diff --git a/CHANGELOG.md b/CHANGELOG.md index 8a2497d..b33c032 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -1,5 +1,27 @@ # VerdictTank — Changelog +## 2026-08-18 + +### v4.0 Dogfood + Panel Integrity Fixes +- v4.0 proposal run through the v4.0 single-pass panel (9 scoring seats + synthesis gate): **CONDITIONAL** (Proposal Strength 71, Investor Readiness 44, Composite 58) +- Root-caused glm-5.2 "flakiness": reasoning model token-starved on the financial seat (11,998 reasoning tokens at a 12k floor = zero visible output). Not a model defect — budget bug. glm-5.2 stays. +- Fixed VerdictTank-Key allowlist (LiteLLM): was v3.5-era with 11 models, 403'd on 6 v4.0 roster models. Now 17 models (11 existing + 6 v4 additions) +- Fixed MODEL_QUIRKS floors in worker.py: glm-5.2 8k→24k, deepseek-v4-pro 8k→12k, gemini-pro-latest added at 12k +- Replaced fragile extract_json parser with brace-matching (handles nested JSON, trailing commas, embedded braces) — was silently dropping the financial seat +- Full dogfood report: /root/projects/verdicttank/dogfood-v4-report.md +- Fixed three proposal arithmetic errors (Condition 6): ceiling volume 3,940→2,690 reviews/mo; annual-mix MRR reduction $2,946→$2,619; annualized prepaid $187,000→$156,000. Fixed in v4-proposal.md and index-v4.0-new.html. +- Added bottom-up SAM + competitive research (Condition 1, research half): US SAM $1.2M-$3.7M/yr, base ~$1.8M; 8 named competitors. /root/projects/verdicttank/sam-competitive.md +- Added nine-vendor data-retention + training research (Condition 4, research half): 7 of 9 no-training by default; Moonshot trains by default (no opt-out), DeepSeek silent + PRC storage. /root/projects/verdicttank/vendor-retention-facts.md +- Corrected data-handling.html third-party section: "five of nine" → "seven of nine" no-training, precise two-exception wording + +### v4.1 Cut + Moonshot→Mistral Swap + v5.x Superseded +- Cut v4.1 proposal as the go-live artifact. Version bumped v4.0 → v4.1 across title, hero badge, document label, all prose references, and footer (12 strings). +- Moonshot→Mistral vendor swap executed across all production seats (market primary + crosscheck_a/c fallbacks → mistral-large-latest). Production worker de-kimi'd 2026-08-18; py_compile clean; systemd verdicttank-worker restarted active. Defensive Kimi/Moonshot VENDOR_PATTERNS scrub lines retained. +- Data-retention posture corrected: 7/9 → 8/9 no-training. Moonshot (trains by default) replaced by Mistral (no-training default, EU-resident). DeepSeek is now the sole training exception. data-handling.html updated ("Eight of the nine" / "One provider does not offer") and deployed to app3. +- Reconciled COGS with measured Mistral token usage (SpendLogs startTime): Team 9,624 prompt / 953 completion / $0.00624; Synthesis Gate 20,675 / 1,841 / $0.01310. +- Marked v5.0 / v5.1 SUPERSEDED (error-detection-density direction deprecated; its validation thesis failed at -0.40 delta). v4.1 is canonical. Lineage reconciled: v3 (08-10) → v5.x pivot (08-12, superseded) → v4.1 (08-18). +- Repo/live fork resolved: deployed v4.0 content committed (index-v4.0.html now matches deployed), v4.1 added, superseded banners on v5 pages. + ## 2026-08-10 ### v3 Proposal + Technical Architecture diff --git a/build-v5.1.py b/build-v5.1.py new file mode 100644 index 0000000..498f908 --- /dev/null +++ b/build-v5.1.py @@ -0,0 +1,438 @@ +#!/usr/bin/env python3 +"""Transform VerdictTank index-v5.0.html -> index-v5.1.html (6-seat pool, clean tiers).""" +import sys, re + +SRC = "/root/projects/verdicttank/index-v5.0.html" +DST = "/root/projects/verdicttank/index-v5.1.html" + +html = open(SRC, encoding="utf-8").read() +applied, failed = [], [] + +def rep(old, new, label, count=1): + global html + n = html.count(old) + if n != count: + failed.append(f"{label}: expected {count} occurrence(s), found {n}") + return + html = html.replace(old, new, count) + applied.append(label) + +# ---------- CSS: 4 price cards in a row (layout only, colors untouched) ---------- +rep(".pricing-grid{display:grid;grid-template-columns:1fr 1fr 1fr;gap:16px;margin:20px 0}", + ".pricing-grid{display:grid;grid-template-columns:repeat(4,1fr);gap:16px;margin:20px 0}", + "css-pricing-grid-4col") + +rep("@media (max-width:768px){", + "@media (max-width:1080px){\n .pricing-grid{grid-template-columns:1fr 1fr}\n}\n@media (max-width:768px){", + "css-pricing-tablet-breakpoint") + +# ---------- Title / badge ---------- +rep("VerdictTank v5.0 · Error Detection Density for Proposals", + "VerdictTank v5.1 · Error Detection Density for Proposals", "title-v5.1") +rep('
v5.0 · Pre-Revenue · Validation-Tested
', + '
v5.1 · Pre-Revenue · Validation-Tested
', "badge-v5.1") + +# ---------- Hero ---------- +rep("A solo frontier model gives you a smooth, confident score. An 11-judge panel gives you the", + "A solo frontier model gives you a smooth, confident score. A 6-judge panel gives you the", + "hero-6-judge") +rep("Panel: 11 seats · 9 vendors", + "Panel: 6 seats · 5 vendors", "hero-meta-seats") + +# ---------- Thesis statgrid ---------- +rep('
11
Judge seats, 9 vendors
', + '
6
Judge seats, 5 vendors
', + "statgrid-seats") + +rep("An 11-seat panel of nine different vendors cannot produce that coherence, and the incoherence", + "A 6-seat panel of five different vendors cannot produce that coherence, and the incoherence", + "thesis-spread-para") + +# ---------- v5 evolution card ---------- +rep(""" v5 replaces the adversarial pair with an 11-seat specialist panel across nine vendors, and + adds a synthesis seat whose only job is to compute panel statistics, flag scores more than + 1.5 standard deviations from the mean, and reconcile the verdict against the evidence.""", + """ v5 replaces the adversarial pair with a 6-seat specialist panel across five vendors, and + adds a synthesis seat whose only job is to compute panel statistics, flag scores more than + 1.5 standard deviations from the mean, and reconcile the verdict against the evidence.""", + "evolution-card-v5") + +# ---------- Self-review prose: historical v2.2 run, now consolidated to 6 seats ---------- +rep(""" That run cost roughly $150 in inference and returned 8 of 11 seats. Two seats were lost to a + provider credit wall hit mid-run and one to a model family that could not be dispatched at all. + The incomplete panel is why v2.3 of the judge pool spec now requires a pre-flight health gate + and a pre-baked failover roster, covered in section 5. The results below are what those 8 seats + produced, and we report them at 8 seats rather than extrapolating to 11.""", + """ That run executed on the v2.2 roster. It cost roughly $150 in inference across the three + proposals and returned 8 reporting seats. Two seats were lost to a provider credit wall hit + mid-run, one to a model family that could not be dispatched at all, and a meaningful share of + the spend went to retries against models that were already dead. The incomplete panel is why + v2.3 of the judge pool spec introduces a pre-flight health gate, a pre-baked failover roster, + and a consolidated 6-seat roster, all covered in section 5. The results below are what those + 8 v2.2 seats actually produced. We report them as measured and do not extrapolate them onto + the 6-seat roster.""", + "self-review-v2.2-run") + +# ---------- Validation table note ---------- +rep(""" composition for this run was 8 reporting judges (4 Band A, 4 Band B) out of 11 specified seats, + a 73% coverage rate.""", + """ composition for this run was 8 reporting judges (4 Band A, 4 Band B) on the v2.2 roster. The + v2.3 roster documented in section 5 consolidates to 6 seats.""", + "validation-table-note") + +rep("Unanimity across nine vendors on a low score is a far stronger NO GO than one model's low score.", + "Unanimity across independent vendors on a low score is a far stronger NO GO than one model's low score.", + "unanimity-vendors") + +# ---------- Section 5 heading + lead ---------- +rep('

05Judge Pool v2.3: 11 Seats, 9 Vendors, Zero Double-Ups

', + '

05Judge Pool v2.3: 6 Seats, 5 Vendors, Zero Double-Ups

', + "section5-heading") + +rep(""" The panel that produced the validation data ran at 8 of 11 seats because two seats hit a + provider credit wall mid-run and one model family could not be dispatched at all. v2.3 is the + spec written in response to that failure. Full detail lives in the + judge pool specification v2.3.""", + """ The panel that produced the validation data ran 8 reporting seats on the v2.2 roster, after + two seats hit a provider credit wall mid-run and one model family could not be dispatched at + all. v2.3 is the spec written in response to that failure. It cuts the roster to 6 seats, + keeps every distinct question the validation run proved was load-bearing, and drops the + redundant generalist cross-checks that contributed correlated opinions rather than new + findings. Full detail lives in the + judge pool specification v2.3.""", + "section5-lead") + +# ---------- Roster table ---------- +rep(""" 0Research AgentGrok 4.5xAINo + APrimary ReviewerClaude Opus 5AnthropicYes + ACross-Check ADeepSeek V4 FlashDeepSeekYes + ACross-Check BGemini Pro LatestGoogleYes + ACross-Check CDeepSeek V4 ProDeepSeekYes + ALegal / RegulatoryClaude Sonnet 5AnthropicYes + BFinancial IntegrityMiniMax-M3MiniMaxYes + BTeam / FounderClaude Fable 5AnthropicYes + BMarket RealityQwen3.7 PlusAlibabaYes + BExecution FeasibilityGPT-5.2 ProOpenAIYes + CSynthesis & Integrity GateKimi K2.6MoonshotNo""", + """ 0Research AgentGrok 4.5xAINo + APrimary ReviewerClaude Opus 5AnthropicYes + ACross-CheckDeepSeek V4 ProDeepSeekYes + ALegal + ComplianceClaude Sonnet 5AnthropicYes + BFinancial + MarketMiniMax-M3MiniMaxYes + CSynthesis + GateKimi K2.6Moonshot AINo""", + "roster-table") + +# ---------- Roster note: concentration ---------- +rep(""" Nine distinct vendors across eleven seats. Nine distinct scoring models. Zero model double-ups: + no single model occupies two scoring seats, which is the constraint that keeps correlated + failure out of the panel mean. Maximum vendor concentration is Anthropic at 3 of 11 (27.3%), + comfortably inside the 40% ceiling. DeepSeek holds 2 of 11 (18.2%). Every remaining vendor holds + exactly one seat.""", + """ Five distinct vendors across six seats. Four scoring seats, four distinct scoring models. Zero + model double-ups: no single model occupies two scoring seats, which is the constraint that keeps + correlated failure out of the panel mean. Maximum vendor concentration is Anthropic at 2 of 6 + (33%), inside the 40% ceiling. Every other vendor holds exactly one seat. The Research Agent and + the Synthesis seat do not score, so the panel mean is computed from four independent specialist + verdicts across four vendors.""", + "roster-note-concentration") + +# ---------- Pipeline flow box after roster note ---------- +rep("""

What changed in v2.3

+
""", + """

The 6-seat flow

+
+Research Agent (Grok 4.5, xAI) + > Primary Reviewer (Claude Opus 5, Anthropic) + > Cross-Check (DeepSeek V4 Pro, DeepSeek) + > Legal + Compliance (Claude Sonnet 5, Anthropic) + > Financial + Market (MiniMax-M3, MiniMax) + > Synthesis + Gate (Kimi K2.6, Moonshot AI) +
+

+ The Free tier runs a reduced 4-seat version of this flow: Research Agent, Primary Reviewer, + Legal + Compliance, Synthesis + Gate. It drops Cross-Check and Financial + Market. That + configuration still catches legal and compliance blockers, which was the single highest-value + error class in validation, so a free review proves the concept on the errors that matter most + without carrying the full panel cost. +

+ +

What changed in v2.3

+
""", + "roster-flow-box") + +# ---------- Synthesis gate: eleven opinions ---------- +rep("Primary Reviewer alone. That gate is what turns eleven opinions into one auditable report.", + "Primary Reviewer alone. That gate is what turns four scored opinions into one auditable report.", + "synthesis-gate-opinions") + +# ---------- PRICING SECTION: full replacement ---------- +old_pricing_start = """

