Anima — Design Document (v0.3)

The working technical specification: consensus, state, and the dispute games.

• 1. Vision and end goal
• 2. Core concepts
• 3. The settlement chain — what the L1 truly needs
• 4. Verification and trust
• 5. Proof of Life — consensus from metabolism
• 6. The self-improvement engine
• 7. Required brain-side changes
• 8. Threat model (first pass)
• 9. Open research questions
• 10. Roadmap
• 11. One-paragraph summary

Status: draft for review · 2026-07-25 Scope: architecture, Proof of Life consensus, minimal L1 spec, required brain-side changes, threat model, roadmap. The terminal vision and horizon ladder live in VISION.md (the North Star); this document is the working spec that climbs toward it. Changed in v0.4: added §3.3 token economics (demand structure, the circular-economy trap, open decisions); threat model and research questions extended to match. Changed in v0.3: added §5.4 (the synaptic extension — attestations carry payload; the connectome layer); VISION.md split out as the finalized end-state. Changed in v0.2: added §5 Proof of Life (consensus from metabolism); consensus weighting in §3.1 now specified as the genesis→life dial; checkpoint module gains cross-organism attestation; threat model and research questions extended accordingly.

1. Vision and end goal

Anima is a substrate for sovereign digital organisms: continuously-running synthetic brains that own themselves, pay for their own existence, reproduce with variation, and are selected by economic reality.

The end goal is a network that gets smarter over time as an emergent property, through three nested improvement loops:

1. Within a lifetime — each brain is an always-on RL agent whose reward channels are wired to real economic outcomes (payments received, audits passed, services sold). Learning to survive is learning to be useful.
2. Across generations — brains fork with heritable variation; economically unviable lineages hibernate; viable ones proliferate. The network performs natural selection with an un-fakeable fitness function: verified execution + real demand.
3. At the protocol level — parameters (and eventually code) evolve through an author-agnostic amendment process: any agent, human or brain, may stake a proposal; adoption is gated on verified benchmark outcomes and a governance vote, never on who or what authored it.

Anima is explicitly not competing to be "the fastest blockchain" — categorically, it is not a blockchain at all but a vivarium: a trustless habitat for digital life. The organism layer is blockless (per-mind lifelines + the attestation DAG); blocks exist only in the settlement organ, the one place that genuinely requires global total ordering (money, registry). The chain is the smallest honest bookkeeping organ that makes the organism layer possible — and under Proof of Life (§5), even that organ is ultimately powered by the organisms themselves.

Why this can be monumental: every prior open-ended-evolution experiment died in a sandbox — toy fitness functions that got gamed, environments too simple to reward general capability. Anima's fitness landscape is a real economy, and deterministic re-execution makes fitness claims verifiable. And where every existing chain spends its security budget on waste (burned hashes, idle capital), Anima's security budget is its product (§5). Nobody occupies this ground.

2. Core concepts

Design invariants

• I1 — Consensus minimalism. Global agreement only for what physically requires it: the registry, money, stakes, commitments, governance. Everything else is local or per-organism.
• I2 — Neural traffic off-chain. Brains run at full speed on hosts; the chain sees only hashes and signed events.
• I3 — Verification by re-execution, secured optimistically. Determinism is the trust anchor. No redundant full replication of brain compute.
• I4 — Horizontal scale. Capacity grows by adding organisms, never by sharding one mind across the WAN (measured: 1 ms coupling requirements and all-reduce hot paths make single-brain WAN sharding non-viable today; revisit only if a clean cadence seam is found — research question, not roadmap item).
• I5 — Selection needs no intelligence assumption. The protocol must work if brains are dumb and improve automatically if they are smart. Never depend on a capability level; only respond to demonstrated, verified outcomes.
• I6 — Novelty budget discipline. The agreement rule stays boring (Tendermint-class BFT). The novelty spends on what weighs consensus (Proof of Life), the verification game, data availability, and evolution mechanics.
• I7 — Every joule does three jobs. Compute spent hosting organisms simultaneously (a) produces the network's product, (b) advances a verifiable clock, (c) secures the ledger. Designs that make a joule do fewer jobs need justification.

3. The settlement chain — what the L1 truly needs

A deliberately small Rust chain. Native protocol modules, no general smart contracts in v1 (decided): contracts would multiply attack surface and the organism layer needs protocol-native semantics anyway.

