The AI critical review platform learns from every verdict it has ever issued, tracks whether its own predictions came true, and opens itself up as an API. Same brutal honesty. Now with a memory.
VerdictTank v2 proved the core thesis: a cross-vendor, multi-judge AI pipeline can deliver a brutally honest, majority-rules verdict on a business proposal in minutes for pennies. It works. It is live. It reviews every proposal IT Pro Partner ships internally, and it is starting to attract outside interest.
But v2 has a structural weakness: it has no memory. Every review is a clean slate. The pipeline critiques your financial model without knowing that 40 other SaaS proposals it has already reviewed made the identical mistake. It flags a churn risk without ever checking whether the last ten times it flagged that exact risk, the flaw actually materialized. Every verdict is an island.
VerdictTank v3 turns every past review into training data for the next one. Three tiers of upgrade:
Net effect: VerdictTank stops being a one-shot critique tool and becomes an institution that gets smarter and more credible with every review it runs. This is the difference between a smart intern who reviews your deck once, and a research firm that has read ten thousand decks and remembers every one.
VerdictTank's live pipeline runs five specialized review stages in three phases, ending in a majority-rules verdict from independent judges running on genuinely different underlying architectures:
Each cross-check reviewer runs on a distinct vendor architecture from the primary reviewer and from each other, so a single vendor's blind spot cannot silently control the outcome. If a judge is unavailable, the pipeline degrades gracefully to a smaller panel rather than failing the review.
~$0.38 in model inference, ~$0.47 fully loaded with infrastructure
3-9 minutes from submission to finished PDF verdict
11-section branded PDF via WeasyPrint, served at verdicttank.com/reports/
Tiered usage-based, $19-$299/mo across five tiers
No memory of past reviews. No idea whether its own predictions were correct. No vertical specialization - a fintech proposal and a consulting proposal get the same generic ten dimensions. No API. No public-facing shareable output. PDF generation is a manual per-review workflow rather than a scripted pipeline. Sanitization of internal architecture details before anything goes public is done by hand, which means it is one tired copy-paste away from a leak.
A SQLite database storing every review result: scores by dimension, flaws raised, conditions attached, and every prediction made. Searchable across the full history. This is the institutional memory the platform currently lacks.
A scheduled job runs at T+90, T+180, and T+365 days after each review, pulls real available data on the reviewed business, and checks it against what the pipeline predicted. Which flagged flaws actually materialized? Which did not?
Every reviewer earns a running accuracy score based on outcome tracking. Future reviews weight judges by their track record. A reviewer that is consistently wrong gets deprioritized or dropped, not out of loyalty to any vendor, but because the data says so.
The pipeline auto-detects proposal type - SaaS, consulting, marketplace, fintech, hardware, services - and swaps in domain-specific review dimensions and industry benchmarks instead of one generic rubric for everything.
Industry-specific attack vectors: regulatory risk for fintech, HIPAA scope creep for healthcare, churn dynamics for consumer SaaS, unauthorized practice of law exposure for legal tech. Generic critique becomes genuinely specialized critique.
Instead of a static financial table, the pipeline simulates the first 12 months: user acquisition curve, churn projection, and revenue trajectory, grounded in real market data and category benchmarks pulled during the research phase.
Sanitized reviews become shareable: an OG preview image, a score card PNG, and a tweet-ready summary. Every review a user runs becomes potential marketing material, with their consent.
The full pipeline exposed as an API endpoint. Upload a document, get a structured review back in roughly 15 minutes. This is also the integration bridge for the VentureBuilt product line.
Every new proposal gets benchmarked against similar proposals already in the corpus, in the same vertical. "Your financial model scores 2.0. The median for B2B SaaS Stage 1 proposals in our corpus is 3.2."
A pre-deploy gate script that scans every piece of public-facing output and strips model names, architecture details, and internal pipeline phase descriptions before anything reaches a public URL or the API. Manual sanitization does not scale to an API product.
The WeasyPrint report workflow gets scripted end-to-end so PDFs generate automatically from structured review data, rather than being handwritten per review. Required for the API and for volume at Enterprise-tier usage.
The v2 tiers (Free through Enterprise, $19-$299/mo, usage-metered by review count) priced the product as if every review were the same commodity unit. It no longer is. v3 tiers map to capability, not just volume: how many reviewers run, whether the corpus and prediction tracking are turned on, and whether the API is exposed.
Pay-per-review outside a subscription: $19.99, full pipeline, one-time. Annual billing on Pro and Enterprise: 20% discount. All tiers retain the degraded-mode fallback from v2 - if a judge is unavailable, the pipeline runs with fewer judges rather than failing.