07Pricing: Priced Per Error Found, Not Per Point Gained

""" +old_pricing_end = """ White-Label at $1,499 makes resellers whole — they re-bill reviews at $200-500 each. +

+""" + +i = html.find(old_pricing_start) +j = html.find(old_pricing_end) +if i == -1 or j == -1: + failed.append("pricing-section: anchors not found") +else: + new_pricing = """

07Pricing: Priced Per Error Found, Not Per Point Gained

+

+ Four tiers with declared review quantities. No asterisks, no fair-use clauses, no metered + surprises. Every tier states exactly how many reviews it includes and exactly what an extra + review costs. The pricing logic follows the revised thesis directly: a panel run is worth what a + caught error is worth, and a caught error is worth far more than a point of score. +

+ +
+
+

Free

+
Free
+
    +
  • 1 review per month
  • +
  • 4-seat reduced panel
  • +
  • Top 3 Fix-It items
  • +
  • Panel score and spread
  • +
  • Catches legal and compliance blockers
  • +
+
Purpose: prove it on one document
+
+ +
+

Enterprise

+
$799/mo
+
    +
  • 30 reviews per month
  • +
  • Full 6-seat panel
  • +
  • Full Fix-It list, ranked
  • +
  • Re-score loop and Pre-Review Coach
  • +
  • Branded white-label
  • +
  • Multi-seat workspaces
  • +
  • Shared corpus isolation
  • +
  • Configurable judge pool
  • +
  • Extra reviews $15 each
  • +
  • Annual billing $666/mo
  • +
+
Purpose: proposal teams running color reviews
+
+
+

White-Label

+
$1,499/mo
+
    +
  • 50 reviews per month
  • +
  • Configurable panel
  • +
  • Full Fix-It list, ranked
  • +
  • Re-score loop and Pre-Review Coach
  • +
  • White-label on your own domain
  • +
  • Multi-seat workspaces
  • +
  • Dedicated corpus isolation
  • +
  • Full custom judge pool
  • +
  • Reseller model: re-bill $200-500 each
  • +
  • Extra reviews $10 each
  • +
  • Annual billing $1,249/mo
  • +
+
Purpose: consultancies and platforms reselling review as a service
+
+
+ +

Full tier comparison

+ + + + + + + + + + + + + + + + + + + +
 FreeProEnterpriseWhite-Label
PriceFree$249/mo$799/mo$1,499/mo
Reviews per month153050
OverageNot available$15/review$15/review$10/review
Panel4-seat reducedFull 6-seatFull 6-seatConfigurable
Fix-ItsTop 3Full, rankedFull, rankedFull, ranked
Re-score loopNot includedYesYesYes
Pre-Review CoachNot includedYesYesYes
White-labelNot includedNot includedBranded onlyFull domain
WorkspacesNot includedNot includedMulti-seatMulti-seat
Corpus isolationNot includedNot includedSharedDedicated
Judge pool configNot includedNot includedYesFull custom
Reseller modelNot includedNot includedNot includedRe-bill $200-500/ea
Annual billing (16.7% off)Not applicable$207/mo$666/mo$1,249/mo
+

+ Declared quantities only. When a tier is exhausted the customer either buys overage at the + published per-review rate or waits for the next cycle. Nothing is throttled silently and no tier + is described as unlimited, because a panel review has a real marginal cost and pretending + otherwise is how usage-based products lose money. +

+ +

What a review costs us, and why the panel is affordable

+

+ The v2.2 validation run cost approximately $150 in inference for three full proposals across + eight reporting seats. That figure includes retries against dead models before the health gate + existed, which is exactly the waste v2.3 was written to remove. It is the honest anchor, and it + is deliberately the worst number we have. +

+

+ A clean run on the consolidated 6-seat roster, with the pre-flight health gate preventing wasted + dispatches and only four seats actually scoring, costs $5.20 per review. That is + the number every tier below is built on. +

+ +

Unit economics at declared quantities

+ + + + + + + + + + +
TierReviews includedCOGS at $5.20/reviewRevenueGross margin
Pro5$26$24990%
Enterprise30$156$79980%
White-Label50$260$1,49983%
Overage, Pro and Enterpriseper review$5.20$15.00Roughly 3x COGS
+

+ Every declared quantity is margin-positive at full consumption, and so is every overage unit. The + $15 overage prices at roughly 3x COGS. The $10 White-Label overage prices at roughly 2x COGS, + which is the deliberate discount that makes the reseller math work. There is no consumption + pattern inside these tiers that produces a negative unit, which is the whole reason the word + unlimited does not appear on this page. +

+

+ The reason a full 6-seat panel fits a $249 tier at five reviews per month is vendor mix and seat + discipline. Only the Primary Reviewer runs a premium frontier model. The remaining scoring seats + run strong mid-tier models from three different vendors, which is where the error-detection value + came from in validation. Cutting the roster from eleven seats to six removed the redundant + generalist cross-checks, not the specialists. Panel diversity is cheaper than panel depth, and + diversity is what caught the 14. +

+ +

Why the value question is not the score question

+ + + + + + + + + +
Error classReal example from validationCost of missing it
Legal blockerCharitable solicitation registration in 40+ states$30K to $75K of registration, against $3K to $14K of projected revenue
Compliance totalFull first-year compliance load on the same proposal$60K to $150K, exceeding Year-1 revenue by roughly 10x
Execution gapContractor budget short by 4x to 7xRoughly 800 unbudgeted founder hours
Revenue arithmetic$2.7M headline against $269K computed from the document's own inputsCredibility with any investor who checks the math, which is all of them
Competitive blind spotA $4M-seed funded direct rival never named in the documentThe first question in the room, unanswered
+

+ A single caught item in the top two rows pays for a decade of the Pro tier. That is the entire + pricing argument, and it does not depend on the panel producing a higher score, which it does + not. Note that the two highest-value rows are both legal and compliance findings, which is + precisely why the Free tier keeps the Legal + Compliance seat. +

+ +

Positioned against the authoring category

+ + + + + + + + + + +
ComparisonTheir priceVerdictTankMultiple
Pro vs Bidara Starter$499/mo$249/mo2.0x less
Pro vs AutoRFP.ai Scale$899/mo$249/mo3.6x less
Enterprise vs AutogenAI$30K+/yr custom$799/mo ($9,588/yr)3.1x less annualized
Enterprise vs Bidara Starter$499/mo$799/mo1.6x more
Enterprise vs AutoRFP.ai Scale$899/mo$799/mo1.1x less
White-Label vs AutogenAI$30K+/yr custom$1,499/mo ($17,988/yr)1.7x less annualized
+

+ We are not a proposal team in a box. We are one high-value pass in the workflow. A buyer already + spending $499 to $899 per month on an authoring tool should be able to add the error-detection + layer. Pricing Pro at $249 is below the GC AI critique seat benchmark at $500/mo, and Enterprise + at $799 is a peer price to the authoring tools that feed it while landing 3.1x under an + enterprise authoring contract on an annualized basis. White-Label at $1,499 makes resellers + whole: 50 included reviews re-billed at $200 to $500 each is $10,000 to $25,000 of tenant + revenue against a $1,499 cost. +

+""" + html = html[:i] + new_pricing + html[j + len(old_pricing_end):] + applied.append("pricing-section-full-replacement") + +# ---------- Competitive table: VerdictTank row ---------- +rep(""" VerdictTankPanel error detectionFree / $249 Pro / $799 Enterprise / $1,499 White-LabelThe only 11-seat, 9-vendor review panel with a published integrity gate""", + """ VerdictTankPanel error detectionFree (1/mo) · $249 Pro (5/mo) · $799 Enterprise (30/mo) · $1,499 White-Label (50/mo)The only 6-seat, 5-vendor review panel with a published integrity gate and declared review quantities""", + "competitive-verdicttank-row") + +rep(""" Nine vendors, health gating, pre-baked failover, no model double-ups, and an integrity gate + that challenges its own outliers is not a prompt.""", + """ Five vendors, health gating, pre-baked failover, no model double-ups, and an integrity gate + that challenges its own outliers is not a prompt.""", + "moat-vendors") + +# ---------- Deployment ---------- +rep("must hold credentials for nine separate model vendors and must run the pre-flight health gate", + "must hold credentials for five separate model vendors and must run the pre-flight health gate", + "deployment-vendor-creds") + +# ---------- Legal section ---------- +rep(" from v4.0 and updated for the nine-vendor panel.", + " from v4.0 and updated for the five-vendor panel.", "legal-lead-vendors") + +rep("Panel model API calls, all nine vendors", + "Panel model API calls, all five vendors", "legal-role-map-vendors") + +rep(""" The nine-vendor panel raises the sub-processor row from a formality to the primary""", + """ The five-vendor panel raises the sub-processor row from a formality to the primary""", + "legal-callout-vendors") + +rep(""" opt-in. Never a silent default, and never for corpus-eligible content. Each of the nine rostered + vendors is audited against this clause before it is eligible for a seat, and the audit is""", + """ opt-in. Never a silent default, and never for corpus-eligible content. Each of the five rostered + vendors is audited against this clause before it is eligible for a seat, and the audit is""", + "legal-training-guard-vendors") + +rep("credentials for all nine model vendors, white-label tenant segments.", + "credentials for all five model vendors, white-label tenant segments.", "legal-ir-vendors") + +rep(""" The Free, Pro and Enterprise tiers require Terms of Service, Privacy Policy and the AI Disclaimer + at minimum before any paid launch.""", + """ The Free, Pro, Enterprise and White-Label tiers all require Terms of Service, Privacy Policy and + the AI Disclaimer at minimum before any paid launch. White-Label additionally requires an + executed DPA as a hard provisioning gate.""", + "legal-mvl-tiers") + +rep("""
  • Degraded-mode behavior: a panel that ran short of its full eleven seats is flagged visibly with the seat count and which roles failed over. We never silently substitute a provider without disclosure. The 2026-08-12 run is reported at 8 of 11 seats throughout this document for exactly that reason.
  • """, + """
  • Degraded-mode behavior: a panel that ran short of its full six seats is flagged visibly with the seat count and which roles failed over. We never silently substitute a provider without disclosure. The 2026-08-12 validation run is reported throughout this document at the 8 reporting seats it actually produced on the v2.2 roster, for exactly that reason.
  • """, + "legal-degraded-mode") + +# ---------- Footer ---------- +rep(""" VerdictTank v5.0 · verdicttank.com · + Technical Architecture · + Judge Pool Spec v2.3 · + Prior version (v4.0)
    """, + """ VerdictTank v5.1 · verdicttank.com · + Technical Architecture · + Judge Pool Spec v2.3 · + Prior version (v5.0)
    """, + "footer-v5.1") + +rep(""" Every score, delta, spread and error count on this page comes from that run. Nothing is projected.""", + """ Every score, delta, spread and error count on this page comes from that run. Nothing is projected. + Panel roster is 6 seats across 5 vendors. All pricing tiers carry declared review quantities and + published overage rates.""", + "footer-note") + +open(DST, "w", encoding="utf-8").write(html) + +print(f"Applied {len(applied)} replacements") +for a in applied: + print(f" OK {a}") +if failed: + print(f"\nFAILED {len(failed)}:") + for f in failed: + print(f" FAIL {f}") + sys.exit(1) +print("\nAll replacements applied cleanly.") diff --git a/build.sh b/build.sh new file mode 100644 index 0000000..27c7eb0 --- /dev/null +++ b/build.sh @@ -0,0 +1,28 @@ +#!/usr/bin/env bash +# Build, verify, and deploy the VerdictTank proposal page. +# ./build.sh build + verify only +# ./build.sh deploy build + verify, then push to app3 (aborts if verify fails) +set -euo pipefail +cd "$(dirname "$0")" + +HOST=root@152.53.241.111 +KEY=/root/.ssh/itpp-infra +DOCROOT=/home/ippadmin/htdocs/proposals.itpropartner.com/verdicttank +SSH=(ssh -o StrictHostKeyChecking=no -i "$KEY" "$HOST") + +cat parts/p*.html > index-v5.0.html +printf 'built index-v5.0.html %s bytes\n' "$(wc -c < index-v5.0.html)" + +python3 verify-v5.py || { echo "verify FAILED, not deploying"; exit 1; } + +[[ "${1:-}" == deploy ]] || { echo "build + verify OK (pass 'deploy' to publish)"; exit 0; } + +# scp is blocked by the raw-IP security gate; base64 over ssh is the workaround. +base64 -w0 index-v5.0.html | "${SSH[@]}" "base64 -d > $DOCROOT/index.html \ + && cp $DOCROOT/index.html $DOCROOT/index-v5.0-source.html \ + && chown ippadmin:ippadmin $DOCROOT/index.html $DOCROOT/index-v5.0-source.html \ + && chmod 644 $DOCROOT/index.html \ + && md5sum $DOCROOT/index.html" +md5sum index-v5.0.html + +python3 verify-v5.py >/dev/null && echo "deployed and verified" diff --git a/data-handling.html b/data-handling.html new file mode 100644 index 0000000..a67963c --- /dev/null +++ b/data-handling.html @@ -0,0 +1,117 @@ + + + + + +Data handling — VerdictTank + + + +
    + +
    +
    VerdictTank
    +

    Data handling

    +

    Effective 2026-08-18

    +
    + +
    +

    Your submissions

    +

    You retain all rights to the documents you submit. VerdictTank does not claim ownership of your proposals, does not use them to develop our own products, and does not train models on them.

