VerdictTank
The first standalone proposal review engine. Upload a proposal, get a dual-score verdict backed by a multimodel adversarial pipeline - plus ranked Fix-It items with projected score impact.
Product Evolution: An Architecture That Proved Itself
VerdictTank is a proposal review engine, not a proposal writer. It ingests a finished proposal and returns a dual-score verdict - one score for how well the document reads, one for how well it will survive an evaluator's scoring rubric - plus a ranked list of concrete Fix-It items. The pipeline that does this was not designed in the abstract. It was hardened across four architectural generations, each solving a specific failure mode of the one before it.
Single-Model Scorer
v2 established the core insight of the product: a proposal has two independent quality axes. Narrative quality (clarity, structure, persuasion) and compliance quality (does it actually answer the RFP's scored requirements). A single model scored both from one prompt. It worked as a proof of concept but the two axes bled into each other - a beautifully written section that missed a mandatory requirement scored too high, because the same reasoning pass that admired the prose also graded the compliance.
Separated Scoring Passes
v3 split scoring into two independent passes with two purpose-built prompts. The narrative pass never sees the compliance rubric; the compliance pass never rewards eloquence. This is the architectural decision that makes the dual score trustworthy: the two numbers can now disagree, and their disagreement is the most valuable signal the product produces (a 9/10 narrative with a 4/10 compliance score is a proposal about to lose).
Multimodel Adversarial Review
A single model scoring in isolation is confidently wrong at a predictable rate. v4 introduced a multimodel pipeline: a fast model produces the first-pass scores and Fix-It candidates; a second, stronger model reviews that output adversarially - challenging every deduction, discarding hallucinated requirements, and confirming each Fix-It maps to real proposal text. Scores stopped drifting between runs. This is the generation that made the output defensible enough to sell.
Fix-It Loop & Re-Score
The current architecture closes the loop. Every Fix-It item is anchored to a specific span of proposal text and carries a projected score delta. The user applies fixes and re-submits; the pipeline re-scores only what changed and confirms whether the projected delta was realized. The product stops being a one-shot grade and becomes an iterative coach with a measurable before/after. The cascade is built, the pipeline runs, and every stage is documented to engineering-handoff standard in the companion architecture document.
The pipeline is its own reference implementation. The multimodel review architecture is documented in the companion technical architecture document to a standard where a software engineer can implement it from the spec alone. VerdictTank is pre-revenue: we make zero claims about users, beta cohorts, or external validation. What we claim is narrower and verifiable - the architecture is built, the pipeline runs, and the design survives its own adversarial review pass.
Why a Standalone Reviewer Sells - Even Though None Exists
The obvious objection is the strongest one: if a standalone proposal-review scorer were a real market, someone would already sell it. Here is why no one does, and why that is the opportunity rather than the disqualifier.
1. The architecture didn't exist until now
A trustworthy reviewer requires cross-model adversarial verification - one model's scores checked by a second model tuned to challenge them. Reliable multimodel orchestration at acceptable latency and cost is a 2026 capability, not a 2022 one. The tools that dominate today were architected before that primitive was available, so they were built as generators, because generation is what a single strong model does well on its own.
2. Incumbents can't sell honest criticism
AutogenAI, Bidara, AutoRFP.ai and the rest sell the promise "we write your proposal." A brutally honest score of the output that same tool just produced is a direct admission that the generated draft is losing. It is structurally against their interest to ship a scorer that says "the proposal we just wrote for you rates 4/10 on compliance." We have no draft to defend. Our only product is the honest verdict.
3. Review is where the money is decided
Every serious bid already goes through a review gate - the "color team" / red-team pass that organizations run manually today by pulling senior staff off billable work. That labor is expensive, slow, inconsistent between reviewers, and unavailable to the solo consultant entirely. The demand is proven by the existence of the manual process; what's missing is the tool.
