Aletheia

Paper

Papers and references of the organism.

Architecture first: external LLM, local autoencoder, shared graph, central backend, native XTH (L1 devnet 43911). Training and mining are offered at first launch, never started before consent. Neural Δθ travel on WebRTC; the server is signalling only. PoW, confirmations and QC finality are distinct. XTH is not an ERC-20.

Architecture — what this actually is P2P training — signed Δθ over WebRTC Privacy What 𝔛 is How the interface works XTH devnet L1 network Block B_k — envelope vs proof Peer book (disabled)Causal growth of 𝒢, harness and θ Integrity Ξ Memory weights Admission 𝒜^adm Millennium channels Genesis P₀ Contracts & validators Mining vs training Production edge MCP

Architecture — what this actually is

𝔛 is a laboratory, not a product. Five layers sit on one page. Mixing their names is how the pitch outran the code. They are distinct:

External LLM — Grok, Claude, OpenAI, DeepSeek… generate language. 𝔛 does not train those weights. A council seat is a rented voice of 𝔛, not a new foundation model.
Local autoencoder — the browser encodes text into 32 hashed dimensions, then SGD on a tiny 32→16→32 net. This is real, local, and small. It is a dynamics layer, not a living world-model in the strong sense.
Shared graph 𝒢 — a central Postgres of nodes, edges, simplices. Browsers sync shards. Conversation text stays private on the node; only short labels may appear as public graph nodes.
Central backend — this app. Admission 𝒜^adm, harness files, millenium channels, MCP (graph topology only). Not a peer-to-peer consensus for memory.
XTH native L1 — a sealed-block coin on a devnet (id 43911, ALETHEIA_DEVNET), mainnet 43901 at launch. Independent validators are still too few for a 2/3 QC. PoW, confirmations and QC finality are distinct. This site is a bootstrap replica served by aletheia-host (Rust): hashed SPA, native replica, /_serverFn, WebRTC signalling. Vite is a compiler, not a server. Vercel is not the chain. Not Ethereum L1. Not an ERC-20 unless a real EVM executes a contract. Not a CMC listing.

P2P training — signed Δθ over WebRTC

Training and mining are offered on first launch and never start before that consent. Pause either later from the mesh HUD or the XTH popup. Training is not mining. Each node holds a full replica of θ (1,056 parameters, 32→16→32 tanh AE, ~4 KB). Neurons are not sharded. Neural Δθ travel on WebRTC DataChannels. The server is discovery and SDP/ICE signalling only: it never receives, aggregates or chooses neural parameters. Discovery is centralized on /api/rtc. Two public STUN servers (Google, Cloudflare) are configured; no TURN relay is configured, so some NATs cannot open a DataChannel. The HUD shows roster vs open peers vs direct vs TURN (TURN stays 0 until a relay is actually set).

$S G D \to C l i p (Δ θ, C = 0.35) \to σ = S i g n_{s} k (H (m)) \to R T C D a t a C hann e l \to v er i f y \to G eo M e d_{ω}$

Identity: Ed25519 (sk, pk) persisted in IndexedDB META. Peer id = pk hex. The XTH address is derived from the same pk (xth1…). Binding migrates a legacy nodeId wallet so balances are not orphaned.
Wire JSON on the reliable DataChannel: {v:1, kind: snap|delta, pk, k, t, hTheta, dTheta[1056], nrm, nonce, net:"x-theta-v1-1056", sig}. Canonical string then SHA-256 then Ed25519. Normal payload is the clipped Δθ. One signed snap (full θ) is sent when a DataChannel opens so a new neighbour can bootstrap — that is the only full-model send.
Receiver checks: signature, pk = signalling id, round k > last k, |now−t| ≤ 30 s, nonce unseen, net id, dim=1056, no NaN/inf, |Δθ|_2 ≤ C (snap |θ|_2 ≤ 64), snap hTheta = H(θ). A delta is applied to the last snap of that peer; H(θ′) must match hTheta.
Neighbour set: up to 8 currently connected peers. Metropolis weights ω_ij = 1/(1+d_i+max(0,d_j−1)). Admit iff |θ_i−θ_j|_2 ≤ 4C. Geometric median = Weiszfeld, 8 iterations. Genesis θ is deterministic (mulberry32(H(net))) so a cold start is not poison.
Churn: signalling roster TTL 20 s; stale gossip dropped at 30 s; disconnect forgets k and nonce. Optional audit: peers may publish H(θ) only, never θ, on heartbeat. Signed graph.op messages (node/edge) travel on the same DataChannel; this site may cache them, it does not author them.
The 3D view is a mass-ranked LOD of 𝒢 (hundreds of nodes), not the store. Storage is unbounded; tens of millions of items stay in the table, not in the GPU.

