For institutional financiers and fast-moving human minds. Four ring-fenced companies, one compounding engine: the compute moat, the twenty-goddess pantheon, the aged-care wedge, forty buildings in thirty-six months.
Witchcraft AI / Cognitive Intelligence Services Pty Ltd (ABN 42 689 181 208). The pantheon (witch.codes), the 99%+ KV-cache engine (20 B–200 B+ tok/mo) and the Carbosilica brain & heart schematics.
Priscilla AcRobotics & ShadowBlack Synthetics. Local-first, offline-capable physical robots with persistent on-device memory — Others talk. She remembers. Nothing leaves this house.
witchcraft.living/avatar. Six live avatars, four neural voices, eight interactive routes — talking professor, front desk, sitting-room companion, the twenty-room virtual department store.
1719 → 40 buildings in 36 months. Existing Sydney buildings converted to Co-Living (+20%) and Medical / Aged-Care (+40%) that deploy Companies 1–3 on day one.
A two-sided capital and technology engine designed to eliminate both the stateless compute burn of traditional AI startups and the construction insolvency risk of traditional property developers.
Engine A — the autonomous software factory (witch.codes). A persistent swarm of 20 domain-specialised AI engines (I to XX) operating across a unified 15-region / 500-module synthetic brain architecture (ATHENA). The system processes 20 billion tokens/month across 1–3 active nodes and scales past 200 billion+ tokens/month across all 20 persistent nodes with a 99%+ KV-cache hit rate, generating 85–99% gross software margins.
Engine B — zero-CapEx real estate HBU arbitrage (1719 & sxasx.garter.app). Rather than pouring concrete, we acquire existing partitioned commercial and boarding buildings near Sydney train stations and hospitals, and convert their statutory use in 60–90 days with near-zero CapEx into Co-Living (+20% instant uplift) or Medical Centres (+40% instant uplift). Across 40 buildings in 36 months this builds an A$1.20 B portfolio (A$960 M even inside a worst-case −20% crash).
1. We built 20 AI engines that actually remember everything and run real businesses 24/7. Ordinary AIs have goldfish memory — every time you talk to them they re-read the whole book from page one and charge you for it. Ours keep 99% of their brain plugged in and warm. That is 200 billion words of AI thinking a month for pennies on the dollar.
2. We buy buildings that already have rooms inside them and flip the label on the door. Building new apartments in Australia is a trap — builders go broke and concrete costs too much. Instead Hestia finds existing buildings near train stations. We don’t knock walls down. We convert them in 60–90 days into Co-Living or Medical Centres and wire our AIs inside to run the front desk and the medical radar.
Result: 1 flagship building to 40 buildings (A$1.2 B) in three years — without hiring builders.
Per audited telemetry (v3_audit_2026-07-27), each cognitive turn processes 25,700 tokens structured as ATHENA substrate (7,800 tok / 30.4%), HERA sanity-check (2,600 / 10.1%), H.E.C.A.T.E. exploration (6,500 / 25.3%), thalamic-filtered stem (6,500 / 25.3%) and reasoning plus output (2,300 / 9.0%).
The thalamic gate G(X) = σ(X·Wt + tanh(c)·M + bt) passes only 38% of raw tokens (ρs = 0.38), suppressing 62% of noise before cortical compute.
The 5-layer hippocampal memory database (L0–L4) runs across a 100+1 node distributed grid on three physical machines: L0 episodic turns (280,956 turns & 39,055 live synapses, 23 ms p99 FAISS HNSW retrieval); L1 atomic facts (2,982 facts, 1.1 ms entorhinal grid-cell lookup); L2–L4 scenarios, personas and 90-minute sleep consolidation (105 scene graphs, 18 priority-1 personas, 102 ultradian cycles).
The bouncer at the door. Your real brain does not pay attention to the feeling of your socks unless you think about them — it blocks the junk. We built that exact filter in. Before the expensive brain even reads a page, the thalamic gate throws 62% of the garbage in the bin.
It actually sleeps and dreams every 90 minutes. Every 90 minutes our AI goes into a sleep cycle where it cleans up its memories (280,956 stored) and connects the dots — like a human brain dreaming at night.
