A
ANTHOL AI · MELBOURNE · APRIL 2026
The Silicon Backup
of Human Civilization.
Anthol AI is building a permanent archive of expert judgment — the way top practitioners in their field actually think, reason, and decide.
Phase 1 (2026-2028): Anthol partners with world-class living experts under exclusive MELA licenses, trains their authorized silicon personas, and delivers professional consultation to paying users at 1% of the traditional cost.
Phase 2 (2027+): the same infrastructure becomes an open platform — anyone can commission a silicon persona of themselves, or of someone whose judgment they want to preserve for their family.
01 · THE VISION
Why this project exists.
Human civilization has built sophisticated backup systems for almost everything precious. Wikipedia archives facts. Archive.org preserves the open web. The Svalbard Vault safeguards the DNA of our food crops. But the most valuable and most fragile output of human experience — the accumulated judgment of top practitioners, the way they actually reason — has never had a systematic preservation layer. Anthol AI is building that layer.
Wikipedia
Backed up the world's facts. Any of us can now reference what humanity has learned, in seconds, for free. A civilizational infrastructure built in one generation.
Svalbard Seed Vault
Backed up the world's biological diversity. If a crop disappears somewhere on earth, its DNA is preserved inside a mountain in the Arctic, retrievable when needed.
Anthol AI
Backs up the world's expert judgment. The frameworks, priorities, and reasoning patterns of top practitioners — archived under authorization, queryable by anyone, preserved across generations.
What this is not. Not a chatbot. Not a RAG tool. Not a talent marketplace. Anthol AI is an archival infrastructure for a specific class of intangible human output — the kind that disappears when a practitioner retires, and has historically been recreated only by the next generation spending another thirty years learning the same lessons.
Not preserving records.
Preserving reasoning.
— Founding principle
02 · THE BUILD
Two phases. One goal.
Zero shortcuts between them.
The commercial model unfolds in two phases. The first earns the right to exist. The second is what the archive is for. Phase 1 is a curated partnership model with top experts. Phase 2 opens the platform to everyone. The infrastructure that makes Phase 2 safe, legal, and valuable is exactly what Phase 1 builds.
Phase 1 · 2026–2028 · Curated Expert Partnerships
Anthol signs exclusive MELA agreements with verified world-class experts in five domains — Enterprise, Finance, Legal, Healthcare, Education. Their licensed corpus trains an authorized LoRA persona. The persona is deployed to paying subscribers (consumer) and to enterprise seats (B2B).
Revenue streams
· Consumer subscriptions · Scholar $29/mo · Master $199/mo · Legacy $4,990/yr
· Enterprise Copilot seats · ~$8,500 per seat per year
· High-end IP licensing to institutions · discretionary
What accumulates in parallel
· Compliance infrastructure across 5 jurisdictions
· Technical stack for persona training and serving at scale
· Brand trust with first tier experts and regulators
· A signed roster of flagship IP
Phase 2 · 2027+ · Open Archive Platform
Once the compliance and technical layers are mature, Anthol AI opens to anyone. Any individual can commission a silicon persona — of themselves, of a family member whose judgment they want preserved, of a mentor, of a founder, of a family business head.
Pricing model (directional · to be validated by market)
· Storage fee · preserved persona kept queryable · monthly or annual
· Call fee · per-interaction usage for subscribers or family members
· Inheritance subscription · multi-decade service covering family access rights
Who this serves
· Family offices preserving the founder's business reasoning
· Professionals preserving specialized knowledge for apprentices
· Institutions preserving cultural and linguistic knowledge
· Individuals preserving their own thinking for their descendants
The two phases are one product, not two companies. Phase 1 is not "revenue to fund Phase 2." Phase 1 is Phase 2's infrastructure being stress-tested with real users, real regulators, and real experts. When the platform opens in Phase 2, every system — the MELA template, the opinion letter framework, the persona training pipeline, the three-layer memory, the cross-jurisdictional serving — has been proven in production. No competitor can open a comparable platform in 2027 without having spent 2026-2027 building these exact same foundations.
03 · WHY NOW
Possible for the first time.
Only for 24 months.