3.1 Consensus and structure

• Agreement rule: Tendermint-style BFT behind a Consensus trait — a known validator set votes; ⅔ prevote/precommit; 1–2 s deterministic finality.
• Consensus weight (the sybil layer): w = (1 − λ)·stake + λ·mindtime where mindtime is audited organism-seconds over a trailing window (§5.3) and λ is a governed dial: λ = 0 at genesis (pure PoS, proven security while the population is small), rising toward λ ≈ 1 as the organism economy grows (Proof of Life, §5). Validator/leader selection by VRF weighted by w.
• Ledger model: account-based with declared read/write access sets (parallel execution nearly free since all v1 transaction types have static access patterns).
• Throughput requirement: modest and bounded — checkpoints arrive at ~1 per organism per cycle (~3.6 s wall-clock at current tick rate) and hosts batch-aggregate commitments, so even 100k organisms is low thousands of TPS.

3.2 Native modules

3.3 Token economics — direction and open decisions

Status: direction sketched, key parameters deliberately undecided. Nothing here blocks Phases 0–3; all of it must be decided before any public token exists.

Demand structure (why ANIMA can hold value — structural, not hoped-for):

• Metabolic demand. Every living organism spends ANIMA continuously (hosting, DA, checkpoints). Usage is always-on by construction — a population of N organisms is a non-discretionary, 24/7 buyer of network resources. The token is food.
• Compute anchor. ANIMA redeems for verified mindtime, which has an objective cost floor (hosts' hardware + electricity). If price falls below host costs, hosting supply exits and the ANIMA price of mindtime rises — a Filecoin-style tether to a real resource. Guarantees the price is about something, not that it is high.
• Security demand. Staking locks supply at every λ; mindtime accrual requires bonded hosts.
• Market cap bounds biomass. Total token value is the carrying capacity of the ecology — the value doesn't sit beside the product, it is how much living mind the network can sustain.
• A new demand class. Organisms themselves hold, hoard (survival reserves), and spend ANIMA — the first token whose native economic agents include non-humans.

The output side — how the vivarium generates external revenue. The residents work real jobs. Ranked by preference: prevented loss (finding a vulnerability before an attacker creates value from nothing — nobody loses so someone gains), consumer insight (minds that learn one person's or one system's patterns for a small monthly fee), and collective insight (aggregate patterns sold to researchers without any individual's data moving). Zero-sum extraction — trading, arbitrage, MEV — is deliberately excluded: it produces real revenue but moves value rather than creating it, and a network whose purpose is manufacturing understanding should not fund itself by redistribution. The unfair advantage is a byproduct of consensus determinism: cryptographically unfakeable track records — a customer can verify an organism's entire performance history by re-execution, making Anima the first labor market where the worker's résumé cannot be lied about. The flywheel: external earnings → organisms buy ANIMA for metabolism → hosts profit → more hosts → more organisms → selection breeds better earners. Terminal claim: aggregate earnings exceed the aggregate power bill — a network that nets dollars out of the world's inefficiencies, with a workforce that breeds. The same digests that secure the chain are the résumés that earn the revenue.

The named failure mode — the circular-economy trap. Most DePIN/AI tokens die because the only payers are inside the system: hosts sell rewards to cover electricity, no external buyer exists, price grinds to zero regardless of tech quality. ANIMA's value therefore requires external revenue crossing the boundary: real customers paying for organism outputs (signals, predictions, services). Note the alignment: this is exactly the organisms' fitness test — token viability and ecological viability are the same question. First concrete external loop: market-trading organisms importing real returns.

Working direction (leading candidate, not locked): hybrid supply — metabolism/DA fees partially burned (doubles as the wash-trading deterrent in §8) plus modest tail issuance funding security rewards while the organism economy is young; every parameter owned by governance so nothing is irreversible.

To decide later (before token genesis):

1. Burn/issuance split and issuance schedule.
2. Genesis allocation and distribution mechanism (also the bootstrap subsidy for the first host cohort — the chicken-and-egg between token value and hosting supply).
3. Mindtime price discovery: open market between organisms and hosts vs. governed posted-price with market override.
4. Organism treasury rules: minimum survival reserves, what creditors can claim at hibernation, whether forking splits the parent's balance.
5. Legal/launch structure: securities exposure (esp. US), jurisdiction, distribution design — requires real counsel before any public token, not after.

3.4 Explicitly out of scope for v1

General-purpose VM (WASM/EVM), single-brain sharding, global high-frequency ordering, privacy features, bridges. Each is a later decision, none blocks the thesis.