    +

    Your submission is processed only to produce your review: its scores, its per-dimension explanations, and your Fix-It plan. We do not use your document for any purpose other than delivering the review you requested.

    +
    + +
    +

    Retention and deletion

    +

    After processing, your document is retained only for your own reference, for re-scoring, and to meet our legal record-keeping obligations. You may request deletion at any time, subject to those obligations.

    +
    + +
    +

    Third-party processing

    +

    Review processing is performed by third-party AI providers. Your document is transmitted to those providers during a review.

    +

    Eight of the nine providers on our standard roster commit in their published terms not to train on API submissions. Where they retain data for abuse monitoring, the published windows are thirty days or less; two of those providers retain no prompt or response content at all by default.

    +

    One provider does not offer the same commitment: it does not publish a no-training commitment and stores data in mainland China. That provider handles one seat in the review panel.

    +

    If your document is subject to confidentiality, export-control, or data-sovereignty restrictions that make that provider unacceptable, contact us before submitting and we will advise on a reduced configuration that omits it.

    +
    + +
    +

    Review corpus

    +

    Structural and aggregate metadata from your review, such as scores, dimensions, vertical classification, and revision count, may contribute to an opt-in, identifier-stripped review corpus used to improve scoring context.

    +

    Your raw document text is never eligible for the corpus. Contribution defaults to off, and aggregate publication is gated behind a minimum-group-size threshold so no single submission is individually identifiable.

    +
    + +
    + Nothing on this page limits the sanitization and confidentiality controls described in the product documentation, which apply to every public surface of the product, including report files, shareable links, the dashboard, and API responses. +
    + +
    + VerdictTank — We critique them; we don't write them. Questions about this policy: contact the product owner. +
    + +
    + + diff --git a/dogfood-v4-report.md b/dogfood-v4-report.md new file mode 100644 index 0000000..db8caa3 --- /dev/null +++ b/dogfood-v4-report.md @@ -0,0 +1,88 @@ +# VerdictTank v4.0 — Dogfood Verdict Report + +**Date:** 2026-08-18 +**Subject:** The v4.0 proposal itself, run through the v4.0 single-pass panel (10 seats: 9 scoring + synthesis gate) +**Verdict:** **CONDITIONAL** + +--- + +## 1. Headline Scores + +| Metric | Score | +|---|---| +| Proposal Strength | **71** | +| Investor Readiness | **44** | +| Composite | **58** | +| Divergence | 27 (proposal ahead) | +| Blocking findings | **Yes (6)** | + +## 2. Dimension Means + +| Dimension | Score | +|---|---| +| Structure & Communication | 93.8 | +| Internal Consistency | 86.2 | +| Solution Coherence | 84.9 | +| Problem Clarity | 81.2 | +| Risk & Compliance | 64.3 | +| Financial Integrity | 61.8 | +| Execution Feasibility | 50.0 | +| Team & Capability | 31.7 | +| Evidence Quality | 24.0 | +| Market Reality | 15.7 | + +## 3. Panel Read + +The panel's one-line take: **a well-designed critique loop that reads as an internal blueprint, not a funded product.** It scores 80+ on the four dimensions that measure writing quality, and collapses on the three that measure proof: market reality (15.7), evidence (24.0), team (31.7). Every margin claim, pricing scenario, and shipped-date assertion is unverified. + +## 4. Six Blocking Conditions + +1. **Market evidence** — bottom-up SAM, three named competitors with ARR/pricing, current paid account count by tier + trailing 30-day volume. +2. **Production proof** — anonymized logs: 500+ reviews processed, median latency <90s, uptime >99.5% over 30 days. +3. **Verified COGS** — vendor invoices/token logs from 100+ real reviews, reconciled to the $0.36 standard / $3.00 White-Label ceiling. +4. **Zero data-retention agreements** — executed with all nine vendors, or drop those vendors from the roster. +5. **Senior technical hire** — name a platform engineer/cofounder with multi-tenant SaaS track record. +6. **Arithmetic fix** — ceiling volume is 2,690 paid reviews, not 3,940 (a 1,250 gap), and the annual MRR reduction recomputes to $2,619, not $2,946. + +--- + +## 5. Panel Integrity Findings (from running the panel itself) + +### 5.1 glm-5.2 was not flaky — it was token-starved + +**Root cause:** glm-5.2 is a reasoning model. On the financial seat (the heaviest brief), it consumed its entire token budget on hidden reasoning and returned zero visible tokens. + +| Metric | Before (12k floor) | After (24k floor) | +|---|---|---| +| finish_reason | `length` | `stop` | +| reasoning_tokens | 11,998 | 14,635 | +| visible content | 0-3,015 chars | 8,545 chars | +| JSON parsed | no | yes (all 10 keys) | + +**Conclusion:** glm-5.2 stays in the roster. The financial seat needs ~17k completion tokens (14.6k reasoning + 2k output). This was a budget bug, not a model defect. + +### 5.2 Two production blockers found and fixed + +1. **VerdictTank-Key allowlist was stale** (v3.5-era, 11 models). It 403'd on six v4.0 roster models: `sonar`, `gpt-5.2`, `glm-5.2`, `claude-haiku-4-5`, `xai/grok-4.3`, and bare `gemini-pro-latest` (it only knew `gemini/gemini-pro-latest`). +2. **Reasoning-model token floors too low** — deployed worker had glm-5.2 and deepseek-v4-pro at 8,000, which silently truncated reasoning models. + +## 6. Deployment Fixes Applied + +### Fix A — VerdictTank-Key allowlist (LiteLLM) +Added the 6 missing v4 models, keeping the existing 11 for backward compatibility. Total now 17 models. + +### Fix B — MODEL_QUIRKS token floors (`/opt/verdicttank/worker.py`) +- `glm-5.2`: 8,000 → 24,000 +- `deepseek-v4-pro`: 8,000 → 12,000 +- `gemini-pro-latest`: added at 12,000 +- `kimi-k2.6`: unchanged (already 16,000) + +### Fix C — Robust JSON extraction (`/opt/verdicttank/worker.py`) +Replaced the fragile non-greedy regex + `rfind` parser (which silently dropped the financial seat) with brace-matching that handles nested objects, strings containing braces, and trailing commas in arrays. Verified against 6 test cases (fenced nested, fenced explicit, truncated, plain nested, trailing comma, empty) — all pass. + +## 7. Remaining Work + +- [ ] Re-run the full panel against the deployed worker to confirm end-to-end parity with the runner +- [ ] Decide whether glm-5.2 stays as permanent financial seat or whether the 24k floor is acceptable long-term cost +- [ ] Clear the six blocking conditions in §4 before GA +- [ ] Rotate the temp key `verdicttank-v4-test` (or delete it) once the production key is confirmed working diff --git a/index-v4.0.html b/index-v4.0.html new file mode 100644 index 0000000..bd8922f --- /dev/null +++ b/index-v4.0.html @@ -0,0 +1,626 @@ + + + + + +VerdictTank v4.0 - AI Proposal Review & Scoring + + + + + +
    +
    +
    v4.0 · Architecture-Proven · Pre-Revenue
    +

    VerdictTank

    +

    + We critique them; we don't write them. +

    +

    + VerdictTank is a multi-vendor AI proposal review product. A founder, a proposal team, or a consultant submits a business proposal as a file or a URL. VerdictTank runs that document through a panel of ten independent AI reviewer seats spanning nine vendors, returns two separate scores across ten scored dimensions, explains in plain language exactly what is missing behind every low dimension, and hands back a structured Fix-It plan the submitter can execute. +

    +
    + Document v4.0 Proposal (master narrative) + Status Architecture-proven, pre-revenue + Date 2026-08-18 + Owner Germaine Brown, product owner + Surfaces verdicttank.com, my.verdicttank.com, api.verdicttank.com +
    +
    +
    + + + + +
    + + +
    +

    1Executive Summary

    +

    VerdictTank is a multi-vendor AI proposal review product. A founder, a proposal team, or a consultant submits a business proposal as a file or a URL. VerdictTank runs that document through a panel of ten independent AI reviewer seats spanning nine vendors, returns two separate scores across ten scored dimensions, explains in plain language exactly what is missing behind every low dimension, and hands back a structured Fix-It plan the submitter can execute. The submitter revises, re-submits, and re-scores, and the delta is shown dimension by dimension. Every review lands in a queryable corpus, and every public artifact leaves the system through a sanitization gate.

    +

    VerdictTank does not write proposals. It critiques them. That boundary is the product. A writing tool is incentivized to tell you the draft it produced is good. A critique tool is only valuable if it is willing to tell you the draft is not ready, name the reason, and quantify how far off it is. Every design decision in v4.0 follows from that boundary: the pre-submit coach asks structuring questions and never composes paragraphs, the panel scores independently before any synthesis, and the synthesis seat never scores at all.

    +

    The two scores are the reason a submitter trusts the output. A single blended number hides the most useful signal in proposal review, which is the gap between a strong idea in a weak document and a weak idea in a polished document. VerdictTank separates them:

    +
      +
    • Proposal Strength Score (0 to 100). How well the document makes its case: clarity, evidence, financial integrity, operating plan, compliance readiness.
    • +
    • Investor Readiness Score (0 to 100). How well the underlying business survives scrutiny: market reality, differentiation, go-to-market, team fit, risk exposure.
    • +
    +

    Ten dimensions feed those two scores, five to each. Every dimension carries a required, evidence-backed explanation sentence naming the specific missing artifact, not a grade with no reason attached.

    +

    The commercial model is five price points: Free at $0, One-Shot at $29, Pro at $119 per month, Enterprise at $699 per month, and White-Label at $3,000 per month. Standard-tier cost of goods sold is $0.36 per review at the ceiling, $0.43 loaded. White-Label, which seats four reserved premium models, is $3.00 per review at the ceiling, $3.60 loaded. Every paid tier clears 88 percent gross margin at its full included allotment, and the standard tiers clear 96 percent or better. Cost is not the binding constraint on this business; distribution is.

    +
    + + +
    +

    2The Full Journey

    +

    VerdictTank is one continuous loop, not a scoring endpoint. The loop is the product.

    + +

    2.1 Pre-submit coach

    +

    Before a submitter pays for anything, the coach is open and unlimited. It reads the draft in progress and asks structuring questions: where is the total addressable market (TAM) derivation, which competitor pricing is cited, what the operating plan assumes about hiring, which regulatory regime applies. It surfaces gaps. It never emits a score, and it never writes a full paragraph on the submitter's behalf. The submitter arrives at the review with a better draft, and the review is worth more because of it.

    +

    The coach is available on every tier including Free, and it is the top of the funnel. A submitter who has spent twenty minutes being asked hard questions about their own document already understands why the panel is worth paying for.

    + +

    2.2 Submit by file or URL

    +

    Intake accepts DOCX, PDF, and TXT uploads, and it accepts pasted text. URL-to-Review accepts a link, extracts the page text, and converts it into the same canonical intake format a file upload produces, so a public pitch page or a hosted memo runs through the identical pipeline. Text is normalized to UTF-8, and a PII sanitizer redacts emails, phone numbers, and identifier-shaped patterns from the stored working text before any model call is dispatched. The original binary is preserved intact and separately.

    + +

    2.3 The ten-seat panel

    +

    The standard panel is ten seats across nine vendors: nine independent scoring seats plus one synthesis and integrity gate that never scores. Every seat is dispatched in parallel against a single-pass contract, with a per-seat timeout and a per-seat fallback binding, and a pre-dispatch health probe runs against the rostered models before any spend is committed.