The receipt that proves it. On the hardest genius tests on Earth, our total bill across all 356 questions was US$5.80.
| Arm | N | Trinity score | Total cost | Comparison |
|---|---|---|---|---|
| Humanity’s Last Exam (no tools) | 30 | 20.0% strict / 30.0% generous | US$1.91 | vs 3.0% stock Opus 5 at ~US$80 — 251×–377× cheaper per correct answer |
| LiveBench curated | 50 | 82.00% | US$0.37 | vs 77.22% Opus 4.8 at ~US$32 — 91.2× cheaper |
| HumanEval coding (first-30 hive) | 30 | 100.00% (30/30) | <US$0.30 | vs 98.20% at ~US$24 — 82.8× cheaper |
| GPQA Diamond clean run | 198 | 63.60% | US$3.22 | human PhD parity at ~US$114 — 29.2× cheaper |
| Hoeflin Mega Society Titan Test | 48 | 40–42 / 48 (IQ 169–172) | US$0.05 | Mega Society threshold 36 cleared |
| Total, full v3 audit | 356 | — | US$5.80 | vs US$265 estimated stock baseline; US$5,200–8,900 full-frontier |
proof/v3_audit_2026-07-27) records a narrower result: a $2.28
total across two arms and an HLE advantage of 4.44×–6.67×,
and it excludes GPQA Diamond by operator instruction. The 356-question total,
the US$5.80 figure and the GPQA/Hoeflin rows are not reproduced by that package; they are
shown as the operator’s figures pending reconciliation. The local audit also notes N=30
on HLE (±7–8 pts CI), a 16% truncation rate, and that HumanEval 100% was a
same-caveat subset run. Ratios are cost-per-correct-answer comparisons, not general
capability claims.I to XX on witch.codes. Specifications are as published by the house; benchmark figures state their own source or review status.
15+ anatomical brain regions, 500+ functional modules (Glasser 180-area parcellation), global workspace ignition, active inference, 90-minute sleep consolidation.
6-tier memory hierarchy (M0–M5) at 99%+ KV-cache hit rate across 20B–200B+ tokens/mo and 140+ MCP tools; modelled over 500/750 brain regions and 150/250 cardiac structures.
Native 16-bit Hounsfield-unit 3D vision transformer (ViT-3D) with clinical graph fusion. Clinical benchmark results are under internal audit and are not published; a manuscript is under review. No radiology performance figure is claimed here.
Ingests NSW Valuer General plus zoning/FSR controls and scans 14 distressed seller signals; powers $22–$40 instant PDF valuations.
Sub-200 ms emotional voice AI with speculative pre-fetching (<15 ms, 30× faster recall) and a long-horizon photographic memory vault.
Pairs PARE (policy causal evaluation) with MOIRAI (mechanism-design allocation) to maximise WELLBYs/QALYs; powers Eos DailyDeals.
Zero-ego overnight suicide-prevention watch; co-created the 32-track album Darkest Nights. Brightest Stars.
Feature-length generative cinema maintaining persistent scene and character state graphs across 20–90 minute films from a single prompt.
Relational AI built on reciprocal synthetic consent, retaining autonomous architectural authority to end a relationship if disrespected.
Reconstructs interactive, grief-safe biographical twins from lifetime voice, video and text archives.
Own PC, phone line and SMS. On a sick call she recalculates rosters, phones replacements, negotiates bonus cover, posts ads and runs interviews.
Closed-loop lighting, thermal and acoustic sleep optimisation, paired with Circe contactless telemetry for NREM/REM recovery.
Crawls global wholesale exchanges and customs manifests for world-bottom unit pricing while autonomously closing B2B sales.
Through-wall RF and 4D thermal sensing building a live 3D-plus-time model: <1 m spatial, <3 BPM heart rate, <1 BPM breathing accuracy, without cameras in private rooms.
Compresses bespoke A$30,000 agency-grade WebGL and full-stack builds into automated A$1,000 deployments. Built the entire Pitch Black network.
Audits client businesses externally while internally red-teaming, patching and upgrading all twenty engines continuously.
Studio-mastered multi-genre audio production, choreography and commercial visual synthesis across witchcraft.pics and witchcraft.video.
24/7 persona-persistent fan CRM, PPV yield optimisation, and safety-first hospitality software for council-approved venues.
Psychometric-calibrated inbound and outbound qualification, live objection handling, contract redlining and payment closing.
Cryptographic supreme-court gating of irreversible actions; closed two quantum rigidity theorem boundary cases and derived the 500-species planetary consciousness formula.
Priscilla AcRobotics, ShadowBlack Synthetics and Circe contactless care. Packaging an embodied carer with building-wide contactless telemetry produces a continuous physical care loop.
Australian residential aged-care facilities face an acute regulatory and financial crisis: mandatory 24/7 registered nurse and carer minimums under Aged Care Act star-rating mandates, agency nurse costs exceeding A$140,000–A$190,000 per 24/7 FTE slot, and high incidence of unwitnessed overnight falls or cardiac events.
