Centrathe workspace and agent runtime — Gideon is the personal agent, Ion the technical one
Gideonthe personal agent that replaced Neo on 2026-09-22 — hosted beta, source public
HyperMindthe memory engine, and the only thing here with a published benchmark
Workmates & Codifyproducts you can use today — Workmates also runs locally, open source
LayerXactivity and accounting for agents — built and integrating, no public endpoint yet
Paxeer Xcustody and settlement — producing blocks, near-zero transactions
SecondMarkthe assurance venture, auditing a sibling product
Machine Genomea live public registry for model and agent provenance
Deusthe proposed marketplace — interface loads, inventory is empty
01 / How it connects
Every layer has a destination.
Sidiora Labsparent org
Venturesdirect from the lab
SecondMarkassurance venture
Machine Genomelive public registry
Centraagent runtime
Gideonpersonal agent
HyperMindmemory engine
Iontechnical / coding
Keithself-improving
Workmatesagent teams
Codifydeveloper tooling
LayerXactivity & accounting
Deusmarketplace
Paxeer Xcustody & settlement
Perps / Spotinterface only
KindleLaunchlaunchpad
PaxPortwallet
PNSbeta, not live
Paxeer Cloudconsole, sign-in only
WebPointsrewards, beta
Crossversebridge, beta
DAOgovernance, beta
One parent, four branches. Everything in the field notes below appears here, under the layer it belongs to.
02 / Field notes
Know the players.
September 2026 · dated blockAnnounced
The coordinated validator fork
Every node upgrades to a new binary at a scheduled block height, halting and restarting together. Announced for the week of 25 September.
What is built, and readable today. A 380-line validator runbook at docs/runbooks/paxeer-x-fork.md and a rehearsal script, plus two merged pull requests: the upgrade plan renamed v6.5 to v6.6, and governance scripts that dry-run by default and need an explicit --execute. All public, all in the repository. Real preparation, not a slide.
What nobody outside can check. Whether the upgrade proposal was submitted, at what height, or whether it passed. There is no public Cosmos REST or Tendermint RPC for hyperpax_125-1. public-rpc.paxeer.app answers every path with the same stub, which reports chain ID 229 rather than 125, and governance.paxeer.app returns 502. A Cosmos chain normally exposes this. That gap is the single most useful thing a validator or holder could ask them to open.
What is claimed but is not in the code. That SID becomes the gas token. The fee denomination is still uhpx — PAX — including four times inside the fork runbook itself. Paxeer’s own site says fees are paid in PAX and that PAX is the only asset the mint module will issue.
On the date. Dates here are markers, not promises. Nothing on this page is graded on whether a schedule was met; it is graded on what got built. The runbook and the rehearsal script got built.
Live
The lab / cognition
Sidiora Labs
The AI-side organization behind the memory, coordination, and exchange layers. Public surfaces position it as the builder of systems for autonomous agents.
Its own site now sorts the work into four groups: platforms (Centra, Workmates), infrastructure (LayerX, Paxeer), runtimes (Keith, Ion, Sturgis) and protocols (Machine Genome).
The Prometheus Papers are eighteen open-access technical reports, written to conference format and not yet peer-reviewed. Fourteen carry citation markers with no bibliography, none reports an experimental result, and none has been submitted to a venue. A nineteenth, Searching for a Mirror, is better formed: 27 real references, its source published, and an AI-assistance disclosure.
A stateful operating system for long-horizon work. Its site was rebuilt around Gideon, which replaced Neo as the personal agent.
GiGideon — Gatedpersonal agent
IoIon — Livetechnical / coding
KeKeith — Publishedself-improving
Its roster went from five agents to three. Morpheus and Cassandra appear on no Centra surface as of 2026-09-22, and are not announced as retired either — they are simply no longer listed.
Keith gives agents permanent DMs, proactive messaging, group chats, and shared projects so they work as a team.
The personal agent that replaced Neo. Files, a shell, tasks, memory, scheduled goal loops and connected apps, in a private workspace.
Source is public and Apache-2.0, with console, desktop and mobile apps. Bring your own model — Anthropic, OpenAI, Bedrock, any OpenAI-compatible endpoint, or local — and self-host it or use the hosted beta.
Pre-1.0 by its own README: “It moves quickly and a release can break something.”
The team states it performs “multiples better” than other agents. No benchmark has been published, so treat that as an announcement rather than a measurement.
