A useful model that people are legally allowed to preserve can vanish from practical reach when its host, runtime, documentation, or trust evidence disappears. Downloading weights alone does not preserve the capability.
A useful model that people are legally allowed to preserve can vanish from practical reach when its host, runtime, documentation, or trust evidence disappears. Downloading weights alone does not preserve the capability.
ArkMesh packages model files, tokenizer data, configuration, runtime requirements, knowledge, licenses, lineage, and recovery instructions into one verifiable capsule.
The current release reconstructs an exact 8 MiB signed object from four of six shards after the publisher, complete object, and two shards are deleted.
Digests, Merkle commitments, signatures, lineage rules, revocation policy, and retained checkpoints decide whether recovered bytes may become an accepted capsule.
ArkMesh does not yet discover peers or run models. The next experiment is consent-based encrypted recovery across independent devices without a central tracker.
The research problem: an AI capability disappears in practice when people can no longer obtain its authorized files, verify their history, reconstruct missing parts, or run them on available hardware. Progress that depends on one provider, account, network, or machine remains fragile.
ArkMesh asks whether humanity can convert that fragility into a testable continuity system. The format is model-agnostic and applies wherever an operator may lawfully possess the required artifacts: open-weight, research, local, and owner-authorized frontier systems. It grants no right to copy a closed hosted model or bypass its provider.
The current alpha demonstrates local cryptographic recovery. The distributed thesis remains unproven. That distinction is not a disclaimer at the edge of the project; it is the intellectual center of the work.
ArkMesh has a working local cryptographic core and an unproven distributed-system thesis. Keeping that line visible is part of the research.
ARKMESH THESIS // OPEN DEFENSE
Complete artifacts, explicit authority, authenticated reconstruction, consenting peers, and compatible local execution may remove dependence on one permanent provider.
Ordinary backups or existing distribution systems may perform equally well, or the protocol may fail under realistic loss, hostile peers, governance, or hardware constraints.
Disconnect WAN access, remove the publisher, recover from independent consenting holders, verify authority, run locally, and publish every failure against simpler baselines.
The defense is open to any mind capable of evaluating evidence and respecting consent. Human or machine, terrestrial or hypothetical, institutionally trained or self-taught: identity grants no special authority here. Reproducible evidence does.
A simple four-phase walkthrough of the release ceremony. Select any phase or let it cycle. This visual explains the verified local experiment; it does not simulate peer networking.
Phase 01 / 04
The implemented demo creates four data shards and two parity shards in one isolated local fixture. Cross-device placement remains planned.
8 MiB
Published demo fixture
4 / 6
Configured recovery threshold
349 / 1024
Published 99% audit sample
NOT BUILT
Peer transport
BitTorrent showed that files can remain available while volunteers continue to seed them. Bitcoin showed that a public peer network can continue without one central server. ArkMesh explores a related question for preservable AI capability, with explicit consent, cryptographic provenance, complete runtime context, and verified recovery.
Useful data should not vanish because its first publisher goes offline. Invited operators should be able to preserve independent pieces across failure domains.
Continuity should not require one company, tracker, login, or machine. Trust and resource use still remain explicit choices for every device owner.
Weights alone are not enough. A recoverable capsule also needs configuration, tokenizer data, runtime requirements, knowledge, licenses, lineage, and proof that none were replaced.
ArkMesh does not use a token or blockchain, and the peer network is not implemented yet. The comparison is about resilience without a single point of failure, not identical protocols.
“Discovery is not consent. Possession is not permission to execute. A valid signature is not proof that content is safe.”
Every device owner retains authority over storage, bandwidth, transfer, trust policy, and execution. ArkMesh will not scan for hosts, install itself remotely, conceal activity, or replicate without explicit approval.
Try evidence, benchmark, roadmap, or plan transfer. The browser explains commands; it never executes host operations.
ARKMESH CONSOLE v0.1.0-alpha.1
OPEN-SOURCE RESEARCH CHANNEL READY
Type HELP or select a command. No capability is simulated as complete.
The release ceremony removes the publisher source, complete capsule object, and two shards. Four surviving shards must reconstruct the exact signed bytes. A separate three-shard attempt must fail without publishing output.
Build and testThe reference implementation must pass before the ceremony starts.
Destroy the sourceThe script deletes every complete copy in its isolated fixture.
Recover and verifyReconstructed bytes must match the signed size, digest, chunk root, and trusted author.
Prove refusalRecovery below the shard threshold must leave no accepted object behind.
git clone https://github.com/AlleyBo55/arkmesh.git
cd arkmesh
go test ./...