The idea of "a digital version of a specific person whose judgment you can query" has been speculated about for 40 years. What changed in 2025-2026 is that three independent enabling forces matured simultaneously. They open a window. On current trajectory, that window begins to close around 2028 — after which the space becomes commoditized, over-regulated, and dominated by whoever established the category first.
Models can hold a person
Foundation models now reproduce a specific practitioner's way of reasoning, not just their facts. Extended thinking, long context, and stable persona consistency crossed a threshold in late 2025. "An AI that reasons like Dr. X" is a technical reality, not a demo.
Training costs $50, not $5M
LoRA / DoRA fine-tuning produces a 50-500MB adapter that holds a persona, trains in hours, and costs $15-50 in compute. Three years ago a comparable persona model required $5M+ and a research team. The economics of per-person AI are genuinely new.
The law finally recognizes this
Three major regulations landed in 2025-2026: California AB 2602/1836, China GB 45438-2025, EU AI Act Article 50. For the first time, "an authorized digital persona" is a category the law names and protects — no longer a grey zone that could be shut down overnight.
The window is open — for 24 months
By 2028, these three forces become table stakes. General-purpose platforms will have retrofitted compliance. Persona fine-tuning will be a commodity API. Major regulators will have issued enforcement precedents. The period to establish a category-defining position — the Anthol of authorized silicon personas — runs from now through mid-2028. After that, the space is crowded and locked down, and the only durable assets are early authorization libraries and early compliance estates. Both of which take 18-36 months to build from scratch.
04 · TECHNOLOGY
Not building models.
Building the containers that hold them.
Foundation models are utilities now — capable, abundant, and dropping in price 30-60% every six months. Building one is a $50M+ bet with no durable lead. Anthol lives one layer above: turning a generic model into a specific authorized person, then routing each query to whichever underlying model meets the quality bar at the lowest cost. The three architectural decisions at the core.
Decision 1 · The persona is a LoRA adapter, not a fine-tuned model
Each authorized persona is a 50-500MB LoRA / DoRA adapter trained on a curated corpus — the expert's public work, proprietary material licensed under MELA, and expert-reviewed synthetic dialogues. Training cost per persona: $15-50. Recalibration cadence: monthly, with the expert in the loop during Phase 1 · with the designated reviewer during Phase 2. The adapter is the asset Anthol owns. The foundation model underneath is a pipe, rented from whoever prices it best.
Model-agnostic Adapter-centric Reviewer-in-loop Platform-portable
Decision 2 · Every query routes to the right model
One user message, the cheapest model that meets the quality bar. A brief confirmation runs on Haiku ($0.03/message). A nuanced strategic question routes to Opus or a GPT flagship. A Chinese-language query lands on Qwen3-Max. High-volume simple tasks route to DeepSeek V3.2.
Blended API cost is 60-75% lower than single-flagship routing. As the field keeps dropping prices, the cost curve bends with theirs automatically — no re-architecting required.
Decision 3 · Three memory layers, not one context window
A chatbot has a context window. A relationship needs memory.
L1 · Knowledge graph over the expert's corpus (LightRAG).
L2 · Semantic user facts, updated after each session (Mem0).
L3 · Temporal graph of the user's goals, preferences, decisions over time (Zep Graphiti).
Memory is what turns "useful" into "trusted". It is also what makes the Phase 2 platform feasible — each user's interactions with each persona compound into a permanent, portable asset.
The six-layer stack
06
Application · safety · compliance
Product surface · C2PA watermark · GB 45438 dual-watermark · MELA audit trail · EU AI Act Art.50 disclosure. Compliance built in, not retrofitted.
05
Perception · presentation
TTS · STT · real-time audio-video · digital-human rendering. End-to-end first-token latency target < 1200ms.
04
Agent orchestration
State-machine routing · tool calls · multi-step reasoning. Open-source agent framework, auditable state graph.
03
Memory · retrieval
LightRAG knowledge graph + Mem0 user facts + Zep Graphiti temporal graph. Three layers, each doing what it alone does well.
02
Model · LLM core + LoRA
Foundation model routed per message. Expert LoRA adapter loaded for persona. The adapter is the asset.