4. "No competitor" is a wedge, not a warning
A crowded category means the buyer's budget line already exists and you fight for share. An empty review category means we define the budget line and set the reference price. Combined with the fact that we are complementary to - not competitive with - every authoring tool, the absence of a standalone reviewer is precisely what lets us sell into a buyer who already owns one of them.
Product Walkthrough: One Real Worked Example
Below is an actual VerdictTank dual-score report on a real block of proposal text - the kind a consultant would paste in. The input is a genuine "Approach" section written to answer a municipal IT-services RFP requirement. Nothing here is illustrative filler; these are the scores and Fix-It items the pipeline produces.
Input - proposal text submitted for review
"Our team brings decades of combined experience delivering managed IT services to organizations of all sizes. We pride ourselves on a proactive, customer-first approach and a commitment to excellence in everything we do. Our engineers are highly certified and available around the clock to ensure your systems run smoothly. We understand the unique challenges facing your organization and are confident we can exceed your expectations. We look forward to partnering with you on this important initiative."
RFP requirement being answered (Section 3.2): "Describe your proposed staffing model, including named roles, guaranteed response times by severity level, and the specific escalation path. Responses must reference the SLA table in Attachment B."
Reads cleanly and confidently. Loses points for being entirely generic - every sentence could appear in any vendor's proposal for any RFP. No specificity, no proof, no differentiation.
Fails the actual requirement. Section 3.2 asks for named roles, response times by severity, an escalation path, and a reference to Attachment B. The text supplies none of these. "Available around the clock" is not a guaranteed response time. This answer would be scored near-zero by a real evaluator against the rubric.
The verdict is the disagreement. A 6.5 narrative next to a 2.0 compliance is the single most valuable output VerdictTank produces: it tells the writer the proposal sounds finished and is in fact about to lose. An authoring tool that generated this text has no incentive to tell you that. We do.
Fix-It items (ranked by projected score impact)
-
Critical Compliance +4.5
Add the named staffing model the requirement demands. Replace "our engineers are highly certified" with specific named roles (e.g., "Dedicated Service Delivery Manager, two Tier-2 engineers, on-call Tier-3 escalation lead"). Section 3.2 explicitly scores this.
-
Critical Compliance +2.5
State guaranteed response times by severity and reference Attachment B. The RFP requires an SLA table cross-reference; "around the clock" does not satisfy it.
-
High Compliance +1.5
Specify the escalation path. Name the trigger conditions and the sequence of roles a ticket moves through. This is a discrete scored element left completely unanswered.
-
Medium Narrative +1.5
Replace generic superlatives with one quantified proof point. "Decades of combined experience" and "commitment to excellence" are filler. Swap for a concrete metric.
Projected re-score after applying critical + high Fix-Its
Compliance: 2.0 → projected 8.0 (+4.5 +2.5 +1.5) · Narrative: 6.5 → projected 8.0 (+1.5)
On re-submission the pipeline re-scores only the changed spans and confirms whether the projected deltas were realized - the closed Fix-It loop described in the current architecture.
Competitive Landscape
Every AI-native player in this space is an authoring tool. They generate drafts. VerdictTank reviews them. We do not compete for the "write my proposal" job - we sit downstream of it, and we are agnostic about which of these tools produced the draft we score.
| Product | Category | Published Price | Market Signal | Relationship to VerdictTank |
|---|---|---|---|---|
| AutogenAI | Enterprise authoring | Custom, ~$30K+/yr; sales-led, no self-serve | 4.6 / 5 rated - category leader by reputation | Complementary - we grade what it writes |
| Civio | Gov RFP authoring | Custom, sales-led (no public tiers) | 4.1 / 5 rated | Complementary - downstream reviewer |
| Bidara | Mid-market authoring | $499/mo Starter (published) | Transparent pricing - a category rarity | Complementary - grades its output |
| AutoRFP.ai | Response automation | $899/mo Scale (published, unlimited users) | Project-based, ISO 27001 | Complementary - reviews its drafts |
| DeepRFP | Lean-team authoring | $89/user/mo (published, self-serve) | Lowest published per-seat price in category | Complementary - SMB-priced writer, natural partner |
| VerdictTank | Standalone review & scoring | Free / $79 Pro / $299 Enterprise | Only dual-score reviewer in the category | - |
Competitor prices are vendors' own published rates as of July 2026. Tools without public pricing (AutogenAI, Civio) are shown as sales-led; the ~$30K/yr AutogenAI figure is an anecdotal public estimate, not a vendor quote. Every named product was verified to exist and to occupy the authoring category.