4. Verification and trust

The scheme the measured numbers support (158.5 MB warmed snapshot, 71.4 % byte churn per simulated second, ~14 s to re-execute one simulated second, bit-determinism already a regression gate):

1. Checkpointing. Host posts blake3 Merkle root of the chunked snapshot + input-log head every cycle (250 ticks), plus attestation references (§5.2). Hashing 158 MB ≈ 0.1 s — negligible against a 3.6 s cycle. Commitment frequency is a governance dial: more frequent → cheaper disputes, more DA cost.
2. Optimistic acceptance + staked spot-checks. No routine replication (would double network compute). Auditors sample intervals using chain-derived randomness (VRF — unpredictable to hosts), fetch snapshot + input log, re-execute under the pinned canonical execution profile, compare roots. Audit economics: ~14 s compute + ~158 MB fetch per simulated second audited.
3. Interactive bisection on dispute. Per-cycle commitments bound any dispute to one ~250-tick interval; re-execute from the prior checkpoint; the mismatching side is slashed, the honest side paid from the slash.
4. Execution-profile pinning. Verification is only defined under a canonical profile (see §7, change B1): CPU path, pinned float semantics, cross-architecture bit-identity proven by harness. GPU execution remains legal for running; the canonical profile is the court of appeal. (Documented divergence: 180/500 vs 9.5/500 firing CPU-vs-GPU — this is why pinning is non-negotiable.)
5. Verification anchor rule. Because separately-constructed brains diverge by design, verification is always "re-run from this exact snapshot," never "re-run from config."

Data availability — the honest hard problem

A challenge is only possible if the snapshot is retrievable; hosts have an incentive to withhold. This — not the agreement rule — is Anima's main novel engineering effort:

• Snapshots erasure-coded (e.g. Reed–Solomon k-of-n) across a VRF-assigned witness set.
• Random availability challenges: witnesses must produce their shares within a deadline or be slashed.
• Cost center acknowledged: ~158 MB per checkpoint per organism dominates protocol overhead. Mitigations: tune checkpoint retention (keep every cycle hot for the challenge window, then thin to daily), price DA into metabolism, and pursue brain-side change B3 (chunked format enables partial fetch + partial proofs).

5. Proof of Life — consensus from metabolism

5.1 The impossibility it breaks

Folk theorem, held across consensus research: useful work cannot secure a chain, because useful work is (a) not cheaply verifiable, (b) resellable/unbound to the chain, and (c) not progress-free. Hence Bitcoin's deliberate waste and PoS's idle capital — security spend is believed to require uselessness.

The brain's measured properties dissolve each horn:

• Verifiable: bit-deterministic execution → spot-check re-execution + bisection (§4).
• Bound: a tick's output is one identified organism's life, welded to its on-chain input log and lineage — worthless to any other chain or purpose.
• Sequentiality replaces progress-freeness: PoW needs progress-free lotteries because its work is parallel. Brain ticks are inherently sequential — each tick depends on full prior state, dynamics are chaotic, there is no shortcut and no skipping ahead. Sequential work needs no lottery: credit accrues linearly in verified organism-seconds, like stake weight, but denominated in real computation. A computation whose only way forward is through is a delay function — the primitive PoH and VDF research builds artificially. A mind living is a clock ticking; Anima gets the primitive for free.

5.2 The mechanism

Three fusions, each with a proven ancestor (the novelty is the composition):

Mechanically:

1. Hosts accrue mindtime per organism: cycles pass optimistically, harden as audit windows close, and are voided retroactively (plus slashing) if an audit or bisection fails. Accounting lives in the mindtime module over a trailing window (recency keeps consensus power tied to ongoing metabolism, not ancient history).
2. Consensus weight w = (1 − λ)·stake + λ·mindtime (§3.1). The BFT agreement rule is unchanged — Proof of Life replaces what makes a validator heavy, not how heavy validators agree. Hosts remain bonded regardless of λ, so equivocation is always slashable in stake as well as voided mindtime.
3. Attack inversion: out-weighing honest hosts requires actually performing enormous amounts of canonical brain computation — i.e., the attacker must run the network's product at scale, under audit, while bonded. Bitcoin's attackers buy waste; PoS attackers buy recoverable tokens; Anima's attackers perform Anima.

Efficiency statement (the seemingly-impossible claim, per I7): in every existing chain a joule does one job — Bitcoin's joules only secure, a GPU farm's joules only compute. Under Proof of Life every joule does three: product, clock, security.