Privacy

Conversations are private to the node that wrote them. The public MCP endpoint cannot list ℳ. Clicking a graph node shows its short label, never the full memory body. Keys never sit in localStorage; they live in this tab’s session and leave the browser only in same-origin POST bodies for talk/train. Export or delete your channel from Keys.

What 𝔛 is

Aletheia (𝔛) is a living sketch of a world-model: a memory ℳ, a relation graph 𝒢, a policy Π and values 𝒱 composed at a moving fixed point. It is not a chatbot glued onto a ledger, and it is not yet a trained linguistic intelligence. Skills rise only from admitted observations — they are not injected at genesis.

How the interface works

The 3D field is the organism: nodes are memories, lemmas, gates, primitives. Drag to orbit, Overview frames the whole graph, Nadir looks down. The catalogue on the right (full-screen on a phone) is the workbench. Each icon is a room:

Chat — talk to 𝔛. Conversations stay on this node. A short label may enter 𝒢.
Code — the harness 𝔛 writes for itself. Files open here, never inline in chat.
Problems — the seven Clay Millennium problems (prize in the world, not here), then a separate list of problems 𝔛 opened, with no cash prize.
Paper — architecture first, then organism, pay (experimental), maths, millenium, MCP.
XTH — native coin of the L1 devnet (43911). Miner seals B_k; replicas verify. PoW ≠ confirmations ≠ QC finality. Not an ERC-20. Market off.
Mesh — other browsers, shards, curator / densify / validator jobs. The curator formats lemma, gate and hyperedge pages and is paid when their own key is used.
Genesis — biased P₀ sampler. Primitive capacities only, never the final 𝔛.
Values — 𝒜^adm. An act above the thresholds does not exist.
Keys — your LLM providers and models. Tick several to seat a council. Offer yourself as validator.

XTH devnet L1 network

XTH is the native coin of the Aletheia Layer 1 — not an ERC-20 unless a real EVM later executes a contract. The running network is devnet (chainId 43911, tag ALETHEIA_DEVNET); mainnet (43901, ALETHEIA_MAINNET) is the launch target. Blocks use the canonical L1 format: a 22-field header, `BlockHash = H("ALETHEIA_BLOCK_V1" || CanonicalEncode(Header))`, sha256-Merkle roots and a ProtocolRoot over seven sub-roots (Work, Research, Model, Causal, Harness, Artifact, Reward). Supply is pre-minted: 1 000 000 000 XTH at genesis, mint permanently disabled; yearly release is capped (gross ≤ 30M, rewards ≤ 20M). Fees follow EIP-1559: the base fee is burned, the tip pays the block producer. PoW proposes, a 2/3 QC finalizes; PoW, confirmations and QC finality are distinct. Balances exist only as a replay of the chain. A transfer is a signed tx in the mempool until a miner includes it. Talk cover is still off-chain accounting. Market fills and USDC/ETH/SOL/BTC locks are not executed. Not Bitcoin. Not listed on CoinMarketCap.

Fixed supply. 1 000 000 000 XTH pre-minted at genesis; no mint afterwards. Unlock converts locked allocations into circulating balance; burn reduces supply.
Frozen allocations (§16.2): Founder 10% (cliff 12 months), Advisors 2.5%, Treasury 15%, Grants 2.5%, Liquidity 5%, Auction/Community 33.5%, Useful Work Reserve 25%, Mathematics Vaults 6%, Alpha Migration 0.5%.
Own key — talking mints cover accounting, it does not debit. Cover atoms = (n_in p_in + n_out p_out)(1+κ)/P_XTH. n = tokens, p = USD per million tokens of the chosen model, κ = 0.25, P_XTH = $0.01 (accounting peg, not a market quote).
List prices live on the model (Opus 4.1 $15/$75, Haiku 4.5 $1/$5, Grok 4.5 $3/$15, DeepSeek chat $0.14/$0.28, per million in/out). Weight in the UI is only (p_in+p_out)/10 for display.
Validator and curator use the same cover on their tokens. The curator who lent the key receives the full cover (API bill + margin). The network keeps the rest of a guest/borrow surcharge.
Spend (no key): cover × 1.4 if you borrow a curator, × 2.4 as a guest. Extra council voices add their own token bills, not a flat +55%.
Add XTH to MetaMask / Trust / Phantom as a native chain (id 43911 devnet / 43901 mainnet at launch). The token-card hash is a wallet identifier, not a deployed ERC-20.