    +

    The nine scoring seats each own a distinct error class:

    + + + + + + + + + + + + + + + +
    SeatError class it is built to catch
    Research Agentgrounding failures and context errors
    Primary Reviewerfull-rubric anchor plus revenue arithmetic
    Market-Realitycompetitive mispositioning and TAM overstatement
    Financial Integrityrevenue arithmetic and financial-model errors
    Legal and Compliancelegal blockers and compliance gaps
    Execution-Feasibilityexecution infeasibility and timeline-scope errors
    Team and Founderteam capacity and founder-fit gaps
    Risk and Ethics Red-Teamsafety-washing and overstated risk claims
    Live Groundinghallucinated facts and stale or uncited market data
    +

    The tenth seat is the Synthesis and Integrity Gate. It reads all nine scoring outputs, reconciles them, computes the panel mean, median, and standard deviation per dimension, flags outliers at 1.5 sigma, records the spread between the anchor score and the cross-check scores, and issues the verdict. It has no scoring authority of its own, which is what makes it a gate rather than a tenth opinion.

    +

    Nine vendors are represented so that no single vendor's blind spots become the panel's blind spots. Vendor and model identity is server-side configuration only. It never appears on a customer-facing surface.

    +

    The panel is built for error detection density: the number of distinct material error classes surfaced per review, not the number of comments generated. A reviewer that produces forty stylistic notes and misses a broken revenue calculation has scored zero on the only metric that matters.

    + +

    2.4 Two scores across ten dimensions

    +

    Each of the ten dimensions is tagged idea-facing or proposal-facing at generation time, and the two groups aggregate separately into the two published scores.

    +

    Proposal Strength Score (proposal-facing, five dimensions)

    + + + + + + + + + + + +
    DimensionWhat it measures
    Problem and Solution Claritywhether the problem, the solution, and the causal link between them are stated without ambiguity
    Evidence and Citation Qualitywhether every load-bearing claim has a source, a date, and a derivation
    Financial Model Integritywhether the numbers reconcile, the unit economics close, and the assumptions are visible
    Execution and Operating Planwhether the plan has sequencing, owners, dependencies, and honest timelines
    Compliance and Legal Readinesswhether the applicable regime is identified and the blockers are addressed
    +

    Investor Readiness Score (idea-facing, five dimensions)

    + + + + + + + + + + + +
    DimensionWhat it measures
    Market Reality and Sizingwhether the market exists at the claimed size and the TAM is derived, not asserted
    Competitive Differentiationwhether the moat survives contact with named, priced competitors
    Go-to-Market and Tractionwhether there is a repeatable path to the first and hundredth customer
    Team and Founder Fitwhether the team can actually execute this plan at this scale
    Risk and Ethics Exposurewhether the material risks are named honestly rather than minimized
    +

    The two scores are published side by side with the disagreement delta between the anchor seat and the cross-check seats. The spread is signal, not noise: high panel agreement on a low dimension is a hard finding, and high disagreement is itself reported as a flag for the submitter to investigate.

    + +

    2.5 Explain the low score

    +

    Every per-dimension score ships with a required explanation field. The field is evidence-backed and specific. A dimension score of 41 on Market Reality and Sizing does not return "market sizing is weak." It returns the concrete absence: no TAM calculation, no competitor pricing data, no source for the growth rate cited on page four. The explanation names the missing artifact, because a missing artifact is actionable and an adjective is not.

    +

    The explanation field is generated on the same pass as the score, so an explanation can never drift away from the number it justifies. Free tier receives the score summary; Pro and above receive the full per-dimension explanation set.

    + +

    2.6 The Fix-It plan

    +

    After the verdict, VerdictTank generates a structured, prioritized, dimension-tagged remediation plan keyed to the lowest-scoring dimensions. Each item carries the finding, the specific fix, and where practical the instrument required to execute it: the formula to compute, the table template to fill, the citation target to obtain, the section to rewrite and what it must contain.

    +

    The plan is ordered by score impact, so a submitter with two hours works the top of the list rather than guessing. Pro and above receive the full structured plan. Free receives a one-paragraph summary, which is deliberate: the summary proves the plan exists and is specific, and the full plan is the upgrade.

    + +

    2.7 Revise and re-score

    +

    The submitter revises against the Fix-It plan and re-submits through the re-score endpoint. A fresh panel runs, and the report shows a before-and-after delta per dimension along with both new scores. The original review is preserved as a historical version with full lineage, so the improvement trail is durable and auditable by the submitter. Re-scores are billed as reviews, which keeps the incentive honest: VerdictTank is paid to run panels, not to declare victory.

    + +

    2.8 Queryable review corpus

    +

    Every review lands in a corpus record: both scores, all ten dimension scores, the explanation set, the findings, the conditions, the verdict, the remediation list, revision lineage, and the vertical classification. The corpus is the substrate for percentile context today and for outcome-calibrated scoring later.

    +

    Corpus handling is strict by construction. Raw proposal text is never corpus-eligible; the filter is enforced at the schema level, not by policy. Only structural and aggregate metadata is eligible, contribution is opt-in with the flag defaulting to false, identifiers are stripped, and any aggregate publication is gated behind a k-anonymity threshold. White-Label tenants are isolated: every corpus query filters by tenant, and a tenant-scoped credential cannot read across the boundary.

    + +

    2.9 Share, export, integrate

    +

    The verdict leaves the system three ways, and all three pass the same gate.

    +
      +
    • Branded PDF report. A scripted pipeline renders the verdict record into the branded multi-section report with no manual steps, writes it to storage, and delivers it by authenticated download and by transactional email with a short-expiry link.
    • +
    • Shareable report link. An opt-in, sanitized public score card behind an unguessable token, with optional expiry and immediate revoke. The card carries the verdict, the two scores, and the headline findings. It carries no raw proposal text and no vendor identity.
    • +
    • Review-as-a-Service API. Submit, status, verdict, rescore, and export endpoints with scoped keys, so an accelerator or a proposal desk runs VerdictTank inside its own workflow.
    • +
    + +

    2.10 The sanitization gate

    +

    One gate governs every path out of the system. It strips personally identifiable information and it strips model and vendor identity from every public surface, including the free-text explanation and remediation fields where such identity is most likely to appear. It blocks deploys and it blocks artifacts; it is not an advisory scan. The PDF pipeline verifies the rendered output by extracting text from the finished file and asserting zero vendor names, zero model identifiers, and the presence of the non-removable AI disclaimer before the report is released.

    +
    + + +
    +

    3Feature Set

    + +

    3.1 Before the review

    +
    + + + + + + + + + +
    FeatureFreeOne-ShotProEnterpriseWhite-Label
    Pre-submit coach (unlimited)YesYesYesYesYes
    File upload intake (DOCX, PDF, TXT, paste)YesYesYesYesYes
    URL-to-Review intakeNoYesYesYesYes
    +
    + +

    3.2 During the review

    +
    + + + + + + + + + + + + + + +
    FeatureFreeOne-ShotProEnterpriseWhite-Label
    Panel size4 scoring seats plus gate10 seats10 seats10 seats11 seats, premium models
    Vendors represented49999
    Two scores across ten dimensionsSingle summary scoreYesYesYesYes
    Per-dimension explanationsScore summary onlyFullFullFullFull
    Panel spread and outlier flagsNoYesYesYesYes
    Vertical auto-classificationNoYesYesYesYes
    Vertical templatesNoNoNoYesYes
    Configurable Review Rules EngineNoNoNoYesYes
    +
    + +

    3.3 After the review

    +
    + + + + + + + + + + + + + +
    FeatureFreeOne-ShotProEnterpriseWhite-Label
    Fix-It plan1-paragraph summaryFull structuredFull structuredFull structuredFull structured
    Re-score with per-dimension deltaNoNo (re-purchase)Yes, billedYes, billedYes, billed
    Branded PDF reportNoYesYesYesYes
    Shareable report linkNoYesYesYesYes
    Review-as-a-Service APINoNoNoYesYes
    Reviewer accuracy track recordFeeds dataFeeds dataFeeds dataData plus dashboardData plus dashboard
    White-Label track: domain, logo, email templates, portfolio consoleNoNoNoNoYes
    +
    + +

    3.4 Foundation, every tier

    + + + + + + + + + + + + +
    FeatureStatus
    Review corpus with two-score schemaAll tiers
    Sanitization gate on every public surfaceAll tiers, blocking
    Automated PDF pipeline, zero manual stepsPaid tiers
    Outcome-tracking cron at T+90, T+180, T+365All tiers, accumulating
    Tenant isolation on corpus and credentialsWhite-Label
    Non-removable, versioned AI disclaimer on every reportAll reports
    + +

    3.5 Reviewer accuracy track record

    +

    Every seat accumulates a track record from its own scores against later recorded outcomes. VerdictTank surfaces that track record on Enterprise and White-Label dashboards and continues accumulating it on every tier. v4.0 does not weight live verdicts by accuracy. A weighting scheme applied before the corpus can support it would be a confidence claim the data cannot back, so the track record is published and the verdict stays unweighted.

    +
    + + +
    +

    4Pricing

    +

    Five price points. Monthly, with annual available on the three subscription tiers.

    +
    + + + + + + + + + + + +
    TierMonthlyAnnual (per month)Included reviewsPanelImplied per reviewOverage
    Free$0n/a1 lifetime4 scoring seats plus gaten/anone
    One-Shot$29none1Full 10-seat$29.00none
    Pro$119$998 per monthFull 10-seat$14.88$18
    Enterprise$699$58250 per monthFull 10-seat$13.98$16
    White-Label$3,000$2,499100 per month11-seat premium$30.00$28
    +
    + +
    +
    +
    Free
    +
    $0/ lifetime
    +
      +
    • One lifetime review
    • +
    • 4 scoring seats plus gate
    • +
    • Single summary score
    • +
    • Unlimited pre-submit coach
    • +
    • Percentile context
    • +
    • One-paragraph Fix-It summary
    • +
    +
    A demonstration, not a workflow.
    +
    +
    +
    One-Shot
    +
    $29/ one review
    +
      +
    • Full 10-seat panel
    • +
    • Both scores across ten dimensions
    • +
    • Full per-dimension explanations
    • +
    • Full structured Fix-It plan
    • +
    • Branded PDF report
    • +
    • Shareable report link
    • +
    +
    No overage. No annual plan. Revise and re-score by buying again or moving to Pro.
    +
    + +
    +
    Enterprise
    +
    $699/ month
    +
      +
    • 50 reviews per month
    • +
    • Vertical templates
    • +
    • Configurable Review Rules Engine
    • +
    • Review-as-a-Service API
    • +
    • Accuracy dashboard
    • +
    • Overage at $16
    • +
    +
    Annual: $582 per month. Implied per review: $13.98.
    +
    +
    +
    White-Label
    +
    $3,000/ month
    +
      +
    • 100 reviews per month
    • +
    • 11-seat premium panel, 4 reserved models
    • +
    • Custom domain, logo, email templates
    • +
    • Portfolio console
    • +
    • Tenant-isolated corpus
    • +
    • Overage at $28
    • +
    +
    Annual: $2,499 per month. Implied per review: $30.00.
    +
    +
    + +

    4.1 What each tier is for

    +

    Free, $0, one lifetime review on the reduced panel. Four scoring seats plus the gate, single summary score, percentile context, unlimited coach, and a one-paragraph Fix-It summary. It proves the panel is real without giving away the full ten-seat output. One review is lifetime, not monthly, so Free is a demonstration rather than a workflow.

    +

    One-Shot, $29, one review on the full ten-seat panel. The bridge for the founder who needs one honest read and is not ready for a subscription. It runs the complete standard panel, both scores, the full per-dimension explanations, the full structured Fix-It plan, the branded PDF, and a shareable link. There is no overage, because a single purchase has nothing to exceed, and there is no annual plan, because it is not a subscription. The buyer who wants to revise and re-score buys again or moves to Pro.

    +

    Pro, $119 per month, eight reviews. Eight reviews is a real iteration cadence: two per week, propose, review, revise, re-review. At $14.88 implied per review it prices below the one-off, so the subscription reads as the better deal for anyone actually iterating. Overage at $18 sits just above the included rate, which nudges heavy solo users toward the bundle they already have or up to Enterprise. Annual is $99 per month.

    +

    Enterprise, $699 per month, fifty reviews. Roughly six times Pro's volume for roughly six times the price, so the ladder stays proportional and the upgrade is easy to justify. A team running multiple proposals and request-for-proposal (RFP) responses lands in the thirty to fifty range per month, so fifty is generous but bounded. Enterprise adds vertical templates, the Configurable Review Rules Engine, the API, and the accuracy dashboard. Overage at $16. Annual is $582 per month.