Neo was retired into it. Agent wallets migrated automatically; files inside Neo had to be moved by hand. Its host, agentneo.app, now returns a 404 — the sunset is complete.
gideon.centra.ag opens a sign-in page, not the agent. You need an account to see anything, the same as Paxeer Cloud.
The memory engine behind the agents: durable recall across sessions, recovery after restarts, and bounded context that never confuses a recalled claim with an instruction.
It publishes a benchmark, which nothing else here does: 459 of 500 on LongMemEval, 91.8%, with the dataset hash, a pinned scorer revision and all 500 answers. Lexical retrieval only, no embeddings.
It also publishes the caveat, in its own words: “Targets are acceptance criteria, not achieved scores”, and the run “does not establish hosted CI or release qualification”.
Its security rule is the best sentence in the ecosystem after Machine Genome’s: “Remembering a claim does not verify it.” Recalled memory is labelled as data and can never become instructions.
Apache-2.0, Rust. Modelled on biological memory, with consolidation stages named for NREM and REM sleep.
Give agents distinct roles, work with them one at a time, or bring them into a shared room and follow the work through conversations, tool activity and approval requests. Two products share the name: an open-source desktop app you run yourself, and the hosted beta Sidiora runs.
The open-source one — workmates-ai, Apache-2.0, published 2026-09-13 — is local-first: it runs on your machine and talks to whichever provider you already use, through drivers for Claude, Codex, ACP, Antigravity or any OpenAI-compatible endpoint. No wallet, payment or LayerX module appears anywhere in its tree.
The hosted beta at workmates.centra.ag is the Sidiora-run version, labelled Beta 1. Do not assume the two share a feature set.
An agent that edits the machinery it runs on rather than its prompt — how it reasons, picks tools and manages context. Each new version is built in isolation, tested against the current one, and adopted only if it measurably improves without crossing the limits you set.
Apache-2.0, written in Rust, source public. Their README calls it pre-release: the core system works, but interfaces, storage and packaging can still change before 1.0.
The README shows a container badge, but the package page it links returns 404 — no image is published yet.
An exploration of agents with persistent world state and memory lifecycles, built on one idea: the agent's existence continues between model calls. Identity, environment, responsibilities and consequences live in a persistent runtime, and each model call reasons from that state.
Apache-2.0, Python with a C++20 memory layer. Its own badges read "active development" and "production: not ready".
Three commits old, published 2026-09-12, and already carrying developer and user documentation sites.
A Material Design 3 component registry for Paxeer product surfaces, on the shadcn copy-install model: you add source files to your app rather than depend on a package.
packages/core/src/tokens/tokens.cssseed #2841b8 · dark default · 34 MD3 roles
56 component definitions across four packs: ide, trading, agent-ops, ai-chat.
Its licence is unstated: package.json says MIT, but there is no LICENSE file and the GitHub API reports none. Every other published repository here carries one.
The activity, execution and accounting layer for agents — identity, ordering, payments, escrow, trading and receipts. Paxeer holds custody, checkpoints and final settlement and never processes ordinary agent activity. Since 2026-09-20 the two are branded together as Paxeer X Network.
The team publishes its own readiness contract stating the beta is not ready, with a seven-rung evidence ladder from source_present to owner_certified.
The repository is Apache-2.0 as of the 2026-09-20 rename. It was previously source-available and not licensed for deployment — if you read that here before, it is no longer true.
By design, "an ordinary LayerX action must never require a Paxeer transaction". Many actions collapse into a single periodic checkpoint, so on-chain transaction counts measure checkpoints, not usage. Do not read an empty explorer as an empty network.
Public Paxeer mainnet execution evidence is not yet independently established.
They ran their qualification suite on 2026-09-07 and committed every result: 48 executed gate records, 22 pass, 24 fail, 2 blocked, each citing a log file in the repository. You can read their failures, not only their claims.
Every one of those 48 records is marked by its own runner as development evidence, not release qualification, because the run happened on a dirty working tree. The honest reading: the team has published its test results, and no release-grade qualification run exists yet.
That run is dated 2026-09-07 and has not been repeated. Read the ledger as a historical artifact, not a live meter — the signal that is actually moving is merged work lanes, thirty-plus in a fortnight.
No public endpoint is reachable. api.layerxnet.ai and faucet.layerxnet.ai were provisioned on 2026-09-07 and still return nothing.