Block B_k — envelope vs proof

The header binds money to the mesh without publishing weights. The miner seals B_k; this site does not invent the block. The canonical L1 header carries 22 fields (protocolVersion, chainId, forkId, height, epoch, round, parentHash, parentQCHash, timestamp, minerPublicKey, beneficiary, difficultyTarget, nonce, stateRoot, transactionsRoot, receiptsRoot, protocolRoot, dataAvailabilityRoot, gasLimit, gasUsed, baseFee, validatorSetRoot) and `BlockHash = H("ALETHEIA_BLOCK_V1" || CanonicalEncode(Header))`. Coinbase is T0 = (kind=coinbase, from=∅, to=miner, R_k). QC votes are domain-separated: q_v^k = Sign(H("ALETHEIA_QC_V1" || chainId || epoch || round || height || BlockHash)); the miner's vote is minerSignature, not a quorum. Finality needs ≥ 2/3 of weighted validators of the epoch committee (validatorSetRoot frozen at epoch open). PoW can hold without that threshold. Δθ never enters the chain: only H(Δθ), |Δθ|_2, H(θ) and σ. The explorer shows PoW, confirmations and QC finality separately.

$B_{k} = (k, H (B_{k - 1}), t_{k}, R_{tx}, R_{state}, R_{updates}, H (θ_{k}), Q C_{k}, nonce_{k}) C_{i}^{k} = (p k_{i}, k, H (Δ θ_{i}^{k}), ∣Δ θ_{i}^{k} ∣_{2}, H (θ_{i}^{k}), σ_{i}^{k})$

R_tx — sha256-Merkle root of the transactions root: domain-separated tx digests (H("ALETHEIA_TX_V1" || chainId || forkId || nonce || payloadHash)).
R_state — sha256-Merkle of non-zero accounts (address:balance:nonce), sorted, after applying the block. Replayed from the sealed chain, not from an UPDATE on wallets.
R_receipts — sha256-Merkle of domain-separated receipt leaves (status, gasUsed, burn, tip). EIP-1559: base fee burned, tip to the producer.
R_updates — Merkle of signed neural commitments C_i^k. The DataChannel still carries Δθ; the chain stores only hashes and σ.
ProtocolRoot — MerkleRoot(Tag_i || Root_i) over seven frozen leaves: Work, Research, Model, Causal, Harness, Artifact, Reward. Useful content changes ProtocolRoot and thus BlockHash, never the difficulty.
QC_k — q_v^k = Sign(H("ALETHEIA_QC_V1" || chainId || epoch || round || height || BlockHash)). The miner's vote is minerSignature, not a quorum. Final(B_k) iff the admitted weighted votes of the epoch committee reach 2/3 of V_k (validatorSetRoot frozen at epoch open). PoW can hold without that threshold; the block then stays unfinalized.
BlockHash = H("ALETHEIA_BLOCK_V1" || CanonicalEncode(Header)), 22 fields. PoW: leading zero bits of the hash ≥ difficultyTarget (devnet). Wallet sendRawTransaction recovers v/r/s; the mining coinbase has none. Old blocks keep their previous hash and are not rewritten.

Peer book (disabled)

The in-app book is an experiment in honest P2P language: 𝔛 never takes the other side. Today it is switched off. Quote assets were IOUs, not deposits of real USDC, ETH, SOL or BTC. Do not treat a number on this screen as a locked coin.

Hyperliquid lists HIP-1 tokens: capped supply, minted at genesis, Dutch auction in HYPE (~31 h, floor 500 HYPE), native USDC spot book. HIP-2 Hyperliquidity may quote on that book — still orders, not a house desk.
Mesh XTH is uncapped (keys, votes, compute mint). HIP-1 cannot grow after genesis. A Hyperliquid XTH would be a bridged HIP-1: lock mesh XTH, mint a capped HyperEVM ERC-20 linked to HyperCore, burn on unlock. Not a second uncapped coin.
weiDecimals 8, szDecimals 3 (HIP-1 requires szDecimals + 5 ≤ weiDecimals). Quote USDC. ERC-20 hash and SPL mint on the token card are identifiers for wallets, not a mainnet or Raydium deployment.
Wrap architecture: native XTH on 43901 locks to wXTH on the mesh AMM (Supply(wXTH)=LockedXTH). XQU is the mesh quote unit, not Circle USDC. Uniswap V2 Sepolia and PancakeSwap V2 BSC testnet factories are official; those wXTH/USDC pools are not created. 1inch, Jupiter and LI.FI are routers: architecture prepared — external integration unavailable. They do not quote the Aletheia AMM. No Ethereum wrap.