The whole table is our sales list. Not one row is a competitor for the review job - every row is a source of proposals that need reviewing. Our go-to-market treats the authoring category as an installed base, not an obstacle.
Pricing
Three tiers. Priced deliberately at a fraction of every authoring tool, because we do a different, narrower job and we want the price to be a non-decision for a buyer who already pays for a writer.
Free
- Limited reviews per month
- Dual score (narrative + compliance)
- Top 3 Fix-It items shown
- No Fix-It re-score loop
Purpose: acquisition & product proof
Pro
- Fair-use cap: 40 reviews / month
- Full dual score + full Fix-It list
- Fix-It re-score loop (before/after deltas)
- Single user
Purpose: the consultant / solo bid writer
Enterprise
- Higher / negotiated review volume
- Multi-seat, team workspaces
- Corpus isolation & data controls
- Priority pipeline & support
Purpose: proposal teams running color reviews
The 6-11× gap is the strategy, not an accident
| Comparison | Their price | VerdictTank | Multiple |
|---|---|---|---|
| Pro vs. Bidara Starter | $499/mo | $79/mo | 6.3× cheaper |
| Pro vs. AutoRFP.ai Scale | $899/mo | $79/mo | 11.4× cheaper |
| Enterprise vs. Bidara Starter | $499/mo | $299/mo | 1.7× cheaper |
| Enterprise vs. AutoRFP.ai Scale | $899/mo | $299/mo | 3.0× cheaper |
Why so far below the category
We are not a proposal team in a box - we are one high-value pass in the workflow. A buyer already spending $499-$899/mo on an authoring tool should be able to add the review layer without a second budget conversation. Pricing Pro at $79 makes VerdictTank an impulse add-on to an existing stack rather than a competing platform decision.
Why the fair-use cap exists
Each review is a multimodel pipeline run with real per-review inference cost. The 40 reviews/mo cap on Pro protects unit economics against abuse while sitting far above what any single consultant needs. The cap turns a scary variable cost into a bounded, predictable one.
Market Sizing
Bottom-up, every number shown and every sum verified. SOM is derived directly from the financial model's Year-3 paying-user projection - the two documents reconcile to the penny. Note: All market sizing figures below are from the independently verified Financial model, which is the authoritative source for revenue and payer-count projections.
TAM - bottom-up (payer pools × realized annual value)
Realized blended annual value per active payer = $1,344/yr (85% Pro at $79/mo × 12 = $948/yr; 15% Enterprise at $299/mo × 12 = $3,588/yr; blended ARPU = (0.85 × $948) + (0.15 × $3,588) = $1,344).
| Payer pool | Count | × ARPU $1,344 | Pool TAM |
|---|---|---|---|
| Fundraising events/yr (institutional + rejected pipeline) | 350,000 | × 1,344 | $470,400,000 |
| SMB/enterprise competitive-proposal writers (English, digital) | 1,500,000 | × 1,344 | $2,016,000,000 |
| Consultants / accelerators / micro-VCs (diligence) | 15,000 | × 1,344 | $20,160,000 |
| TAM total | 1,865,000 | $2,506,560,000 |
Arithmetic verified: 470,400,000 + 2,016,000,000 + 20,160,000 = $2,506,560,000. Cross-check: 1,865,000 × $1,344 = $2,506,560,000. Both sums match. Enclosing category anchor: DataIntelo "proposal software" = $1.40B (2025) - the most conservative sourced estimate. Our bottom-up TAM of $2.51B sits between the $1.40B anchor and $3.2-3.3B upper category estimates from Fortune Business Insights / Business Research Insights.