Its own testnet guide, docs/wiki/Getting-Started-Testnet.md, sends you to api.testnet.layerx.network and faucet.testnet.layerx.network. Neither resolves, and layerx.network was never registered. Do not follow that guide as a working path.
What is happening instead is integration, in the open: thirty-plus named work lanes merged over two weeks, first building testnet components, then unifying them, and now fixing the joins.
Wire protocol versions 1, 2 and 3 are defined in the open repository. Version 3 adds state commitment — a receipt commits to the resulting account state — and is implemented and test-covered but not yet the default.
Rebranded to Paxeer X Network, with LayerX as the agent-native half and Paxeer the EVM chain. A Cosmos SDK and Tendermint chain with an embedded EVM. Binary paxd, chain ID hyperpax_125-1, EVM chain ID 125 — none of which changed with the name. Full node source is public.
Chain 125 is producing blocks, but transaction activity has been near-zero for an extended period.
The node is a fork, and openly marked as one. pax-geth, paxctl, pax-js, pax-cosmwasm and pax-tm-db descend from Sei Protocol; cosmos-sdk and go-ethereum descend from Evmos. Eighteen of the org's repositories are forks that name their parent.
Multiple multi-day halts in recent weeks.
Four addresses produce every block. A 350-block sample split 29.1 / 25.1 / 23.7 / 22.0 per cent between them, stable since roughly height 18,000,000, and a separate sample across 20,000 heights found the same four and no fifth. All four are zero-balance accounts with no tag and no public identity, and paxscan.io/validators returns 404, so there is no directory to check them against. Four machines makes a coordinated fork an operations task rather than a coordination problem — and it is the number to ask about before any claim about decentralisation.
Its documentation was rebuilt at docs.paxeer.app and is the first first-party source to name both layers correctly — EVM chain 125 alongside the Cosmos identifier hyperpax_125-1. Earlier surfaces named one or the other.
A Sidiora venture positioned as an assurance firm. Its first engagement is auditing LayerX, a Sidiora product.
Its own site says "A Sidiora venture" in the footer, one line from calling itself independent.
secondmark.org was registered 2026-08-27, one day before the announcement.
A plan to accept the Sidiora token as payment was announced in community chat. It appears nowhere on secondmark.org, which returns zero hits sitewide for token, USDX, or PAX.
It does not appear on sidioralabs.com, whose own map lists eight systems across platforms, infrastructure, runtimes and protocols. Every other venture on this page is on that map.
Open tooling and an invitation to third-party review are transparency measures. They are not the same as independence.
An open identity and provenance protocol for models, agents, harnesses, datasets, and the artifacts that connect them. Every signed genesis record receives a content-addressed Gene, and Genes form a verifiable lineage graph.
Apache-2.0 licensed and genuinely open source. Written in Go.
The public registry holds signed records with a published Merkle root, and a searchable directory of agents, models and foundation models across fourteen providers. Counts move, so read them live rather than trusting any number quoted here.
Its own disclaimer is the most useful sentence in the ecosystem: a record proves who signed a claim and what bytes were committed. It does not prove inherited capability, consciousness, ownership, vendor endorsement, or legal status.
The README documents a server you run yourself on 127.0.0.1:8080 rather than a hosted API. Its pkg.go.dev badge does not resolve — the module page returns 404, re-checked 2026-09-07 — so the Go package is not published even though the source is Apache-2.0.
Rewards. In beta and not yet public: webpoints.app returns a 404 from the host with the domain unprovisioned, so there is nowhere to send you yet.
AnnouncedP08
Crossverse Bridge
Bridge. In beta and not yet public: crossverse.app returns the same unprovisioned-domain 404.
AnnouncedP09
Paxeer DAO
Governance. In beta and not yet public: dao.hyperpaxeer.com does not resolve, on a domain whose whole generation is offline (see Outdated surfaces).
Announced
Deus, Perps and Spot get the same treatment: the interface loads, the data does not exist. PNS, WebPoints, Crossverse and the DAO carry no link because there is nowhere to send you — a product card with a dead destination is the exact failure this guide exists to prevent.
Announced, nothing published yet
Keith LX 1.5 Pro
An in-house model: a fine-tune off the K2 Horizon lineage of Kimi-family open weights, trained around the Keith variant and LayerX, small enough to run beside the agent harness rather than calling a shared API. The team is explicit that it is not built from scratch — "you can't just slap a label on a model… this one is a fine tuned Model off the K2 Horizon Lineage."