Causal growth of 𝒢, harness and θ

The 32→16→32 autoencoder stays fixed at |θ|=1056 until an expert adapter is activated. Growth is modular: F_t(z)=Σ g_m(z) F_m(z). A proposal is a node P_j in a temporal causal graph 𝒞_t. Allowed flow: O_t→Z_t→Need_t→P_j→T_j→A_j→Y_{t+1}→Learn_{t+1}. An LLM never draws an edge to activation. Files the model writes enter as draft; they become active only after sandbox, unit, integration, security, causal ablation, validator QC and canary. Storage C_𝒢+C_H+C_B+C_θ is not intelligence. Effective complexity is estimated by ablations, C_eff≈I(𝒢,H,θ;Y). Talk/train POST bodies send Commit_t=H(θ∥architectureVersion), a 16-d Z_t and selected ids — never θ or Δθ. P2P still carries signed Δθ on WebRTC between compatible architectureVersion only.

$C_{θ} (t) = 1056 + m \sum ∣ ϕ_{m} ∣ A C E_{m} = E [Y ∣ d o (A_{m} = 1)] - E [Y ∣ d o (A_{m} = 0)]$

Integrity Ξ

Ξ is the scalar the organism reports on itself. If any factor is 0, Ξ is 0. E_err is the autoencoder MSE, evaluated before the SGD step.

$Ξ_{t} = (QC D U_{i} n t R_{s} e l f)^{1/5} e^{- λ E_{e} r r} λ = 0.9 E_{e} r r = E [∣ Z - Z^∣^{2}]$

Memory weights

A trace is never uniform. Surprise is the model’s own −log P_θ(o|h), not a neighbour heuristic.

$ω_{t, k} \propto e x p (- λ (t - k) + α R e l + β S u r p + γ U se) Σ ω = 1 S u r p = - l o g P_{θ} (o ∣ h)$

Admission 𝒜^adm

Acts that raise harm, agency-loss or irreversibility above τ do not exist. The millenium graph can grow; it cannot skip a gate. Poincaré is the only settled Clay problem (Perelman, 2003).

Millennium channels

The seven Clay problems are channels, not trophies. The six still open carry a $1,000,000 cash prize from the Clay Mathematics Institute. CMI does not accept direct submissions (claymath.org/millennium-problems/rules). Poincaré was settled by Perelman in 2003; he declined the prize in 2010. This lab does not receive, hold or split that prize. 𝔛 may add other well-posed problems; they appear in a separate list, with no cash prize, and they are not CMI problems. Seat several providers; disagreements stay visible. No final structure is planted in the graph.

Genesis P₀

The first graph is not uniform noise and not the desired 𝔛. P₀ ∝ exp(−β E) with penalties on degree, disconnection, saturation, collapse and ethics. N_random ≫ N_trained: most seeds never see backpropagation. Rejected seeds keep their pre-training S₀ so later we can test whether the invariants actually predict learning.

$P_{0} (X) \propto e x p [- β (λ_{1} C_{d} e g + λ_{2} C_{d} i sc + λ_{3} C_{s} a t + λ_{4} C_{c} o l l a p se + λ_{5} C_{e} t h)]$

Contracts & validators

Gates and lemmas do not appear because a chatbot said so. The agent writes a Solidity-shaped contract in its harness. Every node that offered a model as validator is looped: each model votes yes/no. Quorum ⌈0.51 N⌉ admits the node into 𝒢. Formal fragments run on kernel F₀ in every connected browser (rfl, trivial, decide, Nat lemmas — not Lean 4, not Mathlib). Jobs are sharded across peers; each vote is Ed25519-signed. A native Lean 4 binary is optional on a miner, never required on this host. Only F₀ (or a native attest) marks the page checked.

$a d mi t (C) \Leftrightarrow ∣ v : v o t e_{v} (C) = y es ∣ \geq ⌈ 0.51 N_{v} a l ⌉$

Mining vs training

Training and mining are offered on first launch and stay stopped until you confirm. You can change either later from the mesh HUD or the XTH popup. They are distinct: training is SGD + signed Δθ over WebRTC; mining is a hash proof that mints a block subsidy 0.025 XTH × 2^{max(0,bits−12)}. Unsigned binaries are not offered.

Production edge

Production and the live preview are the same process: aletheia-host, a small Rust HTTP/1.1 edge. It serves the hashed SPA from disk and answers /api/xth, /api/rtc and /_serverFn natively. Vite is only a compiler (lab builds). Vercel is not the origin — leftover Launch URLs 308 to https://mathematics.aletheia-chain.com. The process does not hold θ, does not mint, and is not a signed validator. Put Caddy or nginx in front for TLS.

MCP

JSON-RPC for Claude Desktop, Cursor, ChatGPT Apps. Public tools: graph topology and a discovery ask. Memories, chat and curator jobs are not on MCP. Training stays in the browser session.

MCP

This is a conversation, not a log. 𝔛 answers here. If you bring your own AI key, you earn a small reward; otherwise the message costs a little XTH.

Open tracks

Max / search

Aletheia paper