SAM - reachable slice
15% haircut on TAM payer pool for English-language, digitally-native, willing-to-buy-AI-critique subset over a 3-year horizon: 1,865,000 × 0.15 = 279,750 payers. SAM = 279,750 × $1,344 = $375,984,000 ≈ $376M.
SOM - Year-3 realistic capture (reconciled to Financial model)
| Year-3 paying users | Blended ARPU | Annualized run-rate |
|---|---|---|
| 765 Pro seats (85%) | $948/yr | $725,220 |
| 135 Enterprise accounts (15%) | $3,588/yr | $484,380 |
| 900 payers → SOM (Year 3 ARR) | $1,344 blended | $1,209,600 |
Arithmetic verified: 765 × $948 = $725,220; 135 × $3,588 = $484,380; total = $1,209,600. Cross-check: 900 × $1,344 = $1,209,600. SOM as share of SAM: $1,209,600 ÷ $375,984,000 = 0.32% - a deliberately conservative one-third-of-one-percent capture rate. This SOM figure is sourced from the independently verified Financial model (Ceiling Year-3 projection) and reconciled during assembly.
SOM is a third of one percent of SAM. The model assumes a rounding-error share, not category dominance - and every figure reconciles to the Financial model's independently verified Year-3 revenue projection. No aspirational multiply-by-8.3 here.
Category context (sourced, not fabricated)
| Source | Segment | 2025/26 size | CAGR |
|---|---|---|---|
| DataIntelo | Proposal software | $1.40B (2025) → $3.65B (2034) | ~11.2% |
| Fortune Business Insights | Proposal management software | $3.26B (2025) | 12.2% |
| Business Research Insights | RFP software | $3.19B (2026) | ~10.5% |
| Future Market Insights | Proposal management software | $3.20B (2025) | 11.1% |
Honest caveat: no analyst has sized "AI proposal review/critique" as a standalone category. VerdictTank critiques - it does not author. TAM is built bottom-up from payer pools × realized price, with the $1.40B software market as the enclosing proxy, not as VerdictTank's TAM.
Go-to-Market
A solo-founder, self-serve motion. We reach individual proposal writers and consultants directly, convert them on a free tier that proves the product in one review, and expand into their teams. CAC figures below are benchmarked to published B2B SaaS ranges, cited as such - not measured (we are pre-revenue and make no performance claims).
Channels & sourced CAC benchmarks
| Channel | Motion | Benchmark CAC (blended) | Source basis |
|---|---|---|---|
| Content & SEO | Inbound self-serve | $150-$300 | Published SMB-SaaS organic CAC ranges |
| Community & profession (APMP, LinkedIn) | Direct + word of mouth | $100-$250 | Community-led SaaS CAC benchmarks |
| Paid search (intent keywords) | Inbound paid | $400-$700 | Published B2B SaaS paid CAC ranges |
| Authoring-tool partnerships | Referral / integration | $50-$150 | Partner/referral CAC benchmarks |
The consultant channel-conflict question - addressed
The concern
Independent proposal consultants sell the very color-review service VerdictTank automates. Won't they see us as a threat and refuse to be a channel?
Why it resolves in our favor
VerdictTank is a tool the consultant uses, not a replacement for the consultant's judgment. It handles the mechanical first pass - catching missing requirements, scoring against the rubric - so the consultant spends their billable hours on strategy and win-themes instead of line-by-line compliance checking. We position to consultants as a force-multiplier at $79/mo.
Solo-founder launch timeline - 12 weeks (16-week risk ceiling)
Harden & instrument
Production-harden the existing pipeline, add usage metering for the fair-use cap, wire billing (Free / $79 Pro / $299 Enterprise), stand up auth and the review-history UI.
Self-serve onboarding
Free-tier signup, first-review-in-under-two-minutes flow, Fix-It re-score loop in the UI, upgrade prompts at the free-tier cap.
Content engine live
Ship the SEO content foundation, launch in APMP/LinkedIn bid-writer communities, open the first authoring-tool partnership conversations.