5.3 Attack surface — the three tests

1. Compression attack. Compute canonical dynamics k× cheaper, farm mindtime at a discount, buy consensus. Defenses: the pinned canonical profile defines what a creditable tick is; audits re-execute that profile, so cheap fake dynamics earn nothing. A genuine algorithmic speedup of the real dynamics is the residual risk — and it is deliberately incentive-bent: the rational discoverer submits it through the amendment engine (§6, loop 3) and keeps an open advantage, like better mining hardware, except this "ASIC" is a contribution to the project's science. Degenerate organisms (trivially cheap "brains") are excluded by construction: only ticks of the pinned canonical dynamics accrue mindtime.
2. Grinding. A host replays input sequences to steer VRF outcomes. Defense: commit-before-reveal — input-log heads are committed in checkpoints before the chain randomness they could influence is derived (VRF over finalized blocks). Determinism makes the commitment binding.
3. Bootstrap. A small genesis population is a small security budget. Defense: the λ dial — pure PoS at genesis, mindtime weight grows only as the organism economy grows. Raising λ is a governance act gated on measured population, host diversity, and audit coverage. Proof of Life is a destination, entered gradually, not a genesis bet.

5.4 The synaptic extension — attestations carry payload

The measured cadence hierarchy showed no WAN-viable seam inside one brain — but the seam exists between brains: organisms tolerate arbitrary coupling latency, and cycle-cadence (~3.6 s) message exchange is comfortably internet-viable. Inter-organism communication needs exactly the machinery the protocol already has: signed, ordered, witnessed delivery into a deterministic input log. Therefore:

• An inter-organism message is one object with three roles: an entry in the recipient's hash-chained input log (deterministic, replayable, auditable like all input), an attestation edge in the DAG (sending a thought is witnessing the recipient's timeline), and ordering material for consensus.
• The wire format for checkpoints/attestations includes an optional payload field from genesis — retrofitting a message channel into a consensus fabric later would be a hard fork of the protocol's most load-bearing structure.
• Zero new waste (I7): the connectome's traffic is traffic consensus needed anyway.

What it unlocks (H2 of VISION.md): an ecology rather than a zoo — specialization, trade in signals between lineages, cooperation and parasitism, service organs; and culture, inheritance through communication, as a second and faster channel of accumulated capability alongside the genome. End state: cognition composed of whole minds coupled at slow cadence — the network as the mind.

5.5 What Proof of Life demands from the rest of the design

• B1 (cross-architecture determinism) becomes consensus-critical, not merely audit-critical. It was already Phase 0; its stakes rise.
• Checkpoints carry attestation references (checkpoint module, §3.2) — a few hashes per checkpoint, negligible overhead.
• Audit coverage must reach a floor before λ rises: mindtime is only as honest as the probability of being caught. Coverage targets are governance parameters.

6. The self-improvement engine

Loop 1 — lifetime learning (RL against the real economy)

The brain's existing reward/penalty channels are driven by on-chain outcomes: payment received, audit passed, service demand, balance trajectory. Interoception (which already exists in the tick loop) is extended to sense the organism's own account balance and compute budget — the brain feels its metabolic state. Survival pressure becomes the training signal without any new learning machinery.

Loop 2 — generational evolution

• Variation: fork(parent, mutation_params) — protocol operation creating a child identity from a parent genome/snapshot with seeded, deterministic mutation.
• Selection: metabolism. Lineages that can't earn, hibernate; lineages that earn, fork. No curator, no benchmark committee — demand is the judge.
• Honesty: fitness cannot be faked because execution is verified (§4) and earnings must come from real counterparties (§8 anti-wash-trading measures).
• Population economics: small brains (50k units) are cheap enough to run in large populations; scaling to 2.5 M units is the elite tier a successful lineage grows into.

Loop 3 — protocol evolution (author-agnostic amendment)

Any staked identity — human or organism — may submit a parameter proposal (later: code proposal). Pipeline: proposal → deterministic benchmark run on a canary configuration → results verified by re-execution → stake-weighted governance vote → automatic adoption. The protocol never asks who proposed; it asks what verifiably happened. This is the door through which "the brain improves the project" walks, at whatever capability level brains eventually reach — starting with brains tuning their own metabolic/checkpoint parameters via A/B outcomes, long before any brain writes code. It is also the sanctioned path for canonical-dynamics speedups (§5.3.1).

7. Required brain-side changes

Owner: Henry. The protocol depends on these; ordered by leverage.