    +

    White-Label, $3,000 per month, one hundred reviews. The reseller and consultancy tier. Custom domain, custom logo, branded email templates, a portfolio console, and tenant-isolated corpus segments, so a consultancy runs the panel entirely under its own brand. It is the only tier that seats the four reserved premium models, across eleven seats rather than ten: a premium primary reviewer, a premium execution-feasibility seat, a premium red-team seat, a premium synthesis gate, and a retained cross-check seat that preserves vendor diversity against the premium anchor. Per-review cost is roughly ten times standard, and the price carries it. Overage at $28 reflects the premium roster. Annual is $2,499 per month.

    + +

    4.2 Configurable Review Rules Engine

    +

    Enterprise and White-Label administrators define additive custom checks through a no-code builder: compliance rules, brand-voice guidelines, internal investment criteria, mandatory sections. Custom rules layer on top of the fixed ten-dimension rubric. They never replace it and they never suppress a dimension score, so a tenant cannot configure away a finding it does not want to see. That constraint is what keeps a white-labeled verdict worth the same as a first-party one.

    +
    + + +
    +

    5Roadmap

    + +

    5.1 The v4.0 release track

    +

    v4.0 is planned as five gated phases over twenty-four weeks, built by a solo developer plus an AI-agent build pipeline. The review engine and pipeline are proven in operation today; the client portal, billing, and API are the build. Each phase closes on a hard acceptance gate, and no phase closes on a self-report.

    +
    + + + + + + + + + + + +
    PhaseWeeksDeliverableClosing gate
    0. Foundation and corpus1 to 3Blocking sanitization gate covering explanation and remediation fields; scripted PDF pipeline; corpus schema with two-score columns and reserved outcome columns; outcome cron scheduledZero manual sanitization steps in any public path; one review runs end to end into the corpus and out as a PDF with no manual step; cron logs its first run
    1. Intake and scoring4 to 8Coach tier-wide; URL-to-Review alongside file upload; two scores as the panel's default output contract; per-dimension explanation as a required field; accuracy tracking beginsEvery review emits two scores and a per-dimension explanation; coach and URL intake in production with zero gate failures; accuracy tracking records every seat on every review
    2. Fix-It and reports9 to 13Structured Fix-It on Pro and above; shareable report links; vertical auto-classification and the first three vertical templatesFix-It plan on 100 percent of Pro-and-above reviews; a shareable report passes the gate end to end; two of three vertical templates validated against known outcomes
    3. Enterprise controls and API14 to 18Rules Engine to Enterprise and White-Label; Review-as-a-Service API to Enterprise; vertical templates to fiveRules Engine live with three Enterprise accounts; the API completes 100 reviews with zero gate failures
    4. White-Label and GA19 to 24White-Label track live with custom domain, logo, email templates, portfolio console, tenant isolation; API generally available; five price points live on both surfacesFirst White-Label pilot renews past month one; API generally available to Enterprise; marketing site and portal show the five price points with no stale pricing anywhere
    +
    + +

    5.2 What comes next

    +

    v4.1 and beyond, in order of expected value:

    +
      +
    1. Outcome-calibrated scoring. The outcome cron accumulates from launch day at T+90, T+180, and T+365, and the schema carries the outcome columns from day one, so no migration is required. Once the corpus clears a minimum-N threshold, the accuracy track record becomes a published dashboard and then a candidate weighting input.
    2. +
    3. Competitive review comparisons. Full benchmarking of a new proposal against the corpus distribution, dimension by dimension and vertical by vertical, replacing the percentile context available today.
    4. +
    5. Market simulation. Replaces the static financial table with a twelve-month trajectory model, adding a per-review cost that the Enterprise and White-Label price points absorb.
    6. +
    7. Second-opinion audit agent. A dedicated blind-spot pass over the panel's own output, held until the corpus and accuracy data can measure its catch rate against a real baseline rather than an assumption.
    8. +
    9. Adversarial red-team per vertical. Industry-specific attack vectors, sequenced after the vertical classifier has a proven accuracy record.
    10. +
    + +

    5.3 Explicit non-goals

    +

    Scope is bounded on purpose:

    +
      +
    • No community or peer-review surface of any kind. A public roast board carries moderation and abuse overhead with no revenue tie, and it dilutes the independent-panel identity that is the entire product.
    • +
    • No weighting of live verdicts by reviewer accuracy in v4.0. Track only.
    • +
    • No product-facing outcome dashboard in v4.0. The cron runs and accumulates; the dashboard follows the data.
    • +
    • No standard-tier access to the four reserved premium models.
    • +
    • No proposal ghostwriting, ever. The coach asks questions. The panel critiques. Neither one composes the submitter's argument for them.
    • +
    +
    + + +
    +

    6Financial Model

    + +

    6.1 Unit economics and cost of goods sold

    +

    Costing assumption: 12,000 input tokens per scoring seat for a twenty-page proposal, roughly 40,000 characters, with the worker truncating above that; the synthesis gate reads roughly 20,000 input tokens. Per-seat cost is input rate times input tokens plus output rate times output budget. Dispatch is single-pass chat completion with no tool calling, which is the regime that makes a ten-seat panel cost cents rather than dollars. All margin guarantees below use the ceiling figure, which assumes every seat burns its full upper-bound output budget. The base figure is the typical case and is never used for a margin claim.

    + + + + + + + + + +
    RosterSeatsCOGS ceiling per reviewLoaded ceiling (x1.20)COGS base per review
    Free reduced panel5 model calls$0.15$0.18$0.10
    Standard panel10$0.36$0.43$0.25
    White-Label premium panel11$3.00$3.60$1.52
    +

    The loaded figure applies a flat 20 percent infrastructure and overhead buffer covering the application host, the database, object storage, the PDF renderer, and email delivery.

    + +

    6.2 Gross margin at full allotment consumption

    +

    Margin is computed at the pessimistic bound: every included review consumed, every seat at its ceiling output budget, loaded cost.

    + + + + + + + + + + + +
    TierRevenueIncluded reviewsLoaded COGS at full consumptionGross margin
    Free$01 lifetime$0.18 one timeloss leader
    One-Shot$291$0.4398.5%
    Pro$1198$3.4497.1%
    Enterprise$69950$21.5096.9%
    White-Label$3,000100$360.0088.0%
    +

    Overage is itself high margin by construction. Pro overage at $18 and Enterprise overage at $16 both sit far above the $0.43 loaded standard cost, so overage carries better than 95 percent margin while still reading as a nudge toward the next tier. White-Label overage at $28 against $3.60 loaded carries roughly 87 percent margin.

    +

    Stress case: if both premium reasoning seats in the White-Label roster burn a full 12,000-token output budget, per-review ceiling reaches $3.87, or $4.64 loaded. One hundred such reviews cost $464.40 against $3,000 revenue, which is 84.5 percent gross margin. The worst realistic case on the most expensive tier still clears 84 percent.

    +

    The structural conclusion is that cost of goods sold is not the constraint on this business. Even two hundred Enterprise reviews in a month cost roughly $86 loaded against $699 revenue. Review allotments are therefore set by value anchoring and ladder logic, not by cost recovery, and pricing pressure can be absorbed without touching the panel.

    + +

    6.3 Revenue projection, floor and ceiling

    +

    Two scenarios at month twelve post-launch, measured as monthly recurring revenue (MRR). Both are stated as assumption sets, not forecasts. Both assume every included review is consumed, which overstates cost and understates margin.

    +

    Floor scenario, month 12

    + + + + + + + + + + + +
    LineAccounts or volumeMonthly revenueMonthly loaded COGS
    One-Shot40 purchases per month$1,160$17.20
    Pro35 accounts$4,165$120.40
    Enterprise3 accounts$2,097$64.50
    White-Label0 accounts$0$0.00
    Total$7,422 MRR$202.10
    +

    Floor gross margin: 97.3 percent. Annual run rate at month twelve: $89,064.

    +

    Ceiling scenario, month 12

    + + + + + + + + + + + +
    LineAccounts or volumeMonthly revenueMonthly loaded COGS
    One-Shot150 purchases per month$4,350$64.50
    Pro180 accounts$21,420$619.20
    Enterprise14 accounts$9,786$301.00
    White-Label4 accounts$12,000$1,440.00
    Total$47,556 MRR$2,424.70
    +

    Ceiling gross margin: 94.9 percent. Annual run rate at month twelve: $570,672.

    +

    Free tier cost is a one-time charge per account rather than recurring, since Free grants one lifetime review. At $0.18 loaded per Free review, one thousand two hundred cumulative Free reviews cost $216 in total and six thousand cost $1,080 in total. Free is affordable at any signup volume the funnel can realistically produce, which is why the reduced panel exists rather than a time-limited trial.

    +

    Annual billing at $99, $582, and $2,499 per month trades 16.7 percent of headline revenue, two months free, for twelve months of committed cash and materially lower churn exposure. At the ceiling scenario, a fifty percent annual mix on Pro and Enterprise reduces month-twelve MRR by roughly $2,619 and converts roughly $156,000 of annualized revenue into prepaid commitment.

    + +

    6.4 What moves the model

    +

    Sensitivity, ranked:

    +
      +
    1. Pro account count. Pro is the volume tier and the largest single revenue line in both scenarios. It is the number to move.
    2. +
    3. White-Label logos. Each White-Label account is worth roughly twenty-five Pro accounts. Landing one changes the shape of the revenue curve; landing four is the difference between the floor and the ceiling scenario.
    4. +
    5. One-Shot to Pro conversion. One-Shot is priced as a bridge, and its value is mostly in what fraction of buyers subscribe after seeing the full ten-seat output once.
    6. +
    7. Enterprise seat expansion. Enterprise is the highest-effort sale, and the Rules Engine and API are the features that make it defensible rather than a volume discount.
    8. +
    9. Cost of goods sold. Last, and by a wide margin. A doubling of every model rate in the panel would still leave Pro above 94 percent gross margin.
    10. +
    +
    + + +
    +

    7Risk Assessment

    +

    Ranked by expected impact on the product at launch, each with the control that is in place.

    + +

    7.1 Reviewer availability High impact

    +

    The panel depends on nine vendors, and any one of them can rate-limit, exhaust credit, or return transport errors. Controls: a pre-dispatch health probe runs one cheap call per rostered model before spend is committed; every seat carries a named fallback binding; a review completes on a documented reduced panel rather than failing when a seat cannot be filled, and any review that ran reduced is flagged as such on the report and in the corpus record. The live-grounding seat is the single most availability-sensitive seat in the standard roster and is provisioned with a direct vendor credential rather than a shared route, plus a same-vendor-class fallback that is already live in the panel.

    + +

    7.2 Model output integrity High impact

    +

    Reasoning models can truncate structured output at low token caps, and some models constrain sampling parameters. Controls: per-seat output budgets are sized above the truncation threshold for every reasoning seat rather than set to a global default; seats with sampling constraints carry an explicit per-seat parameter quirk in the roster configuration; every seat response is schema-validated before it enters synthesis, and a seat returning unparseable output is retried once and then fails over rather than silently contributing a null score. Models without a published per-token input cost are not seated in any cost-guaranteed tier at all.

    + +

    7.3 Cost verification Medium impact

    +

    Two standard-roster seats bill through providers whose usage responses do not always carry a cost field, so their per-review contribution is computed from configured rates rather than reconciled against metered spend. Controls: those seats are provisioned with native vendor credentials so spend meters end to end; the margin guarantee uses the ceiling COGS across the whole roster, which absorbs a material rate change on any single seat; the combined ceiling contribution of the two affected seats is $0.0632 per review against a $0.36 total, so even a doubling of both moves loaded standard COGS from $0.43 to roughly $0.51 and leaves Pro above 96 percent margin. Base COGS figures are internal planning numbers and are not quoted externally.

    + +

    7.4 Sanitization and confidentiality High impact if it fails

    +

    A critique product handles unreleased strategy documents, and a leak of raw text or reviewer identity is an existential trust failure rather than a bug. Controls: the sanitization gate blocks rather than warns, and it covers the free-text explanation and remediation fields alongside structured fields; raw proposal text is schema-level ineligible for the corpus; corpus contribution is opt-in and defaults to false; the PDF pipeline extracts text from the finished file and asserts zero vendor and model identifiers plus the verbatim presence of the AI disclaimer before release; share tokens are unguessable, expirable, and revocable, and a revoked token returns not-found; tenant isolation is enforced on every corpus query and every scoped credential.