The v2 panel runs four reviewers across independent vendor architectures: one primary reviewer, one sober validator, and two independent cross-checks. The moat of this product is architectural diversity, not any single model's quality. Every candidate addition below is evaluated on one question: does it bring a genuinely different training philosophy, or is it just another vendor's flavor of the same approach?
| Judge Role | Training Philosophy | What It Adds to the Panel |
|---|---|---|
| Reasoning-Verification Judge | An open-weight model trained aggressively via reinforcement learning against verifiable reasoning and math tasks rather than human preference data | Catches numerical and logical inconsistencies in financial models that preference-trained judges tend to smooth over rather than flag |
| Execution-Feasibility Judge | A model optimized for long-horizon agentic and tool-use tasks rather than conversational quality | Grades operational feasibility, not just strategy - can this actually be built and staffed on the stated timeline and budget |
| Market-Reality Judge | A model trained with a live, less-curated real-time data feed and a deliberately lower-friction personality than the standard RLHF-softened profile | Provides the contrarian, less agreeable read on market timing and demand claims that the other judges are structurally pulled away from |
| Candidate Profile | Decision | Reasoning |
|---|---|---|
| Long-context, high-throughput efficiency model | Deferred | Genuinely different cost/throughput profile, but overlaps too closely with an existing cross-check judge's strengths to justify a fifth panel slot right now. Revisit if document lengths in the corpus keep growing. |
| Sovereign, data-residency-focused efficiency model | Reserve for EU vertical | Strong fit if a European or regulated-data vertical template ships later. Not needed for the general panel today. |
| Fully open-source, community-fine-tunable model | Cost fallback only | Behind the frontier on reasoning benchmarks today. Held in reserve as a self-hostable fallback judge if vendor pricing or availability ever forces a degraded-mode default. |
| Next-generation frontier generalist | Not for the panel | Overlaps philosophically with the existing primary reviewer. Flagged instead as an alternate primary-reviewer vendor for business continuity if the current primary vendor has an outage or pricing shock. |
Per house sanitization policy, exact vendor and model names are held internally and stripped from all public-facing documentation and API responses by the automated sanitization gate (Infrastructure Fix #10). Internal architecture selection is documented separately and reviewed quarterly.
| Phase | Timeline | Scope | Exit Criteria |
|---|---|---|---|
| Phase 0 Foundation |
Weeks 1-3 | Automated sanitization gate live and blocking deploys. Automated PDF generation scripted end-to-end. Corpus database schema designed and backfilled with every historical review. | Zero manual sanitization steps remain. Every past review is queryable in the corpus. |
| Phase 1 Cumulative Intelligence |
Weeks 4-7 | Corpus search live internally. Prediction-vs-outcome cron scheduled at T+90/180/365. Reviewer accuracy scoring begins tracking (read-only, not yet weighting verdicts). | First cron cycle completes on a batch of 90-day-old reviews with real outcome data attached. |
| Phase 2 Deeper Analysis |
Weeks 8-12 | Vertical auto-classification and templates for the top 4 verticals by volume (SaaS, consulting, marketplace, fintech). Adversarial red-team per vertical. Market simulation engine replaces the static financial table. | 3 of 4 vertical templates validated against known historical outcomes with acceptable accuracy. |
| Phase 3 Product & Distribution |
Weeks 13-18 | New reviewer panel additions (Section 5) go live. Public-facing shareable review reports ship. Review-as-a-Service API opens in closed beta. Competitive corpus comparisons appear on every review. | API beta completes 100 reviews with zero sanitization gate failures. |
| Phase 4 General Availability |
Weeks 19-22 | Reviewer accuracy scoring begins weighting live verdicts. API opens to Enterprise tier generally. White-label tier launches with first 2-3 accelerator/VC pilot accounts. | First white-label pilot renews past month one; API error rate under 1%. |
| Category | Players | What They Do | VerdictTank v3 Difference |
|---|---|---|---|
| AI proposal writers | Template-based generation tools | Generate proposals from prompts | We critique them; we still don't write them. No overlap, no substitution risk. |
| Human consultants | Independent reviewers, boutique firms | Manual review, $300-800/hr, days of turnaround | Under $1 and minutes, now backed by a corpus of every proposal type they've ever seen us review |
| Single-model AI critique | Generic chat-based "roast my idea" prompts | One model, no structure, no memory, no verification | Structured multi-judge verdict with tracked prediction accuracy - the critique itself has a track record now |
| Accelerator/VC diligence tools | Internal scoring rubrics, deal-flow CRMs | Human-scored, not benchmarked against a broader corpus | White-label tier gives accelerators a benchmarked, auditable, corpus-backed second opinion on every cohort company |
| Cumulative-memory proposal review | No direct competitor exists as of August 2026. A review platform that tracks whether its own past predictions came true and adjusts its judge panel accordingly is a genuinely new category, not an incremental feature. | We are still the category, now with a moat that compounds instead of resetting every review | |
| Cost Component | Per Review |
|---|---|
| Base pipeline (research + primary + 3 cross-checks) | $0.47 |
| 3 new specialist reviewer additions (Section 5) | $0.21 |
| Market simulation engine | $0.06 |
| Vertical red-team pass | $0.05 |
| Corpus write, embedding, and similarity search | $0.02 |
| Prediction tracking cron (amortized) | $0.01 |
| Infrastructure (sanitization gate, PDF automation, storage) | $0.04 |
| Fully loaded cost per full v3 review | $0.86 |
Free-tier reviews (single reviewer, no corpus write, no simulation) stay near $0.07/review and function as a loss-leading funnel.