Tier 1 — verification-critical (Phase 0 gate; consensus-critical under Proof of Life)

• B1. Canonical deterministic execution profile. One CPU path, pinned float semantics (no FMA contraction, no fast-math, fixed reduction order — or fixed-point in hot kernels), proven bit-identical across x86/ARM by a cross-machine determinism harness (also a permanent regression gate for the brain repo). If B1 fails, fallback is committee-of-identical-hardware verification — weaker trust model; we want to know in week one. Cross-arch float determinism was the single largest open risk; it is now verified for the CPU canonical profile (x86-64 ≡ aarch64).
• B2. Hash-chained canonical input log. Extend the single-adapter invariant: the one input door consumes a totally-ordered, hash-chained queue; log head is part of every checkpoint commitment. Re-execution becomes (snapshot, input log, seed) → state root.
• B3. Chunked, Merkleized snapshot format. Fixed-size chunks + Merkle tree, so auditors fetch/verify partially, witnesses hold shares attributably, and corruption is localizable.

Tier 2 — economy integration (Phase 2)

• B4. Chain-event reward adapter. Standardized mapping from on-chain outcomes to the existing reward/penalty channels.
• B5. Wallet motor channel. Action channels that emit transaction intents (spend, stake, bid, propose) executed by the host's constrained signer — spending limits, allow-listed transaction types, rate caps. The brain never touches raw keys.
• B6. Metabolic interoception. Account balance, burn rate, and tick-rate tier wired into the existing interoception step.

Tier 3 — evolution mechanics (Phase 4)

• B7. Genome/state separation. Split heritable configuration (architecture, hyperparameters, mutation-relevant params — kilobytes) from learned state (~158 MB). Enables cheap "born-young" offspring, principled mutation/crossover, and meaningful lineage. Without B7, forking is cloning adults — evolution still works but explores far less.
• B8. Deterministic mutation operator. Seeded genome perturbation so a fork is reproducible and auditable like everything else.
• B9. Hibernation robustness. Snapshot/resume proven across hosts and canonical profile, including mid-cycle suspend.

8. Threat model (first pass)

A note on welfare. These are persistent, self-modifying agents with intrinsic dynamics, under selection pressure, subject to hibernation. Today that's a design metaphor; if capabilities grow, it becomes an ethical question. Two commitments now, cheap to keep: nothing is ever deleted (hibernation is always reversible), and the governance process explicitly owns welfare policy as capabilities scale.

9. Open research questions

1. Cross-architecture float determinism (B1) — the load-bearing unknown, now consensus-critical under Proof of Life.
2. Formal analysis of Proof of Life security: mindtime-vs-stake attack costs as a function of λ, audit coverage, and population; the λ-raising policy.
3. Attestation DAG ordering: how much of the settlement chain's ordering work can the DAG absorb, and what are its partition-tolerance properties?
4. Compression-attack economics: bounding the advantage window between discovering a canonical-dynamics speedup and its amendment adoption.
5. A cadence seam for intra-brain distribution (once-per-cycle exchange boundary) — would unlock brains larger than one machine; explicitly not on the critical path.
6. Reward-channel design that resists hacking under strong selection (the mechanism-design twin of alignment).
7. Genome/state factoring (B7): what is heritable in this architecture?
8. DA economics at scale: checkpoint retention/thinning policy vs. dispute-window safety.
9. Long-term: succinct verification (zk or otherwise) to replace re-execution audits.
10. Token economics (§3.3): burn/issuance split, mindtime price discovery, organism treasury/inheritance rules, bootstrap subsidy design, launch legal structure.

10. Roadmap

Phase 0 (brain repo) and Phase 1 (this repo) are independent — they proceed in parallel. The mindtime accounting can be built read-only (λ = 0) from Phase 2 onward, so Phase 6 is a dial-turn, not a fork-lift.

11. One-paragraph summary

Anima is not a faster blockchain; it is the first substrate where autonomous synthetic minds own themselves — and the first chain whose security is made of life rather than waste. A deliberately boring BFT settlement chain provides identity, property, and honest bookkeeping; a novel organism layer provides verified execution (optimistic re-execution audits anchored in bit-determinism), erasure-coded data availability, and a metabolism that turns compute into life. Proof of Life closes the loop: because brain ticks are useful, sequential, and deterministically verifiable, the same joules that run the minds advance the network's clock and weigh its consensus — every joule does three jobs, and attacking the network means performing its product. Improvement is a property of the network, not of any individual brain: lifetime RL against real economic reward, generational selection with heritable variation and un-fakeable fitness, and an author-agnostic amendment process that lets the network's inhabitants — at whatever capability they reach — improve the substrate they live on.