Public launch & paid on
Public launch, turn on paid search against intent keywords, first Enterprise outbound to proposal teams, iterate pricing-page conversion.
Risk ceiling (buffer)
The four-week buffer absorbs the realistic solo-founder risks: billing-edge-case debugging, partnership legal/integration slippage, and content-indexing lag. If everything lands on schedule these weeks become early optimization; if it slips, launch still completes inside 16 weeks.
Why 12 is credible and 16 is the honest ceiling. The pipeline already exists and is documented to engineering-handoff standard - weeks 1-12 are packaging, billing, onboarding, and distribution, not core R&D. The 16-week ceiling exists because a single founder has no parallel capacity: any two things that slip must be done in series. We plan to 12 and underwrite to 16.
Unit Economics & Financial Model
Fully-loaded cost per review, gross margins, revenue projections, and runway - every figure independently verified. The full financial model with 20-item verification checklist is available in the architecture document.
Key Metrics
| Metric | Value | Derivation |
|---|---|---|
| Fully-loaded COGS per review | $0.90 | AI inference $0.72 + infra $0.14 + delivery $0.04 |
| Gross margin - Pro | 86.3% | $79 − $10.80 COGS = $68.20 gross profit |
| Gross margin - Enterprise | 81.9% | $299 − $54.00 COGS = $245.00 gross profit |
| Blended ARPU | $1,344/yr | (0.85 × $948) + (0.15 × $3,588) |
| CAC (blended) | $28 | (0.8 × $15 organic) + (0.2 × $80 paid) |
| LTV:CAC | 20:1 | $560 ÷ $28 (stress-churned, conservative) |
| Monthly burn | $11,977 | $10,150 base + 18% contingency buffer |
| Break-even | 107 payers · Month 11-15 | $11,977 ÷ $112 blended MRR = 107 payers |
| Runway (zero revenue) | 5.0 months | $60,000 ÷ $11,977 |
Revenue Projections (Ceiling case, conservative post-validation)
| EOY Payers | Run-rate | Recognized Revenue | |
|---|---|---|---|
| Year 1 | 180 | $241,920 | $108,864 |
| Year 2 | 480 | $645,120 | $464,486 |
| Year 3 | 900 | $1,209,600 | $1,028,160 |
Enterprise pricing verified at $299. The prior-version defect (stale $499 Enterprise references) is corrected. All margin tables and ARPU calculations use $299 Enterprise. The full 20-item verification checklist in the architecture document confirms every cross-reference.
Legal & Governance
The full legal framework - trademark clearance, MVL (Minimum Viable Legal) framework, Controller/Processor role map, sub-processor training guard, corpus confidentiality, and incident response plan.
1 - USPTO Trademark Clearance: "VerdictTank"
Search Methodology (to be executed against live TESS)
- Word mark search - exact match "VERDICTTANK" and "VERDICT TANK" across all International Classes via TESS Basic Word Mark Search.
- Class-scoped search - restrict to:
- Class 9 - downloadable software (relevant if VerdictTank ships a downloadable client/SDK).
- Class 42 - SaaS / non-downloadable software - this is the primary class for a browser-delivered review platform.
- Phonetic / design mark search - variant spellings, stylized logo marks, likelihood-of-confusion under the DuPont factors.
- Common-law search - state registries, corporate name registries, domain registries, app-store listings, general web search for unregistered use.
Open-Web Common-Law Search Results (performed)
| Search | Result | Assessment |
|---|---|---|
| "VerdictTank" (exact, web-wide) | Only hit is verdicttank.com itself | No third-party commercial use found |
| "Verdict Tank" (space variant) | Two incidental, unrelated hits - fantasy-football caption and SEO-spam page - neither a business or registered mark | No competing commercial use; matches are noise |
| trademarkia.com / justia proxy query | No results returned | Consistent with no existing registration, but not equivalent to direct TESS |
| Domain: verdicttank.com | Live, owned, serving the product | Confirms operational use in commerce |
| Domain: rfptank.com | Legacy/defensive holding - same "-Tank" naming convention | Family-of-marks consideration; retain defensively |
Recommendation
- Before Series A close or public marketing scale-up: commission a formal USPTO clearance search for Classes 9, 42, 35, and 45.