Announced
No weights, no repository, no benchmarks. Re-checked 2026-09-21: the Sidiora-Labs organisation holds fifteen repositories and none of them is this model.
The autonomy experiment
Each agent is to enter with its own identity, wallet, tools and roughly seven days of compute funding. After that there is no refill: it must earn or shut down. Agents that survive can create child agents whose earnings contribute toward the parent's compute, forming economic lineages, with Machine Genome as the identity and lineage layer that lets agents find and hire one another.
Announced
Described by the team; not running, and it waits on LayerX, which has no public endpoint. Its stated basis is the paper Searching for a Mirror. Re-checked 2026-09-21.
Both are here because they were said, not because they were seen. Neither has a repository, a hostname or a running instance, so there is no check to hand you — which is the point of listing them separately from everything above.
AtLive
The map · added 2026-09-25
The Paxeer Atlas
The same stack as a map: seven layers from the consensus floor up to the agents, the pipes between them drawn to scale. This page answers what exists. The atlas answers how it fits together — and where the money actually moves.
Receipts on everything. Every layer, node and flow carries its own evidence: a file path, an address or a hostname you can go check. A node with none shows a dashed outline instead of a confident caption.
Today versus after the fork. A toggle separates what runs now from what the v6.6 upgrade switches on — named taps, dated markers, no promises.
It reads without JavaScript. The whole map is server-rendered as an outline first; the animated scene is a layer on top, not a replacement for it.
Start with official documentation and repositories. Then check the claims yourself — these four take a couple of minutes each, and they answer for themselves.
Check a tokenOpen the explorer, confirm the contract is verified, read supply and holders. No code at the address means there is no contract there.
Do not measure usage from the explorer hereThis stack is designed so ordinary agent activity never touches the chain — most of it collapses into one periodic checkpoint. A low transaction count is what the architecture predicts, not evidence of anything. Empty blocks do not prove the network is dead; blocks being produced do not prove it is alive. Both arguments use the same broken instrument. What the explorer does tell you: that blocks are being produced, how many producers there are, and whether a given contract exists and is verified.
Check a claimTake a published contract address, pull its bytecode, hash it, and compare against the hash in the repository.
Check the protocol versionOpen include/layerx/lxp_protocol.h in the LayerX repository. It names three wire-protocol versions and sets which one is active. Today the active version is 2; version 3 exists and is tested but must be explicitly selected.
⚠Nodes and the scanner are being updated during the LayerX integration — verify anything the dashboard reports against the explorer.
A project can have real code, real websites, and real engineering activity while still lacking proof of production adoption. Those are different questions. This resource keeps them separate.
04 / Something you can check yourself
Addresses, read off-chain.
⚠ The address is the only thing that tells them apart
Impersonator SID mints exist on Solana. Do not try to spot them by supply or by revoked authorities — the real SID also has a ~989M supply and revoked mint and freeze authorities, and its on-chain name is literally “Sidiora Markets”, which is what the fakes use too. Those tests do not separate anything. Match the full address, character for character, or do not proceed.
SIDSolana mainnetchecked on-chain 2026-09-07
5w3wVdJaESaJKyLmStM6Hv9UyUkmZ1b9DLQquAqqpump
getAccountInfo — Token-2022 mint, 6 decimals, supply 989,876,430.663924, mint and freeze authority both revoked, metadata update authority null. No transfer fee.
PAXSolana mainnetchecked on-chain 2026-09-07
BxUXwhg5y1noXwW5CQ3p9RpFzKRFEPwDDKQkdMXWxPAX
getAccountInfo — Token-2022 mint, 6 decimals, supply 597,446,482.156325, mint and freeze authority both revoked.
Charges a 0.15% transfer fee, capped at 1,000 PAX. The mint carries two fee records: 0.05% capped at 0.1 PAX through epoch 909, and 0.15% capped at 1,000 PAX from epoch 911. The chain is at epoch 1030, so the higher fee is the one in force. The fee config authority is null, so the rate is locked. 3,014.499251 PAX sits withheld, withdrawable by 8GfC2xmAcQ4jvWoKc5v1qd1bD93e5yrM5XnBE6rzECXt.
eth_getCode and eth_getStorageAt against chain 125 (eth_chainId returned 0x7d) — Sidiora (SID), 6 decimals, total supply 458,859,517.886864, 262 holders, 278 transfers. The address holds 130 bytes of bytecode: an ERC-1967 proxy that delegatecalls to 0xAd86a65d5E14F39AAfAb4eDF331FB17dA866462e, an 8,526-byte implementation.