01
Infrastructure
Cloud · GPU · storage · CDN. Dual-region deployment. Australian data sovereignty primary, secondary regions per jurisdiction.
Architectural thesis, not a shipping claim. Anthol's V1 production stack is being built now. What is above are the decisions already made about how it will be structured. Production engineering details, latency benchmarks, and infrastructure topology become public when V1 ships to paying users.
05 · PRODUCT
What a paying user
actually experiences.
The product is a three-step experience. A user arrives and sees the authorized experts. They pick the channel they already use daily. They enter a space that learns them, session after session, until consulting the expert becomes ambient. This is Phase 1's consumer surface. In Phase 2, the same surface works for any person who has been granted a silicon persona — and for the families who preserved them.
1 · Five domains at launch
One MELA-authorized flagship per domain. The Enterprise flagship is signed. The other four are in active MELA negotiation with verified candidates. Identity and full credentials of each candidate are available in the investor data room.
E
Enterprise
Org design · CEO coaching
MELA S · Signed
F
Finance
Portfolio · risk
MELA A · MoU
H
Healthcare
Clinical reasoning
MELA A · MoU
L
Legal
Contract · triage
MELA A · MoU
ED
Education
Learning paths
MELA A · MoU
— See it in motion
Explore the living interface prototype.
Twelve screens tracing a founder's first six months with the Enterprise flagship E.: the consultation space, Day 1 through Month 6 of relationship evolution, Memory transparency, cross-channel presence, voice call, Phase 2 commission entry.
Open the prototype →
12 screens · ~6 minutes to walk through
06 · COMPLIANCE FIRST
Forty percent of the first $1M
goes to law. Before growth.
Most first-time founders spend angel capital on customer acquisition. Anthol reverses the priority. Of the $1M raised, $400K goes into the legal and compliance layer — before the first customer, before the first ad, before the second engineering hire. This is the order of operations that makes Phase 2 possible.
Where the $400K goes
1
MELA contract architecture · $180K
Melbourne Tier-1 firm drafts the master agreement. Five-layer jurisdictional design (commercial / IP / data / dispute / succession). This is the template every future user of the Phase 2 platform will sign. Drafted once at this depth, then reused. The alternative is years of amending bespoke contracts one at a time.
2
Four jurisdictional opinion letters · $120K
US (San Francisco) · HK · CN (Beijing-Shanghai) · MY. Each opinion letter certifies MELA enforceability under local law. Without these letters, the MELA is enforceable only in Australia. With them, Anthol operates — and Phase 2 onboards users — across every jurisdiction its experts, users, and future persona commissioners live in.
3
CLCO as co-founder · $60K + equity
Chief Legal & Compliance Officer joins as co-founder-grade equity holder. Not a hire, not a retainer. A structural partner whose compensation is tied to the long-term compliance asset. Senior regulatory lawyer in their fourth decade — not hireable at angel salaries without the equity seat.
4
Technical compliance build · $40K
C2PA content credentials on every persona message. GB 45438-compliant dual watermark for China-facing outputs. EU AI Act Article 50 disclosure layer. Audit trail for every response. Compliance as infrastructure, not post-launch features.
Why the order matters
Compliance is not Anthol's moat. Compliance is Anthol's oxygen. Without it, Phase 1 cannot take enterprise contracts; without it, Phase 2 cannot open to the public. Every other layer can be catching-up work. This one cannot.
Product can be rebuilt in 3 months
A better web client can ship in one funded quarter. Design taste, engineering speed, feature parity — all catchable with budget.
A brand can be rebuilt in 3 months
Launch narratives, hero copy, visual system — re-doable with $500K and a good agency. A competitor with $10M matches the surface within two quarters.
A 5-jurisdiction contract cannot
Five-jurisdiction MELA drafting, executed agreements with top-tier experts, and a co-founder-grade CLCO take 18 to 36 months to replicate — with money, connections, and luck. The only part of Anthol that cannot be compressed by more funding.