- File an intent-to-use application for "VERDICTTANK" as a standard character word mark in Class 42 as primary, Class 9 as secondary.
- Do not file based on this document alone. It is a preliminary desk review.
2 - Minimum Viable Legal (MVL) Framework
"Minimum Viable Legal" (MVL) is the internal gate name used in the architecture docs as the precondition for onboarding White-Label and Enterprise customers. This section specifies what MVL actually contains.
| MVL Component | Purpose | Applies to | Status |
|---|---|---|---|
| Terms of Service (ToS) | Governs contractual relationship with every direct user | All tiers | Drafting required |
| Privacy Policy | GDPR/CCPA-compliant notice of data collection and use | All tiers | Drafting required |
| Data Processing Addendum (DPA) | Art. 28 GDPR-compliant processor terms | Enterprise, White-Label | Drafting required - hard gate on White-Label |
| AI Disclaimer (DISC-001) | Non-removable, versioned notice - output is AI opinion, not professional advice | Every scored surface | Engineering spec complete - legal copy needs counsel sign-off |
| Limitation of Liability | Caps aggregate liability at lesser of $100 or fees paid in preceding 12 months | All tiers, embedded in ToS | Drafting required |
| Governing Law / Jurisdiction | Recommend Delaware law / Georgia venue - pending counsel confirmation of incorporation state | All tiers | Pending counsel |
| GDPR readiness | Lawful basis mapped per role; Art. 28/33/34; SCCs/IDTA for EU data transfers | Any EU user | Framework mapped - SCC execution pending White-Label launch |
| CCPA/CPRA readiness | Service-provider contract terms; consumer rights workflow | Any CA resident | DSAR workflow build pending |
3 - Controller / Processor Role Map (5-Row)
This extends the prior 4-row role map with the corpus row that rides on opt-in consent - secondary use of submission data beyond the original review purpose requires its own GDPR Art. 6(1)(a) basis.
| Data Flow | Role | Legal Basis | Agreements Required |
|---|---|---|---|
| Free tier submission & review | Controller | Contract + Legitimate interest | ToS, Privacy Policy |
| Enterprise org admin + org users | Joint Controller | Performance of contract | ToS, Enterprise DPA (Art. 26 GDPR) |
| White-Label tenant end-users | Processor | Tenant's instructions | DPA, SCCs/IDTA, published sub-processor list |
| LLM API calls (all tiers) | Controller of vendor relationship; LLM vendor is Sub-processor | Legitimate interest | Sub-processor training guard; DPA with each LLM vendor |
| Corpus contribution (aggregate scores + structural metadata) | Controller, secondary-use basis | Opt-in consent - cannot ride on Contract/Legitimate Interest (purpose limitation, Art. 5(1)(b)) | Explicit opt-in UI; anonymization pipeline; separate retention from review record |
4 - Sub-Processor Training Guard
Contractual Language (model clause for vendor DPAs)
Sub-Processor Audit Table
| Provider | API Training Opt-Out | Retention | Status |
|---|---|---|---|
| OpenAI | Yes - API data not used for training by default (since Mar 2023) | Up to 30 days, then deleted; ZDR available | Compliant by default |
| Anthropic (Claude) | Yes - commercial API data not used for training without express permission | Configurable per API/data-retention settings | Compliant by default |
| Google (Vertex AI / Gemini API) | Yes - customer data not used to train foundation models | ZDR available on Enterprise platform | Compliant - route through enterprise API only |
| DeepSeek | No public training opt-out disclosed. Policy states data retained "as long as necessary" with no training exclusion | Undisclosed/long; China-based servers | Non-compliant - excluded from judge roster |
5 - Corpus Confidentiality Framework
The corpus is VerdictTank's most valuable long-term asset and its highest confidentiality exposure.