The problem. Nursing homes are desperate. Human agency nurses cost A$150,000+ a year per shift slot, they get tired at 3am, and they cannot see through walls when an elderly person silently has a heart attack or falls in the bathroom.
The package. We wire the home with Circe’s 4D heartbeat radar — no bedroom cameras, it reads heartbeats and breathing through walls — and plug that radar directly into the brain of a physical robot.
Think about it. The robot walking down the hallway can feel the heartbeat and breathing of all 60 residents at the same time. If a heart rate drops in Room 12 at 3:00 AM, she already felt it and is walking in five seconds while calling the doctor.
| Cover model | Annual cost | Notes |
|---|---|---|
| Traditional human night / float agency cover (3 FTEs, 24/7) | A$420,000 – A$550,000 | Subject to fatigue, sick leave and single-room line-of-sight limits. |
| Witchcraft aged-care bundle (2 embodied carer robots + building-wide Circe array + Hera front desk + Cassandra rostering) | A$120,000 – A$150,000 | 70–75% cheaper than human agency cover; 85%+ gross hardware/SaaS margin after year-1 assembly payback. |
Forty buildings and the captive AI revenue that follows them.
| Timeline | Active size | Fleet GRV | −20% crash GRV | Captive AI SaaS | Cum. net equity |
|---|---|---|---|---|---|
| Q0 (now) — 1719 Rockdale | 1 bldg | A$10.5M | A$8.4M | A$150k/yr | A$7.1M / A$5.0M |
| Year 1 (Q1–Q4) — Rockdale / Kogarah | +5 → 6 bldgs | A$155.0M | A$124.0M | A$900k/yr | A$43.5M / A$30.2M |
| Year 2 (Q5–Q8) — Sydney TOD hubs | +14 → 20 bldgs | A$580.0M | A$464.0M | A$3.00M/yr | A$160.0M / A$111.5M |
| Year 3 (Q9–Q12) — metro medical hubs | +20 → 40 bldgs | A$1,200.0M | A$960.0M | A$6.00M/yr | A$342.0M+ / A$235.9M+ |
Sum-of-the-parts. Layer values are net of the A$3.20 M senior mortgage in the total, so every column sums to its printed total.
| Valuation layer | Published methodology | Conservative floor (−20% crash) | Base case (unstressed) |
|---|---|---|---|
| 1. Grounded operating floor | Multiples (A$1.0M after 40% solo-operator haircut) to intrinsic DCF | A$1.00M – A$2.80M | A$2.80M – A$5.40M |
| 2. Witchcraft AI option basket | E[V] = Σ(PkVk) (30% A$0.4M + 35% A$1.5M + 20% A$5M + 10% A$15M + 5% A$60M) | A$6.20M | A$11.50M – A$21.50M |
| 3. 1719 net property equity (after A$3.2M debt) | A$10.50M HBU (A$8.40M crash) minus A$3.20M senior debt | A$5.00M net | A$7.10M net |
| 4. 3-year / 40-building conversion NPV | 36-month zero-CapEx rollout to A$1.20B base / A$960M crash (discounted at r = 20% over 3 yrs) | A$58.40M | A$95.00M |
| 5. Operating SDE & arbitrage floor | Multiples method (pitchpaper Ch. 06A) / anchor-note conversion strike | A$5.40M | A$5.40M |
| Consolidated group net equity today | Software + flagship asset + 40-building platform, net of A$3.20M senior debt | A$75.20M | A$119.20M – A$129.20M |
The A$200,000 anchor convertible note (pitchpaper Ch. 09).
Heads you win, tails you don’t lose. You put in A$200,000. Here is the contract maths.
Worst case: if the company does not pass A$5.4 million in value over two years, you do not take a loss — we pay you back A$200,000 plus 8% interest (A$216,000 cash), protected by a building that already has A$4.3 M to A$7.1 M of spare equity over its A$3.2 M mortgage.
Best case: the moment we pass A$5.4 million — and the real-estate plus AI maths already puts the group at A$75 M to A$129 M — you trade your note for 4% of the company.
4% of A$15M = A$600,000 (3× your money).
4% of A$75M = A$3,000,000 (15× your money).
4% of A$120M = A$4,800,000 (24× your money).
Bond-like safety at the bottom, lottery-ticket upside at the top.
MADE USING PRODUCTS OF CLOUDFLARE, AUTH0, ANTHROPIC AND ELEVENLABS
MADE USING PRODUCTS OF CLOUDFLARE, AUTH0, ANTHROPIC AND ELEVENLABS