish the ICC/ESOMAR 2025 Code; this review did not verify a January 2027 publication deadline. Future guidance is not guaranteed procurement demand.
Major	Describing the reference workflow as proof that the product “lacked only the record” skips the actual platform, rights, method and prospective evaluation work. Describe owner-reported workflow experience and explicitly separate unmeasured adoption from controlled usefulness or validity evidence.
Minor	Correct typical seven-cell cost to $90,250, and copy current funding from the canonical model after initial-wave cash front-loading. Do not present added markets priced only at fieldwork cost when staff/review also increase.
Note	The distilled six-action product and candid low-case/margin discussion make the proposition easier to assess. The financial figures are scenario outputs, not revenue forecasts or an independently assessed investment valuation.

7. Brand platform

Grade	Finding and required correction
Major	Functional benefits still promise four monthly studies and a three-week audit, while Products describes enterprise programmes and six-to-eight-week audits. The product table also retains withdrawn prices/scopes. Replace with the current offer, cadence and usage definitions rather than appending another generation.
Major	“People who behave like your audience,” any work within a minute, and quarterly records for every use imply validated behaviour and performance beyond evidence. Label synthetic characters and make speed an input-bounded target; reserve measured claims for the exact tested configuration/task.
Minor	The nine-figure buyer marketing-budget definition is inconsistent with the enterprise example used elsewhere. Buyer procurement motivations are hypotheses; “never their question” overstates the exclusion of research teams.
Note	Purpose, essence, promise, explicit proof requirements, tone and deferred identity direction follow the written template well. The no-first/no-only/no-universal-score rules should govern the investor copy and chart as firmly as this document.

8. Platform and architecture diagram

Grade	Finding and required correction
Major	The first wave is already two-market and agency access starts early, but multi-market capability appears in phase2 and sub-tenancy in phase1. Isolation and multi-market record keys must exist before the first intake; only richer interfaces can wait.
Major	“A wave is a sealed human study” conflates development and validation splits. Specify160/240 initially,80/240 later per cell, with separate custody. Monthly 150-person pulses are per market, not per audience; alternative semiannual full-wave scope changes freshness claims.
Minor	Rank correlation across four stimuli, theme overlap and dispersion do not each automatically have a defensible confidence interval. Use the protocol's estimands and appropriate uncertainty; do not claim cross-concept generalisation from four examples.
Minor	Update the old 14/12screen list and 18SOP count in prose and SVG to the six primary actions, current supporting/operator views and23SOPs. The diagram is a conceptual architecture, not evidence of controls built.
Minor	Regional hosting, SSO, SOC 2 and supplier APIs are implementation/procurement commitments to validate, not assurances already established. The most senior client still cannot see sealed outcomes, unlicensed personal data or another client's content; replace “nothing is hidden” with visibility of all released findings.
Note	The five-part architecture and proposed stack are sound pilot choices. Fixed build and per-client delivery engineering/model costs are separated in finance to avoid double counting. Model versioning, record ownership and common web/API/MCP authorisation are retained in the reconciled specification.

9. Everyday use

Grade	Finding and required correction
Major	The record tests audience-level task outputs, not each fictional person's biography, traits or individual score. Remove the implication that a character's traits were validated or that an engine can tune that character against the final holdout. A changed mapping is developed separately and needs new confirmation.
Major	The chair's “how far to trust it” cannot become measured certainty merely by attaching a past record. Historical validation error is not the confidence interval of a new creative judgement. Keep the result's fresh-human base and transfer uncertainty explicit.
Major	Challenger access must be limited to released findings, never live sealed targets used to revise the configuration. Ask the research and Panel are not required to agree; contradictions may reveal exactly the gap a user needs.
Minor	“Any stimulus,” under-a-minute completion and unrestricted Talk need supported-format, completeness, latency and fair-use qualifications. Calibration status can restrict predictive claims while leaving permitted exploration easy to use.
Minor	The detailed historical week and quoted reactions need a documented permission/anonymisation decision. Owner-reported examples can motivate the product, but unreviewed transcripts or contextual fingerprints should not become public proof.
Note	Leading with Panel, Verdict, Challenge, Talk, Ask the research and Export is the right product emphasis. The usefulness dashboard is now operationalised with event objects, deduplication, denominators, outcome attribution limits and agency-specific views in the reconciled spec.

Handoff and what remains unverified

The Codex lanes now implement the enterprise model, per-cell wave design,23SOPs, six everyday actions, sub-tenants, usefulness events and anchor-pilot MCP. The financial case, calibration protocol and product specification are their canonical details. Finance's eight checks pass; this does not validate demand, statistical performance or product controls.

Claude owns corrections to the reviewed lanes and site deployment. Re-review changed claims against this list before presenting the package as reconciled. Retain adverse findings instead of replacing this review with an unexplained pass. Exact supplier quotes, rights clearance, client commitments, empirical validity, product tests and legal/security assurance remain launch work.