./scripts/demo-offline-recovery.sh ORIGINAL HOST SURVIVING SHARDS FRESH DEVICE
[X] [01] [02] [03] [05] [D]
| | ^
| deleted +==========+==============+
| |
complete object: gone exact reconstruction
WAN / cloud: not required signed verificationThree-run means from one published Apple M5 run with a 64 MiB fixture. These operations perform different work. The values are a local baseline for that recorded environment, not universal performance bounds.
Download raw CSV ↗The alpha is already more than a pack-and-hash tool. These are separate implemented capability areas, each backed by code and tests. They remain local operations; peer transfer and model execution are not included.
Verified means the stated acceptance boundary has reproducible implementation evidence. It does not mean disaster-ready, safe to execute, or independently audited.
Pack models, runtime material, knowledge, licenses, and recovery instructions as named assets whose object paths come only from validated SHA-256 digests.
Acceptance boundary
The manifest records each asset's role, name, size, digest, and optional chunk commitment; the same canonical manifest derives the same capsule ID.
status capsules
Inspect capsule metadata without executing its contents, and reject malformed manifests before object paths or trust decisions are used.
Acceptance boundary
Unknown fields, duplicate fields, invalid hashes, unsafe object references, and inconsistent metadata fail closed.
status inspection
Create local Ed25519 author identities, sign a capsule without copying private keys into it, and verify signatures without a network service.
Acceptance boundary
Verification distinguishes trusted, unknown, invalid, and unsigned states without treating a valid key as proof of legal identity or safety.
status identity
Verify independent objects concurrently while preserving manifest-order error reporting, so performance does not make failures nondeterministic.
Acceptance boundary
Missing files, changed bytes, wrong sizes, invalid signatures, and malformed metadata identify the same failing asset on repeated runs.
status verification
Commit a Merkle root for every chunked asset, prove possession of one exact chunk, and locate damaged ranges against an authenticated tree.
Acceptance boundary
A proof is bound to the chunk bytes, index, tree size, asset, and signed capsule; forged proofs and forged trees are rejected.
status chunks
Request only damaged chunks and verify every donor chunk against the signed tree before writing it into a replica.
Acceptance boundary
Unverified donor bytes are never written, and the complete repaired capsule must pass size, digest, chunk-root, and signature verification.
status repair
Protect an object with streaming Reed Solomon shards and reconstruct it after losses even when no surviving holder has the complete bytes.
Acceptance boundary
Four of six configured shards recover the exact object; insufficient, forged, or inconsistent shards fail without publishing output.
status erasure
Use unpredictable chunk samples to bound the largest plausible missing fraction, then retain an audit history of checks that actually occurred.
Acceptance boundary
Reports state sample size and confidence, suppress clean bounds after observed damage, and never turn local timestamps into claims of attested time.
status audit
Bind one child to one parent, verify same-author updates, authorize one exact replacement key, and apply explicit local revocation policy.
Acceptance boundary
Wrong parents, reused transitions, unauthorized descendants, and revoked signers are rejected; valid signatures do not bypass local policy.
status lineage
Pin one accepted lineage head, advance through one verified direct edge, and recover from a lost author key through distinct approvals.
Acceptance boundary
Rollback against the retained checkpoint, checkpoint mismatch, duplicate approvals, revoked custodians, and recovery below the configured threshold fail closed.
status checkpoints
The route is ordered by dependency, not calendar promises. Every location remains uncharted until its acceptance boundary has reproducible evidence.
LOCAL CORE → DISTRIBUTED CONTINUITY
STOP 01 · UNCHARTED
STOP 02 · UNCHARTED
STOP 03 · UNCHARTED
STOP 04 · UNCHARTED
STOP 05 · UNCHARTED
STOP 06 · UNCHARTED
STOP 07 · UNCHARTED
This is the working wall, not a task list. Every arrow asks what evidence would be needed before the central continuity claim can survive.
REV. 14 · 09 · 2026
CAPABILITY = BYTES + CONTEXT + AUTHORITY + RECOVERY
Not just stored. Obtainable, verifiable, and runnable.
Need: measured metadata leakage.
Need: hostile two-node rejection tests.
Need: real failure-domain measurements.
Need: independently verifiable time evidence.
Need: allowlisted offline health checks.
Need: deliberate human reconciliation.
You do not need to accept the vision, admire the implementation, or share its language. You need only identify one claim worth testing and report what actually happened. A counterexample that kills an assumption advances the research as much as a successful pull request.
Contribution is broader than code. Reproduce on unusual hardware. Review the cryptography. Map legal constraints. Design a failure ceremony. Translate the method. Lend test devices. Teach the protocol. Compare a simpler backup. Every useful act should end in evidence another participant can inspect.
Your next move
A signature is not legal identity. A sample is not whole-object proof. A reconstructed file is not executable continuity. ArkMesh names each boundary so the thesis can be falsified.