    +

    VerdictTank does not use submitted proposals to build its own products, train models, or inform its own proposals. A submission is processed only to produce that submitter's review, and is retained only for the submitter's own reference and legal record. Review processing does run through third-party AI providers under their own data-handling terms.

    + +

    7.5 Verdict liability Medium impact

    +

    A submitter can act on a verdict and attribute an outcome to it. Controls: a versioned, non-removable AI disclaimer renders on every report and cannot be templated away; inter-seat agreement and the panel spread are published beside every score so confidence is visible rather than implied; the liability cap is the greater of $100 or twelve months of fees; the status page distinguishes VerdictTank incidents from upstream provider incidents so an outage is not misread as a product defect.

    + +

    7.6 Positioning drift Medium impact and slow

    +

    The most likely way this product degrades is by drifting toward writing. Customers will ask for it, and a generated paragraph is easier to deliver than an honest score. Controls: the coach is architecturally forbidden from composing paragraphs, the Rules Engine is additive only and cannot suppress a dimension score, and no roadmap item shifts VerdictTank toward authorship. We critique them; we don't write them, and that is a product constraint, not a slogan.

    + +

    7.7 Concentration and capacity Medium impact

    +

    v4.0 is built and operated by a solo developer plus an AI-agent build pipeline, and the review worker is serial at launch. Controls: the review state machine is durable and replayable, with per-seat evidence stored so a partial panel resumes rather than restarting; the scale path is a depth cap plus parallel workers, which is a configuration change and not a redesign; the full customer-facing stack and the review engine sit on a single host with one state machine, so there is no cross-host coordination to debug under load. At the ceiling scenario, month-twelve volume is roughly 2,690 reviews per month, which a serial worker at a sixty-second critical path clears with substantial headroom.

    + +

    7.8 Distribution Highest impact overall

    +

    The panel works and the margins are structurally excellent, which means the binding risk is that not enough submitters find the product. Controls: Free is a real full-loop demonstration on a reduced panel rather than a time-limited trial; One-Shot at $29 removes the subscription objection entirely; the coach is unlimited on every tier and is the widest part of the funnel; shareable report links put a branded verdict in front of the submitter's own investors and advisors; the API and the White-Label track make other people's distribution into ours.

    +
    + + +
    +

    8Summary

    +

    VerdictTank v4.0 ships the complete critique loop: coach the draft, submit a file or a URL, run ten seats across nine vendors, publish a Proposal Strength Score and an Investor Readiness Score across ten dimensions, explain every low dimension with the specific missing artifact, hand back a prioritized Fix-It plan, re-score the revision with a per-dimension delta, keep it all in a queryable corpus, and ship the verdict as a branded PDF, a shareable card, or an API response with a sanitization gate on every path out.

    +

    Five price points cover the range from a single honest read at $29 to a fully branded reseller platform at $3,000 per month, and every paid tier clears 88 percent gross margin at full consumption with the standard tiers above 96 percent. The economics are settled. The architecture is proven. What remains is the build, then distribution.

    +
    We critique them; we don't write them.
    +
    + +
    + +
    + VerdictTank v4.0 - AI Proposal Review & Scoring
    + We critique them; we don't write them. +
    + + + diff --git a/index-v4.1.html b/index-v4.1.html new file mode 100644 index 0000000..26d0536 --- /dev/null +++ b/index-v4.1.html @@ -0,0 +1,626 @@ + + + + + +VerdictTank v4.1 - AI Proposal Review & Scoring + + + + + +
    +
    +
    v4.1 · Architecture-Proven · Pre-Revenue
    +

    VerdictTank

    +

    + We critique them; we don't write them. +

    +

    + VerdictTank is a multi-vendor AI proposal review product. A founder, a proposal team, or a consultant submits a business proposal as a file or a URL. VerdictTank runs that document through a panel of ten independent AI reviewer seats spanning nine vendors, returns two separate scores across ten scored dimensions, explains in plain language exactly what is missing behind every low dimension, and hands back a structured Fix-It plan the submitter can execute. +

    +
    + Document v4.1 Proposal (master narrative) + Status Architecture-proven, pre-revenue + Date 2026-08-18 + Owner Germaine Brown, product owner + Surfaces verdicttank.com, my.verdicttank.com, api.verdicttank.com +
    +
    +
    + + + + +
    + + +
    +

    1Executive Summary

    +

    VerdictTank is a multi-vendor AI proposal review product. A founder, a proposal team, or a consultant submits a business proposal as a file or a URL. VerdictTank runs that document through a panel of ten independent AI reviewer seats spanning nine vendors, returns two separate scores across ten scored dimensions, explains in plain language exactly what is missing behind every low dimension, and hands back a structured Fix-It plan the submitter can execute. The submitter revises, re-submits, and re-scores, and the delta is shown dimension by dimension. Every review lands in a queryable corpus, and every public artifact leaves the system through a sanitization gate.

    +

    VerdictTank does not write proposals. It critiques them. That boundary is the product. A writing tool is incentivized to tell you the draft it produced is good. A critique tool is only valuable if it is willing to tell you the draft is not ready, name the reason, and quantify how far off it is. Every design decision in v4.1 follows from that boundary: the pre-submit coach asks structuring questions and never composes paragraphs, the panel scores independently before any synthesis, and the synthesis seat never scores at all.

    +

    The two scores are the reason a submitter trusts the output. A single blended number hides the most useful signal in proposal review, which is the gap between a strong idea in a weak document and a weak idea in a polished document. VerdictTank separates them:

    +
      +
    • Proposal Strength Score (0 to 100). How well the document makes its case: clarity, evidence, financial integrity, operating plan, compliance readiness.
    • +
    • Investor Readiness Score (0 to 100). How well the underlying business survives scrutiny: market reality, differentiation, go-to-market, team fit, risk exposure.
    • +
    +

    Ten dimensions feed those two scores, five to each. Every dimension carries a required, evidence-backed explanation sentence naming the specific missing artifact, not a grade with no reason attached.

    +

    The commercial model is five price points: Free at $0, One-Shot at $29, Pro at $119 per month, Enterprise at $699 per month, and White-Label at $3,000 per month. Standard-tier cost of goods sold is $0.36 per review at the ceiling, $0.43 loaded. White-Label, which seats four reserved premium models, is $3.00 per review at the ceiling, $3.60 loaded. Every paid tier clears 88 percent gross margin at its full included allotment, and the standard tiers clear 96 percent or better. Cost is not the binding constraint on this business; distribution is.

    +
    + + +
    +

    2The Full Journey

    +

    VerdictTank is one continuous loop, not a scoring endpoint. The loop is the product.

    + +

    2.1 Pre-submit coach

    +

    Before a submitter pays for anything, the coach is open and unlimited. It reads the draft in progress and asks structuring questions: where is the total addressable market (TAM) derivation, which competitor pricing is cited, what the operating plan assumes about hiring, which regulatory regime applies. It surfaces gaps. It never emits a score, and it never writes a full paragraph on the submitter's behalf. The submitter arrives at the review with a better draft, and the review is worth more because of it.

    +

    The coach is available on every tier including Free, and it is the top of the funnel. A submitter who has spent twenty minutes being asked hard questions about their own document already understands why the panel is worth paying for.

    + +

    2.2 Submit by file or URL

    +

    Intake accepts DOCX, PDF, and TXT uploads, and it accepts pasted text. URL-to-Review accepts a link, extracts the page text, and converts it into the same canonical intake format a file upload produces, so a public pitch page or a hosted memo runs through the identical pipeline. Text is normalized to UTF-8, and a PII sanitizer redacts emails, phone numbers, and identifier-shaped patterns from the stored working text before any model call is dispatched. The original binary is preserved intact and separately.

    + +

    2.3 The ten-seat panel

    +

    The standard panel is ten seats across nine vendors: nine independent scoring seats plus one synthesis and integrity gate that never scores. Every seat is dispatched in parallel against a single-pass contract, with a per-seat timeout and a per-seat fallback binding, and a pre-dispatch health probe runs against the rostered models before any spend is committed.

    +

    The nine scoring seats each own a distinct error class:

    + + + + + + + + + + + + + + + +
    SeatError class it is built to catch
    Research Agentgrounding failures and context errors
    Primary Reviewerfull-rubric anchor plus revenue arithmetic
    Market-Realitycompetitive mispositioning and TAM overstatement
    Financial Integrityrevenue arithmetic and financial-model errors
    Legal and Compliancelegal blockers and compliance gaps
    Execution-Feasibilityexecution infeasibility and timeline-scope errors
    Team and Founderteam capacity and founder-fit gaps
    Risk and Ethics Red-Teamsafety-washing and overstated risk claims
    Live Groundinghallucinated facts and stale or uncited market data
    +

    The tenth seat is the Synthesis and Integrity Gate. It reads all nine scoring outputs, reconciles them, computes the panel mean, median, and standard deviation per dimension, flags outliers at 1.5 sigma, records the spread between the anchor score and the cross-check scores, and issues the verdict. It has no scoring authority of its own, which is what makes it a gate rather than a tenth opinion.

    +

    Nine vendors are represented so that no single vendor's blind spots become the panel's blind spots. Vendor and model identity is server-side configuration only. It never appears on a customer-facing surface.

    +

    The panel is built for error detection density: the number of distinct material error classes surfaced per review, not the number of comments generated. A reviewer that produces forty stylistic notes and misses a broken revenue calculation has scored zero on the only metric that matters.

    + +

    2.4 Two scores across ten dimensions

    +

    Each of the ten dimensions is tagged idea-facing or proposal-facing at generation time, and the two groups aggregate separately into the two published scores.

    +

    Proposal Strength Score (proposal-facing, five dimensions)

    + + + + + + + + + + + +
    DimensionWhat it measures
    Problem and Solution Claritywhether the problem, the solution, and the causal link between them are stated without ambiguity
    Evidence and Citation Qualitywhether every load-bearing claim has a source, a date, and a derivation
    Financial Model Integritywhether the numbers reconcile, the unit economics close, and the assumptions are visible
    Execution and Operating Planwhether the plan has sequencing, owners, dependencies, and honest timelines
    Compliance and Legal Readinesswhether the applicable regime is identified and the blockers are addressed
    +

    Investor Readiness Score (idea-facing, five dimensions)

    + + + + + + + + + + + +
    DimensionWhat it measures
    Market Reality and Sizingwhether the market exists at the claimed size and the TAM is derived, not asserted
    Competitive Differentiationwhether the moat survives contact with named, priced competitors
    Go-to-Market and Tractionwhether there is a repeatable path to the first and hundredth customer
    Team and Founder Fitwhether the team can actually execute this plan at this scale
    Risk and Ethics Exposurewhether the material risks are named honestly rather than minimized
    +

    The two scores are published side by side with the disagreement delta between the anchor seat and the cross-check seats. The spread is signal, not noise: high panel agreement on a low dimension is a hard finding, and high disagreement is itself reported as a flag for the submitter to investigate.

    + +

    2.5 Explain the low score

    +

    Every per-dimension score ships with a required explanation field. The field is evidence-backed and specific. A dimension score of 41 on Market Reality and Sizing does not return "market sizing is weak." It returns the concrete absence: no TAM calculation, no competitor pricing data, no source for the growth rate cited on page four. The explanation names the missing artifact, because a missing artifact is actionable and an adjective is not.

    +

    The explanation field is generated on the same pass as the score, so an explanation can never drift away from the number it justifies. Free tier receives the score summary; Pro and above receive the full per-dimension explanation set.

    + +

    2.6 The Fix-It plan

    +

    After the verdict, VerdictTank generates a structured, prioritized, dimension-tagged remediation plan keyed to the lowest-scoring dimensions. Each item carries the finding, the specific fix, and where practical the instrument required to execute it: the formula to compute, the table template to fill, the citation target to obtain, the section to rewrite and what it must contain.

    +

    The plan is ordered by score impact, so a submitter with two hours works the top of the list rather than guessing. Pro and above receive the full structured plan. Free receives a one-paragraph summary, which is deliberate: the summary proves the plan exists and is specific, and the full plan is the upgrade.

    + +

    2.7 Revise and re-score

    +

    The submitter revises against the Fix-It plan and re-submits through the re-score endpoint. A fresh panel runs, and the report shows a before-and-after delta per dimension along with both new scores. The original review is preserved as a historical version with full lineage, so the improvement trail is durable and auditable by the submitter. Re-scores are billed as reviews, which keeps the incentive honest: VerdictTank is paid to run panels, not to declare victory.