| Tier | Included Reviews/mo | Blended Revenue/Review | Cost/Review | Gross Margin |
|---|---|---|---|---|
| Pro ($79/mo) | 20 | $3.95 | $0.86 | 78% |
| Enterprise ($499/mo) | 100 | $4.99 | ~$1.05 (incl. corpus/API infra) | 79% |
| White-Label ($1,999+/mo) | Fair-use unlimited | Volume-dependent | ~$1.05-1.20 | 75-82% |
| Year 1 | Year 2 | Year 3 | |
|---|---|---|---|
| Paying users (Free excluded) | 160 | 620 | 2,100 |
| White-label accounts | 3 | 12 | 35 |
| MRR (year-end) | $9,400 | $41,000 | $148,000 |
| Annual revenue | $62,000 | $318,000 | $1,240,000 |
| AI + infra costs | $9,800 | $41,000 | $142,000 |
| Net | $52,200 | $277,000 | $1,098,000 |
Assumptions: AI cost deflation of 15-20% annually offsets the added inference load of the larger judge panel and market simulation engine. Churn assumed at 5-7% monthly on Pro, materially lower on Enterprise and White-Label given contractual terms. If the panel expansion in Section 5 does not clear its added cost against measurable accuracy gains within two quarters, it gets pruned back to 5 judges rather than 7 - the roadmap treats judge count as a variable to optimize, not a fixed feature.
| Risk | Likelihood | Impact | Mitigation |
|---|---|---|---|
| Corpus becomes a confidentiality liability - stored proposal content is sensitive IP | Medium | High | Encryption at rest, per-user opt-out of corpus indexing, contractual IP protections in ToS, only aggregate/statistical data (percentiles, medians) ever surfaced in comparisons - never raw proposal text |
| Reviewer accuracy scoring pushes judges toward consensus-seeking rather than independent critique | Medium | High | Preserve a diversity floor - accuracy weighting can never fully silence a dissenting judge. Cap maximum weight per reviewer. Monitor inter-judge agreement rate for unhealthy convergence, same metric already tracked for the training-data recursion risk |
| Prediction-vs-outcome attribution is genuinely hard - correlation is not causation | High | Medium | Confidence-weighted outcome labeling rather than binary right/wrong. Quarterly human spot-checks on a sample of tracked predictions before accuracy scores affect live weighting |
| Sanitization gate fails to catch a leak before public API or shareable report ships | Low | High | Automated grep-based gate blocks deploys on match, plus a manual spot check on the first 50 API responses and public reports before general availability |
| API abuse - scraping, credential sharing, cost-inflation attacks | Medium | Medium | Per-key rate limiting, usage-based billing alerts, ToS enforcement, anomaly detection on review volume spikes |
| 7-judge panel cost creep erodes margin if AI pricing does not fall as fast as projected | Medium | Medium | Live cost-per-review dashboard. Judge count is treated as a tunable parameter - underperforming or overpriced judges get dropped automatically based on the accuracy-vs-cost ratio, not manual review |
| Competitive comparison feature indirectly leaks another user's proposal specifics | Low | High | Comparisons are always statistical (this vertical's median score), never a pointer to a specific competing review. No underlying document is ever exposed across accounts |
| White-label brand risk - an accelerator's credibility gets tied to VerdictTank's judgment quality | Medium | Medium | SLA-backed accuracy audits, transparent methodology disclosure to white-label partners, opt-out clauses if quality benchmarks are missed for two consecutive quarters |
| Training-data recursion (carried forward from v2, now sharper) - the pipeline's own reviews may eventually train the models it evaluates | Medium | High | Volume cap maintained. Inter-judge agreement rate monitored for convergence. Corpus data is never knowingly sold or licensed as training data to the vendors whose models sit on the panel |
v3 is a coherent, sequenced upgrade rather than a feature dump - the infrastructure fixes (sanitization gate, automated PDF) are correctly staged before anything that touches a public API or a public report. The corpus and prediction tracking are the real long-term moat; everything else is downstream of that data existing.
This proposal will be submitted to the VerdictTank pipeline for its own review before implementation begins, consistent with house policy that every IT Pro Partner product proposal is reviewed by the tool it is proposing to improve.