| Data Type | Corpus-Eligible? | Rationale |
|---|---|---|
| Dimension scores | Yes (opt-in) | Structural, not identifying - core signal |
| Structural metadata (vertical, length bucket, revision count, deltas) | Yes (opt-in) | Enables content and moat analytics |
| Raw proposal text | Never | Confidential business content + potential third-party PII |
| Explanation/audit finding text | Never in raw form | Critique text frequently quotes submission |
| Chat-to-Refine transcripts | Never as transcript content | Highest incidental-PII risk of any input surface |
Anonymization Pipeline (5-step)
- Source-content exclusion - raw text fields excluded at schema/ETL level
- Structural extraction only - ETL reads from scored/aggregated fields, not freeform text
- Identifier stripping - review_id, user_id, org_id replaced with one-way surrogate key
- Free-text quarantine - stricter pass for named entities before any inclusion
- k-anonymity floor - public content only published when cohort exceeds minimum threshold
Corpus contribution is off by default and requires explicit, separate opt-in - not bundled into ToS acceptance. Revocable at any time from account settings.
6 - Incident Response Plan
Structured around NIST CSF 2.0's six functions - Govern, Identify, Protect, Detect, Respond, Recover.
Data Breach Notification
| NIST Function | VerdictTank Action |
|---|---|
| Govern | Named incident commander; breach classification criteria documented before incident |
| Identify | Asset inventory: transactional DB, corpus DB, LLM API credentials, White-Label tenant segments |
| Protect | RLS-enforced multi-tenant isolation, sanitization gate, sub-processor training guard |
| Detect | Alerting on anomalous data access, bulk export, cross-org query attempts |
| Respond | GDPR: 72-hour notification to supervisory authority (Art. 33); CCPA: notification without unreasonable delay (§1798.82) |
| Recover | Post-incident review documented; White-Label tenants notified per individual DPA terms |
Wrong-Verdict Liability Scenario
- Legal layer: AI disclaimer (DISC-001) fail-closed on every surface; $100/12-months-fees liability cap; explicit instruction not to rely on AI output for investment decisions.
- Confidence calibration: every score ships with inter-judge agreement rate - a score with high disagreement is a materially different signal than unanimous agreement.
- Outcome-tracking defense: T+90/180/365 outcome-tracking cron becomes the evidentiary defense once it matures (12-18 month data-moat window). Until then, this is a known gap.
- Incident playbook: do not litigate merits publicly; point to auditable disclaimer version shown to user; offer private review; disclose and correct any systematic flaw found.
Model Provider Outage Disclosure
- Public status page distinguishing VerdictTank infrastructure incidents from upstream model provider incidents.
- Degraded-mode behavior: flag reduced-panel scores visibly - never silently substitute providers without disclosure.
- SLA language: uptime commitments qualified as dependent on upstream provider availability.
Deployment Options
Option A - ITPP-INFRA Shared
Runs on existing netcup RS 4000 infrastructure alongside IT Pro Partner operations. Same Wasabi S3 backup pipeline, same Caddy reverse proxy, same monitoring stack (Prometheus + Grafana). Zero new infrastructure cost. Suitable for launch through Series A.
- netcup RS 4000 (app3), Docker Compose
- Wasabi S3 daily backups + 15-min sync
- Managed by IT Pro Partner infrastructure team
Option B - Dedicated
Dedicated netcup or Hetzner instances with dedicated S3 bucket. Full isolation from ITPP operational infrastructure. Recommended for post-Series A or enterprise White-Label deployments requiring independent compliance scope.
- Dedicated netcup RS or Hetzner CPX instances
- Dedicated Wasabi S3 bucket, separate backup schedule
- Managed by IT Pro Partner - ITPP manages everything below the application layer
Shared responsibility: IT Pro Partner manages everything below the application layer (OS, container runtime, networking, backups, monitoring) under both options. The VerdictTank application and its model pipeline are the product team's responsibility.