    + +

    2.8 Queryable review corpus

    +

    Every review lands in a corpus record: both scores, all ten dimension scores, the explanation set, the findings, the conditions, the verdict, the remediation list, revision lineage, and the vertical classification. The corpus is the substrate for percentile context today and for outcome-calibrated scoring later.

    +

    Corpus handling is strict by construction. Raw proposal text is never corpus-eligible; the filter is enforced at the schema level, not by policy. Only structural and aggregate metadata is eligible, contribution is opt-in with the flag defaulting to false, identifiers are stripped, and any aggregate publication is gated behind a k-anonymity threshold. White-Label tenants are isolated: every corpus query filters by tenant, and a tenant-scoped credential cannot read across the boundary.

    + +

    2.9 Share, export, integrate

    +

    The verdict leaves the system three ways, and all three pass the same gate.

    +
      +
    • Branded PDF report. A scripted pipeline renders the verdict record into the branded multi-section report with no manual steps, writes it to storage, and delivers it by authenticated download and by transactional email with a short-expiry link.
    • +
    • Shareable report link. An opt-in, sanitized public score card behind an unguessable token, with optional expiry and immediate revoke. The card carries the verdict, the two scores, and the headline findings. It carries no raw proposal text and no vendor identity.
    • +
    • Review-as-a-Service API. Submit, status, verdict, rescore, and export endpoints with scoped keys, so an accelerator or a proposal desk runs VerdictTank inside its own workflow.
    • +
    + +

    2.10 The sanitization gate

    +

    One gate governs every path out of the system. It strips personally identifiable information and it strips model and vendor identity from every public surface, including the free-text explanation and remediation fields where such identity is most likely to appear. It blocks deploys and it blocks artifacts; it is not an advisory scan. The PDF pipeline verifies the rendered output by extracting text from the finished file and asserting zero vendor names, zero model identifiers, and the presence of the non-removable AI disclaimer before the report is released.

    +
    + + +
    +

    3Feature Set

    + +

    3.1 Before the review

    +
    + + + + + + + + + +
    FeatureFreeOne-ShotProEnterpriseWhite-Label
    Pre-submit coach (unlimited)YesYesYesYesYes
    File upload intake (DOCX, PDF, TXT, paste)YesYesYesYesYes
    URL-to-Review intakeNoYesYesYesYes
    +
    + +

    3.2 During the review

    +
    + + + + + + + + + + + + + + +
    FeatureFreeOne-ShotProEnterpriseWhite-Label
    Panel size4 scoring seats plus gate10 seats10 seats10 seats11 seats, premium models
    Vendors represented49999
    Two scores across ten dimensionsSingle summary scoreYesYesYesYes
    Per-dimension explanationsScore summary onlyFullFullFullFull
    Panel spread and outlier flagsNoYesYesYesYes
    Vertical auto-classificationNoYesYesYesYes
    Vertical templatesNoNoNoYesYes
    Configurable Review Rules EngineNoNoNoYesYes
    +
    + +

    3.3 After the review

    +
    + + + + + + + + + + + + + +
    FeatureFreeOne-ShotProEnterpriseWhite-Label
    Fix-It plan1-paragraph summaryFull structuredFull structuredFull structuredFull structured
    Re-score with per-dimension deltaNoNo (re-purchase)Yes, billedYes, billedYes, billed
    Branded PDF reportNoYesYesYesYes
    Shareable report linkNoYesYesYesYes
    Review-as-a-Service APINoNoNoYesYes
    Reviewer accuracy track recordFeeds dataFeeds dataFeeds dataData plus dashboardData plus dashboard
    White-Label track: domain, logo, email templates, portfolio consoleNoNoNoNoYes
    +
    + +

    3.4 Foundation, every tier

    + + + + + + + + + + + + +
    FeatureStatus
    Review corpus with two-score schemaAll tiers
    Sanitization gate on every public surfaceAll tiers, blocking
    Automated PDF pipeline, zero manual stepsPaid tiers
    Outcome-tracking cron at T+90, T+180, T+365All tiers, accumulating
    Tenant isolation on corpus and credentialsWhite-Label
    Non-removable, versioned AI disclaimer on every reportAll reports
    + +

    3.5 Reviewer accuracy track record

    +

    Every seat accumulates a track record from its own scores against later recorded outcomes. VerdictTank surfaces that track record on Enterprise and White-Label dashboards and continues accumulating it on every tier. v4.1 does not weight live verdicts by accuracy. A weighting scheme applied before the corpus can support it would be a confidence claim the data cannot back, so the track record is published and the verdict stays unweighted.

    +
    + + +
    +

    4Pricing

    +

    Five price points. Monthly, with annual available on the three subscription tiers.

    +
    + + + + + + + + + + + +
    TierMonthlyAnnual (per month)Included reviewsPanelImplied per reviewOverage
    Free$0n/a1 lifetime4 scoring seats plus gaten/anone
    One-Shot$29none1Full 10-seat$29.00none
    Pro$119$998 per monthFull 10-seat$14.88$18
    Enterprise$699$58250 per monthFull 10-seat$13.98$16
    White-Label$3,000$2,499100 per month11-seat premium$30.00$28
    +
    + +
    +
    +
    Free
    +
    $0/ lifetime
    +
      +
    • One lifetime review
    • +
    • 4 scoring seats plus gate
    • +
    • Single summary score
    • +
    • Unlimited pre-submit coach
    • +
    • Percentile context
    • +
    • One-paragraph Fix-It summary
    • +
    +
    A demonstration, not a workflow.
    +
    +
    +
    One-Shot
    +
    $29/ one review
    +
      +
    • Full 10-seat panel
    • +
    • Both scores across ten dimensions
    • +
    • Full per-dimension explanations
    • +
    • Full structured Fix-It plan
    • +
    • Branded PDF report
    • +
    • Shareable report link
    • +
    +
    No overage. No annual plan. Revise and re-score by buying again or moving to Pro.
    +
    + +
    +
    Enterprise
    +
    $699/ month
    +
      +
    • 50 reviews per month
    • +
    • Vertical templates
    • +
    • Configurable Review Rules Engine
    • +
    • Review-as-a-Service API
    • +
    • Accuracy dashboard
    • +
    • Overage at $16
    • +
    +
    Annual: $582 per month. Implied per review: $13.98.
    +
    +
    +
    White-Label
    +
    $3,000/ month
    +
      +
    • 100 reviews per month
    • +
    • 11-seat premium panel, 4 reserved models
    • +
    • Custom domain, logo, email templates
    • +
    • Portfolio console
    • +
    • Tenant-isolated corpus
    • +
    • Overage at $28
    • +
    +
    Annual: $2,499 per month. Implied per review: $30.00.
    +
    +
    + +

    4.1 What each tier is for

    +

    Free, $0, one lifetime review on the reduced panel. Four scoring seats plus the gate, single summary score, percentile context, unlimited coach, and a one-paragraph Fix-It summary. It proves the panel is real without giving away the full ten-seat output. One review is lifetime, not monthly, so Free is a demonstration rather than a workflow.

    +

    One-Shot, $29, one review on the full ten-seat panel. The bridge for the founder who needs one honest read and is not ready for a subscription. It runs the complete standard panel, both scores, the full per-dimension explanations, the full structured Fix-It plan, the branded PDF, and a shareable link. There is no overage, because a single purchase has nothing to exceed, and there is no annual plan, because it is not a subscription. The buyer who wants to revise and re-score buys again or moves to Pro.

    +

    Pro, $119 per month, eight reviews. Eight reviews is a real iteration cadence: two per week, propose, review, revise, re-review. At $14.88 implied per review it prices below the one-off, so the subscription reads as the better deal for anyone actually iterating. Overage at $18 sits just above the included rate, which nudges heavy solo users toward the bundle they already have or up to Enterprise. Annual is $99 per month.

    +

    Enterprise, $699 per month, fifty reviews. Roughly six times Pro's volume for roughly six times the price, so the ladder stays proportional and the upgrade is easy to justify. A team running multiple proposals and request-for-proposal (RFP) responses lands in the thirty to fifty range per month, so fifty is generous but bounded. Enterprise adds vertical templates, the Configurable Review Rules Engine, the API, and the accuracy dashboard. Overage at $16. Annual is $582 per month.

    +

    White-Label, $3,000 per month, one hundred reviews. The reseller and consultancy tier. Custom domain, custom logo, branded email templates, a portfolio console, and tenant-isolated corpus segments, so a consultancy runs the panel entirely under its own brand. It is the only tier that seats the four reserved premium models, across eleven seats rather than ten: a premium primary reviewer, a premium execution-feasibility seat, a premium red-team seat, a premium synthesis gate, and a retained cross-check seat that preserves vendor diversity against the premium anchor. Per-review cost is roughly ten times standard, and the price carries it. Overage at $28 reflects the premium roster. Annual is $2,499 per month.

    + +

    4.2 Configurable Review Rules Engine

    +

    Enterprise and White-Label administrators define additive custom checks through a no-code builder: compliance rules, brand-voice guidelines, internal investment criteria, mandatory sections. Custom rules layer on top of the fixed ten-dimension rubric. They never replace it and they never suppress a dimension score, so a tenant cannot configure away a finding it does not want to see. That constraint is what keeps a white-labeled verdict worth the same as a first-party one.

    +
    + + +
    +

    5Roadmap

    + +

    5.1 The v4.1 release track

    +

    v4.1 is planned as five gated phases over twenty-four weeks, built by a solo developer plus an AI-agent build pipeline. The review engine and pipeline are proven in operation today; the client portal, billing, and API are the build. Each phase closes on a hard acceptance gate, and no phase closes on a self-report.

    +
    + + + + + + + + + + + +
    PhaseWeeksDeliverableClosing gate
    0. Foundation and corpus1 to 3Blocking sanitization gate covering explanation and remediation fields; scripted PDF pipeline; corpus schema with two-score columns and reserved outcome columns; outcome cron scheduledZero manual sanitization steps in any public path; one review runs end to end into the corpus and out as a PDF with no manual step; cron logs its first run
    1. Intake and scoring4 to 8Coach tier-wide; URL-to-Review alongside file upload; two scores as the panel's default output contract; per-dimension explanation as a required field; accuracy tracking beginsEvery review emits two scores and a per-dimension explanation; coach and URL intake in production with zero gate failures; accuracy tracking records every seat on every review
    2. Fix-It and reports9 to 13Structured Fix-It on Pro and above; shareable report links; vertical auto-classification and the first three vertical templatesFix-It plan on 100 percent of Pro-and-above reviews; a shareable report passes the gate end to end; two of three vertical templates validated against known outcomes
    3. Enterprise controls and API14 to 18Rules Engine to Enterprise and White-Label; Review-as-a-Service API to Enterprise; vertical templates to fiveRules Engine live with three Enterprise accounts; the API completes 100 reviews with zero gate failures
    4. White-Label and GA19 to 24White-Label track live with custom domain, logo, email templates, portfolio console, tenant isolation; API generally available; five price points live on both surfacesFirst White-Label pilot renews past month one; API generally available to Enterprise; marketing site and portal show the five price points with no stale pricing anywhere
    +
    + +

    5.2 What comes next

    +

    v4.1 and beyond, in order of expected value:

    +
      +
    1. Outcome-calibrated scoring. The outcome cron accumulates from launch day at T+90, T+180, and T+365, and the schema carries the outcome columns from day one, so no migration is required. Once the corpus clears a minimum-N threshold, the accuracy track record becomes a published dashboard and then a candidate weighting input.
    2. +
    3. Competitive review comparisons. Full benchmarking of a new proposal against the corpus distribution, dimension by dimension and vertical by vertical, replacing the percentile context available today.
    4. +
    5. Market simulation. Replaces the static financial table with a twelve-month trajectory model, adding a per-review cost that the Enterprise and White-Label price points absorb.
    6. +
    7. Second-opinion audit agent. A dedicated blind-spot pass over the panel's own output, held until the corpus and accuracy data can measure its catch rate against a real baseline rather than an assumption.
    8. +
    9. Adversarial red-team per vertical. Industry-specific attack vectors, sequenced after the vertical classifier has a proven accuracy record.
    10. +
    + +

    5.3 Explicit non-goals

    +

    Scope is bounded on purpose:

    +
      +
    • No community or peer-review surface of any kind. A public roast board carries moderation and abuse overhead with no revenue tie, and it dilutes the independent-panel identity that is the entire product.
    • +
    • No weighting of live verdicts by reviewer accuracy in v4.1. Track only.
    • +
    • No product-facing outcome dashboard in v4.1. The cron runs and accumulates; the dashboard follows the data.
    • +
    • No standard-tier access to the four reserved premium models.
    • +
    • No proposal ghostwriting, ever. The coach asks questions. The panel critiques. Neither one composes the submitter's argument for them.
    • +
    +
    + + +
    +

    6Financial Model

    + +

    6.1 Unit economics and cost of goods sold

    +

    Costing assumption: 12,000 input tokens per scoring seat for a twenty-page proposal, roughly 40,000 characters, with the worker truncating above that; the synthesis gate reads roughly 20,000 input tokens. Per-seat cost is input rate times input tokens plus output rate times output budget. Dispatch is single-pass chat completion with no tool calling, which is the regime that makes a ten-seat panel cost cents rather than dollars. All margin guarantees below use the ceiling figure, which assumes every seat burns its full upper-bound output budget. The base figure is the typical case and is never used for a margin claim.