What the remaining 60% does. Core team salary ($300K · 4 FTE for 12 months) · first flagship expert signing, corpus licensing, LoRA build ($150K) · AI infrastructure + API costs ($100K) · seed users, brand, founder reserve ($50K). Total $600K. Enough to reach V1 and begin measurable revenue. Not a dollar on paid acquisition in V1.
07 · THE MOAT
Four land-grab positions.
Whoever gets all four wins.
The category of "authorized silicon persona infrastructure" is not a feature race. Winning the category requires being first and decisive on four specific structural positions. Any three out of four and a well-funded competitor catches up. All four, and the next entrant needs 24-36 months and $50M+ just to reach parity.
01 · Compliance estate
First-mover on the legal layer. MELA template drafted across 5 jurisdictions. Four opinion letters in place. CLCO seated as co-founder. C2PA + GB 45438 + EU AI Act Art. 50 shipped Day-1. A well-funded competitor with $10M in 2027 still faces an 18-month clock on this layer alone — by which time Anthol is Pre-A with the MELA template proven in production through multiple completed signings, and the compliance audit trail accumulating.
18-36 mo replacement cost Day-1 asset
02 · First-tier expert IP
Exclusive MELA signatures with top-tier experts. Each signing is a multi-month trust-building process — credentials, legal review, family consultation, NDA chains. The first wave of signings (2026-2027) locks the recognizable names out of any competitor's roster. Later entrants can only sign second-tier names, which markedly changes consumer positioning.
Exclusive through 2035 Brand multiplier
03 · Architecture design
Technical architecture that supports Phase 2 from Day-1. Model-agnostic routing, adapter-centric persona, three-layer memory, cross-channel state sync. Competitors who optimize for Phase 1 alone (single-model, single-channel, single-context) have to re-architect when Phase 2 becomes relevant — losing 6-12 months at a critical moment.
Platform-ready on Day-1 No re-architect debt
04 · Capital + brand cadence
Angel → Pre-A → Series A → IPO sequence. Each round is not just capital — it is brand signaling to the next cohort of expert candidates and enterprise buyers. The companies that complete this cadence in 2026-2028 become the category in investor, expert, and regulator minds. Late followers compete for scraps of the conversation even if they have better tech.
24 mo IPO-ready Category-defining
Three out of four is a fast follower.
Four out of four is the category.
— On running hard on all four simultaneously
08 · BUSINESS CASE
The numbers for the angel round.
The quantitative basis: valuation framework · four return scenarios · four real risks with mitigations. Full cap-table allocation, monthly financial model, comparable-company deep dive, and exit mechanics live in the investor data room under NDA.
Valuation basis · $5M post-money anchor
The angel round
$1M at $5M post-money · 20% dilution · $4M pre.
Post Day-1 cap table: Founders 75% · ESOP 5% · Angels 20%.
Through successive rounds (Pre-A at $25M post / Series A at $100M post / optional Series B), angel equity at IPO: ~10.5% fully diluted.
The category comparable
Harvey AI — the vertical compliance-first AI anchor — trades at ~58× ARR on $11B valuation with $190M ARR (March 2026 round · GIC / Sequoia co-led). In 9 months, Harvey's valuation grew 2.2× and its ARR grew 2.5× — the capital market continues to reward the compliance-first vertical AI category, not a one-time hype cycle.
Cursor · $29.3B / $2B ARR · ~14.6× ARR.
Lovable · $6.6B / $400M ARR · ~16.5× ARR.
Anthol Year 3 target: $22M ARR. A conservative 18× ARR multiple delivers $400M IPO valuation — approximately one-third of Harvey's current multiple, applied to roughly one-eighth of Harvey's ARR scale. Smaller-ARR segments historically carry higher multiples, which leaves room above the 18× anchor in the upside scenario.
Four return scenarios · 3.5-year horizon
HORIZON = ~30 MONTHS TO NASDAQ LISTING + 12-MONTH STANDARD POST-IPO LOCK-UP + ORDERLY DISPOSITION
89
Upside IPO · 89× · ~261% IRR
$850M IPO valuation · 10.5% fully-diluted · 3.5 years. Assumes Phase 2 platform begins material contribution by Y3.