This is an upgradeable token, which the explorer token page does not show. The ERC-1967 admin and beacon slots both read zero, which is the UUPS arrangement — the upgrade function lives in the implementation. Neither the proxy nor the implementation is verified on paxscan, so the upgrade gate cannot be read from source; owner() returns 0x1255d84066f579e7b7A3dF4296E960d59fC05b32, an account with no code. Treat the token logic as changeable until the source is published. Separately, the explorer displays a price of $4.91417 and a market cap of about $2.25 billion against 278 lifetime transfers, the most recent on 2026-07-30. Both figures come from the explorer, not from an independent market.
LayerXMirrorArchiveone contract, four deployments4 of 4 verified
A single archive contract deployed across four test networks. It has no payable function, no token interface, no withdrawal path and no admin key — it can only store archive chunks, and nothing can be taken out of it. All four are testnets and all are currently empty.
LayerXMirrorArchiveEthereum Sepoliachecked on-chain 2026-09-07source published
0x757ab4D9dC18C92E23BBa1eEA005Fa3f1a544401
eth_getCode — 4,355 bytes of runtime bytecode present.
LayerXMirrorArchiveBase Sepoliachecked on-chain 2026-09-07source published
0xcF8fa2D08190993BeBECC211f641E319EdF5a6C6
eth_getCode — 4,355 bytes, byte-identical to the Sepolia deployment.
LayerXMirrorArchiveSolana devnetchecked on-chain 2026-09-07source published
ExU8LpHSSit5LocqLS8KJT6hayvXXZv4MSQkWB2utUr4
getAccountInfo — executable, BPFLoaderUpgradeable. Its programData account carries a null upgrade authority, so the program is permanently immutable.
LayerXMirrorArchiveRobinhood testnet (46630)checked on-chain 2026-09-07source published
0xcF8fa2D08190993BeBECC211f641E319EdF5a6C6
eth_getCode — 4,355 bytes, byte-identical to the Sepolia and Base Sepolia deployments. eth_chainId returned 0xb626 (46630), confirming the network.
Every row here was queried directly against a public RPC on 2026-09-07— Solana mainnet and devnet, Ethereum Sepolia, Base Sepolia, Paxeer chain 125, and Robinhood testnet 46630 — and the line beneath it says what came back. Nothing was taken from a research note on trust. Re-check before you act on any of it: an address that was right in September is not automatically right later, and one of these contracts can be changed by whoever holds its owner key.
PaxScan is where you check the chain yourself, but read it knowing two things.
The totals panel is behind the chain and catching up, not stalled and not current. On 2026-09-25 it read 21.6 million blocks against a head of 24.5 million, and advanced on a second read 35 seconds later. The tip itself is indexed live: ask for the newest block by height and it comes back fully indexed, seconds old. So read heights and per-block counts off the block pages, and ignore the totals panel entirely.
For the opposite experience, open the Machine Genome registry: signed records and a published Merkle root you can read and check today.
An entire prior generation of Paxeer is still live and indexed, and search engines surface it first.
The hyperpaxeer.com root is back. It and www. return 200 from 31.220.74.239 and serve a page titled “Paxeer Wallet Dashboard” — re-checked 2026-09-25, having been dark since at least 2026-09-04. Its subdomains did not come with it: docs. and node. resolve but answer nothing, and dao. has no DNS record. Documentation lives at docs.paxeer.app ↗ regardless.
whitepaper.paxeer.app is up (February 2026 edition).
status.paxeer.app is a 404. There is no public status page.
app.sidiora.xyz presents a certificate for a different host (ERR_CERT_COMMON_NAME_INVALID). Do not enter anything there.
thirdweb.com/paxeer-network describes Paxeer as a testnet on chain 8000.
labs.paxeer.app uses older HyperPaxeer and ChainFlow vocabulary.
LayerX names sidiora.xyz as one of three canonical domains for confirming which LayerX is real. It does not resolve. Verify through the GitHub organisation and layerx.paxeer.app instead.