    + + + + + + + + + +
    RosterSeatsCOGS ceiling per reviewLoaded ceiling (x1.20)COGS base per review
    Free reduced panel5 model calls$0.15$0.18$0.10
    Standard panel10$0.36$0.43$0.25
    White-Label premium panel11$3.00$3.60$1.52
    +

    The loaded figure applies a flat 20 percent infrastructure and overhead buffer covering the application host, the database, object storage, the PDF renderer, and email delivery.

    + +

    6.2 Gross margin at full allotment consumption

    +

    Margin is computed at the pessimistic bound: every included review consumed, every seat at its ceiling output budget, loaded cost.

    + + + + + + + + + + + +
    TierRevenueIncluded reviewsLoaded COGS at full consumptionGross margin
    Free$01 lifetime$0.18 one timeloss leader
    One-Shot$291$0.4398.5%
    Pro$1198$3.4497.1%
    Enterprise$69950$21.5096.9%
    White-Label$3,000100$360.0088.0%
    +

    Overage is itself high margin by construction. Pro overage at $18 and Enterprise overage at $16 both sit far above the $0.43 loaded standard cost, so overage carries better than 95 percent margin while still reading as a nudge toward the next tier. White-Label overage at $28 against $3.60 loaded carries roughly 87 percent margin.

    +

    Stress case: if both premium reasoning seats in the White-Label roster burn a full 12,000-token output budget, per-review ceiling reaches $3.87, or $4.64 loaded. One hundred such reviews cost $464.40 against $3,000 revenue, which is 84.5 percent gross margin. The worst realistic case on the most expensive tier still clears 84 percent.

    +

    The structural conclusion is that cost of goods sold is not the constraint on this business. Even two hundred Enterprise reviews in a month cost roughly $86 loaded against $699 revenue. Review allotments are therefore set by value anchoring and ladder logic, not by cost recovery, and pricing pressure can be absorbed without touching the panel.

    + +

    6.3 Revenue projection, floor and ceiling

    +

    Two scenarios at month twelve post-launch, measured as monthly recurring revenue (MRR). Both are stated as assumption sets, not forecasts. Both assume every included review is consumed, which overstates cost and understates margin.

    +

    Floor scenario, month 12

    + + + + + + + + + + + +
    LineAccounts or volumeMonthly revenueMonthly loaded COGS
    One-Shot40 purchases per month$1,160$17.20
    Pro35 accounts$4,165$120.40
    Enterprise3 accounts$2,097$64.50
    White-Label0 accounts$0$0.00
    Total$7,422 MRR$202.10
    +

    Floor gross margin: 97.3 percent. Annual run rate at month twelve: $89,064.

    +

    Ceiling scenario, month 12

    + + + + + + + + + + + +
    LineAccounts or volumeMonthly revenueMonthly loaded COGS
    One-Shot150 purchases per month$4,350$64.50
    Pro180 accounts$21,420$619.20
    Enterprise14 accounts$9,786$301.00
    White-Label4 accounts$12,000$1,440.00
    Total$47,556 MRR$2,424.70
    +

    Ceiling gross margin: 94.9 percent. Annual run rate at month twelve: $570,672.

    +

    Free tier cost is a one-time charge per account rather than recurring, since Free grants one lifetime review. At $0.18 loaded per Free review, one thousand two hundred cumulative Free reviews cost $216 in total and six thousand cost $1,080 in total. Free is affordable at any signup volume the funnel can realistically produce, which is why the reduced panel exists rather than a time-limited trial.

    +

    Annual billing at $99, $582, and $2,499 per month trades 16.7 percent of headline revenue, two months free, for twelve months of committed cash and materially lower churn exposure. At the ceiling scenario, a fifty percent annual mix on Pro and Enterprise reduces month-twelve MRR by roughly $2,619 and converts roughly $156,000 of annualized revenue into prepaid commitment.

    + +

    6.4 What moves the model

    +

    Sensitivity, ranked:

    +
      +
    1. Pro account count. Pro is the volume tier and the largest single revenue line in both scenarios. It is the number to move.
    2. +
    3. White-Label logos. Each White-Label account is worth roughly twenty-five Pro accounts. Landing one changes the shape of the revenue curve; landing four is the difference between the floor and the ceiling scenario.
    4. +
    5. One-Shot to Pro conversion. One-Shot is priced as a bridge, and its value is mostly in what fraction of buyers subscribe after seeing the full ten-seat output once.
    6. +
    7. Enterprise seat expansion. Enterprise is the highest-effort sale, and the Rules Engine and API are the features that make it defensible rather than a volume discount.
    8. +
    9. Cost of goods sold. Last, and by a wide margin. A doubling of every model rate in the panel would still leave Pro above 94 percent gross margin.
    10. +
    +
    + + +
    +

    7Risk Assessment

    +

    Ranked by expected impact on the product at launch, each with the control that is in place.

    + +

    7.1 Reviewer availability High impact

    +

    The panel depends on nine vendors, and any one of them can rate-limit, exhaust credit, or return transport errors. Controls: a pre-dispatch health probe runs one cheap call per rostered model before spend is committed; every seat carries a named fallback binding; a review completes on a documented reduced panel rather than failing when a seat cannot be filled, and any review that ran reduced is flagged as such on the report and in the corpus record. The live-grounding seat is the single most availability-sensitive seat in the standard roster and is provisioned with a direct vendor credential rather than a shared route, plus a same-vendor-class fallback that is already live in the panel.

    + +

    7.2 Model output integrity High impact

    +

    Reasoning models can truncate structured output at low token caps, and some models constrain sampling parameters. Controls: per-seat output budgets are sized above the truncation threshold for every reasoning seat rather than set to a global default; seats with sampling constraints carry an explicit per-seat parameter quirk in the roster configuration; every seat response is schema-validated before it enters synthesis, and a seat returning unparseable output is retried once and then fails over rather than silently contributing a null score. Models without a published per-token input cost are not seated in any cost-guaranteed tier at all.

    + +

    7.3 Cost verification Medium impact

    +

    Two standard-roster seats bill through providers whose usage responses do not always carry a cost field, so their per-review contribution is computed from configured rates rather than reconciled against metered spend. Controls: those seats are provisioned with native vendor credentials so spend meters end to end; the margin guarantee uses the ceiling COGS across the whole roster, which absorbs a material rate change on any single seat; the combined ceiling contribution of the two affected seats is $0.0632 per review against a $0.36 total, so even a doubling of both moves loaded standard COGS from $0.43 to roughly $0.51 and leaves Pro above 96 percent margin. Base COGS figures are internal planning numbers and are not quoted externally.

    + +

    7.4 Sanitization and confidentiality High impact if it fails

    +

    A critique product handles unreleased strategy documents, and a leak of raw text or reviewer identity is an existential trust failure rather than a bug. Controls: the sanitization gate blocks rather than warns, and it covers the free-text explanation and remediation fields alongside structured fields; raw proposal text is schema-level ineligible for the corpus; corpus contribution is opt-in and defaults to false; the PDF pipeline extracts text from the finished file and asserts zero vendor and model identifiers plus the verbatim presence of the AI disclaimer before release; share tokens are unguessable, expirable, and revocable, and a revoked token returns not-found; tenant isolation is enforced on every corpus query and every scoped credential.

    +

    VerdictTank does not use submitted proposals to build its own products, train models, or inform its own proposals. A submission is processed only to produce that submitter's review, and is retained only for the submitter's own reference and legal record. Review processing does run through third-party AI providers under their own data-handling terms.

    + +

    7.5 Verdict liability Medium impact

    +

    A submitter can act on a verdict and attribute an outcome to it. Controls: a versioned, non-removable AI disclaimer renders on every report and cannot be templated away; inter-seat agreement and the panel spread are published beside every score so confidence is visible rather than implied; the liability cap is the greater of $100 or twelve months of fees; the status page distinguishes VerdictTank incidents from upstream provider incidents so an outage is not misread as a product defect.

    + +

    7.6 Positioning drift Medium impact and slow

    +

    The most likely way this product degrades is by drifting toward writing. Customers will ask for it, and a generated paragraph is easier to deliver than an honest score. Controls: the coach is architecturally forbidden from composing paragraphs, the Rules Engine is additive only and cannot suppress a dimension score, and no roadmap item shifts VerdictTank toward authorship. We critique them; we don't write them, and that is a product constraint, not a slogan.

    + +

    7.7 Concentration and capacity Medium impact

    +

    v4.1 is built and operated by a solo developer plus an AI-agent build pipeline, and the review worker is serial at launch. Controls: the review state machine is durable and replayable, with per-seat evidence stored so a partial panel resumes rather than restarting; the scale path is a depth cap plus parallel workers, which is a configuration change and not a redesign; the full customer-facing stack and the review engine sit on a single host with one state machine, so there is no cross-host coordination to debug under load. At the ceiling scenario, month-twelve volume is roughly 2,690 reviews per month, which a serial worker at a sixty-second critical path clears with substantial headroom.

    + +

    7.8 Distribution Highest impact overall

    +

    The panel works and the margins are structurally excellent, which means the binding risk is that not enough submitters find the product. Controls: Free is a real full-loop demonstration on a reduced panel rather than a time-limited trial; One-Shot at $29 removes the subscription objection entirely; the coach is unlimited on every tier and is the widest part of the funnel; shareable report links put a branded verdict in front of the submitter's own investors and advisors; the API and the White-Label track make other people's distribution into ours.

    +
    + + +
    +

    8Summary

    +

    VerdictTank v4.1 ships the complete critique loop: coach the draft, submit a file or a URL, run ten seats across nine vendors, publish a Proposal Strength Score and an Investor Readiness Score across ten dimensions, explain every low dimension with the specific missing artifact, hand back a prioritized Fix-It plan, re-score the revision with a per-dimension delta, keep it all in a queryable corpus, and ship the verdict as a branded PDF, a shareable card, or an API response with a sanitization gate on every path out.

    +

    Five price points cover the range from a single honest read at $29 to a fully branded reseller platform at $3,000 per month, and every paid tier clears 88 percent gross margin at full consumption with the standard tiers above 96 percent. The economics are settled. The architecture is proven. What remains is the build, then distribution.

    +
    We critique them; we don't write them.
    +
    + +
    + +
    + VerdictTank v4.1 - AI Proposal Review & Scoring
    + We critique them; we don't write them. +
    + + + diff --git a/index-v5.0.html b/index-v5.0.html index f376434..f163fa0 100644 --- a/index-v5.0.html +++ b/index-v5.0.html @@ -56,6 +56,8 @@ td{background:var(--panel);color:var(--ink)} .evolution-card--current{border-color:var(--accent);background:var(--panel2)} .version-badge{font-size:11px;letter-spacing:.1em;text-transform:uppercase;color:var(--warn);margin-bottom:8px} .version-badge--current{color:var(--accent)} +.superseded-banner{background:#fef3c7;border:1px solid #f59e0b;border-radius:8px;padding:.75rem 1rem;margin:0 0 1.5rem;font-size:.85rem;text-align:center;color:#7c2d12} +.superseded-banner a{color:#b45309;font-weight:600} .breakthrough-label{font-size:12px;color:var(--accent2);font-weight:600;margin-bottom:6px} .callout{background:var(--panel2);border:1px solid var(--line);border-radius:10px;padding:16px 20px;margin:20px 0} .callout--proof{border-left:4px solid var(--accent)} @@ -158,6 +160,11 @@ a{color:var(--accent)}
    + +
    + ⚠ SUPERSEDED. This v5 error-detection direction was deprecated; its validation thesis failed (-0.40 delta). The current product is the v4 single-pass scoring panel. Read the current proposal (v4.1) +
    +