42
Base IPO · 42× · ~191% IRR · Featured
$400M IPO valuation · 10.5% fully-diluted · 3.5 years. 18× ARR on $22M Y3 ARR. Conservative against Harvey's ~58× (at $11B / $190M ARR), Cursor's 14.6×, Lovable's 16.5× — and applied to a much smaller ARR base, where category multiples are historically higher.
39
Base M&A · 39× · ~524% IRR
$300M strategic acquisition · Series A stage · 2 years. Short-duration exit delivers the highest IRR of the four scenarios. Likely acquirers: Anthropic / OpenAI (platform extension) · Pearson / TAL (education integration) · Palantir / Adobe (enterprise knowledge vertical).
16
Downside · 16× · ~300% IRR
$125M distressed M&A · 2 years. Even the worst-case scenario returns positively — and its 2-year duration makes its IRR exceed Base IPO's. This is the asymmetry created by compliance-first positioning: the downside still has a market.
Cap-table allocation detail (Pre-A · Series A · Series B · Pre-IPO sequence with dilution percentages) and monthly cash-flow model (Y1 $1.1M ARR build · Y2 $6.8M · Y3 $22M) are disclosed to qualified investors in the data room under NDA.
Four real risks · mitigations
Risks identified through internal review. Stated openly, with specific mitigations rather than generic reassurances.
Risk 01 · Compliance window closes faster
The risk. If EU AI Act enforcement ramps earlier than 2028, or if general-purpose platforms accelerate compliance retrofits, the 18-24 month lead compresses to 12-15 months.
Mitigation. The 24-month IPO-ready plan is itself the mitigation. Once Anthol is listed, "capital market identity" becomes a non-replicable moat that supersedes the compliance window alone. Compounding effects from early expert signings (Y1-Y2) are locked in by MELA exclusivity through 2035 regardless of window length.
Risk 02 · Flagship expert execution
The risk. Y1 revenue depends on the first flagship persona launching on schedule. A 3-6 month delay pushes Y1 ARR from $1.1M target toward $600K, affecting Pre-A valuation and timing.
Mitigation. Candidate redundancy + enterprise revenue hedge. Multiple cross-border candidates are in concurrent NDA / MoU stages — first-signing individual risk is absorbed by the candidate pool. Anthol Copilot enterprise seats ($8,500 ARPU) provide a $300-500K Y1 floor even if consumer launch slips 6 months. Candidate roster and contract status disclosed in DD confidential annex.
Risk 03 · Foundation-model agent generalization
The risk. Claude / GPT / Qwen agents approaching "general expert" quality within 18 months dilutes the raw LoRA-adapter differentiation.
Mitigation. The asset is not the model. It is the authorization. Foundation models cannot legally use the name, likeness, or proprietary methodology of a MELA-signed expert. The regulatory window converts "general AI capability" into "specific named authorization" — exactly what cannot be generalized by a better underlying model.
Risk 04 · IPO window uncontrollable
The risk. Month 30 Nasdaq listing target can be delayed 6-18 months by macro conditions, IPO market freeze, or geopolitical events.
Mitigation. Dual-path + floor. Nasdaq primary (Month 30) with 6-12 month flex. Hong Kong spin-off / secondary listing as Year 4-5 liquidity path. Strategic M&A at $300M base case provides 39× floor return. Founding team has direct operating experience across US and HK capital markets — path switching is not theoretical.
What is in the data room. Full cap table across Day-1 → Pre-A → Series A → Series B → Pre-IPO · monthly cash-flow model for Years 1-3 · unit economics per tier (Scholar 5.8× LTV/CAC · Master 10.0× · Enterprise 28.7×) · comparable-company deep dive · exit scenario mechanics · first flagship expert identity and signed MELA · MELA opinion letter status per jurisdiction. Available to qualified investors under NDA.
ANTHOL AI PTY LTD · MELBOURNE · VICTORIA
THE SILICON BACKUP OF HUMAN CIVILIZATION
Strictly confidential. Prepared for friendly angel investors, family offices, and industry founders under NDA. Forward-looking statements involve risks and uncertainties. Not an offer to sell securities.