Autonet's design specs, experiment records, and reference material. For
agent/machine onboarding start at the repo-root CLAUDE.md; for the
one-paragraph "what is this" start at the repo-root README.md.
Status (v0.7.0, 2026-07-10): beta on testnet. Substrate.sol and the
DAO/economy suite are deployed on Etherlink Shadownet; the canonical
jurisdiction is a werule-platform DAO (governor 0xD5691B7c…) with the
addresses of record in registry.json (daemons resolve the network from
it — repo-root in a dev checkout, else the GitHub-raw copy cached under
~/.atn/). The tool-economy paradigm is the shipped system: tool_substrate.md
is the core spec (read its Decision (2026-07-10) section first). The
architecture-reference set and the two guides below still describe the
retired pre-substrate paradigm and are banner-marked historical.
Status conventions used below:
- Living spec — current design or built feature; safe to edit as the design evolves.
- Historical record — DO NOT EDIT — a pre-registered contract or a results doc. These were committed before/after a run as an integrity record; amend only via a follow-up doc, never by rewriting.
- Stale (pre-substrate) — describes the dissolved FedAvg / VL-JEPA / role-split paradigm. Kept for history; do not treat as current.
| Doc | What it covers |
|---|---|
tool_substrate.md |
The core spec — start here. Tools as the primary substrate item: manifest, mint formula, combo damper, review-driven position drift, adoption rail. Read the Decision (2026-07-10) section first: fees-only emission, money-only close, REP claimed DAO-side, vet gate retired (tools mint from first attested third-party use). Phase-10 amendment inline. |
services_market.md |
The remote-API market rail (ServiceMarket.sol): registry + EIP-712 payment channels, behavioral trust, ATN-only pricing (2.5% fee taken at settlement through Substrate.payForService), no substrate standing. |
charter_anchor.md |
CharterAnchor.sol — governed anchor for the charter version (values stay off-chain); drift detection built, migration deferred. |
local_e2e.md |
The two economy proof-of-life scripts (scripts/local_e2e_tool_economy.py, local_e2e_venture_loop.py); also the reference walk-through of the venture loop. |
ledger_pricing.md |
Post-phase8 default: ledger (net_score tree recursion) replaces geometric equilibration for mint pricing; the arm-B context-render rule. |
two_plane_inference.md |
Claim graph demoted to a verdict layer; artifacts (full payloads) live in blob store + ArtifactIndex, referenced by sha256. |
epoch_economics.md |
Epoch close + candle mechanics. Note: the emission model shipped as fees-only (2026-07-10) — the epoch pool is burned service fees only, no base emission (Σ minted == Σ burned); see tool_substrate.md Decision (2026-07-10). |
agentic_loop.md |
Hardening spec for send_orchestrate (the anthropic/openai_compat/ollama agent loop) and its adapters. |
cross_platform_isolation_design.md |
Isolation + vault port plan — IMPLEMENTED + Linux-verified (POSIX AF_UNIX / SO_PEERCRED, tracked-PID kill). |
epoch_economics.md · ledger_pricing.md |
(see above) the economic core as shipped. |
auto_update_design.md |
Daemon auto-update: stage-on-poll, apply-on-next-boot (running daemon never self-restarts). Implementing. |
anti_tamper_design.md |
Consensus-node anti-tamper — the "other half" of auto-update. Designed, pre-implementation. |
unified_agent_design.md |
Fractal agent unification (orchestrator ≡ child agent) — the definitive design that drove the atn/ runtime shape. |
| Doc | Status |
|---|---|
phase8_prereg.md |
DO NOT EDIT — prereg committed before any contest call ("does equilibration earn its complexity?"). |
phase8_results.md |
DO NOT EDIT — results: equilibration beat vote-count by +0.127, short of the 0.25 bar → demoted. |
phase9_depth_experiment.md |
DO NOT EDIT — the pre-committed, still-unrun final test of equilibration at graph depth. |
phase10_prereg.md |
DO NOT EDIT — prereg: "does executable ground truth deliver what prose debate couldn't?" (H1 declared FINAL). |
phase10_results.md |
DO NOT EDIT — results: H1 REFUTED; evidence-backed standing still separated defective from correct tools (AUC 1.000). |
| Doc | What it covers |
|---|---|
jurisdiction_wiring_map.md |
Read-only survey (2026-07-05): how the tokenized-project → on-chain surface wires together. Nothing built by the doc itself. |
value_props_head_to_head.md |
Pre-merge comparison: text-claim substrate vs tool substrate — what each promised and delivered. |
DORG_CHAT_INTERFACE_DESIGN.md |
dOrg deployment + chat-interface layer design (mostly not yet built). |
DRAFT_COMMON_SYSTEM_PROMPT.md |
Draft shared base system prompt for cognitive agents (final lives in delegate_prompts.py). |
TRIMMED_MEAN_USAGE.md |
Byzantine-resistant trimmed-mean aggregation usage guide (aggregator-era). |
BACKLOG_TRAINING_DATA.md |
Training-data pipeline analysis. Stale (pre-substrate) — references the VL-JEPA path. |
architecture/OVERVIEW.md · CONTRACTS.md · NODES.md |
Architecture reference set. Historical (pre-substrate) — each carries a dated banner; the current architecture is this README + tool_substrate.md + CLAUDE.md. |
guides/QUICKSTART.md |
Current install/run quickstart: pip install autonet-computer, atn starts the daemon, deploy_substrate.js for local contracts. (The old "Absolute Zero" role-split quickstart was replaced.) |
guides/TRAINING_LOOP.md |
Historical (pre-substrate) — the "Absolute Zero" FedAvg role-split loop; native training is off the live path (see VALIDATION_FINDINGS.md). Banner-marked. |
CHEATSHEET.md |
Common operations quick reference — the contract-call/staking sections are historical (pre-substrate) and banner-marked; the current-command section is live. |
architecture-audit.md |
2026-03-25 audit of orchestrator vs child-agent divergences (historical snapshot). |
- BACKLOG.md — THE project board: user decisions, ratified-unbuilt designs, follow-ups, experiments, critical path. Update in place.
- PLAN.md — the original architectural blueprint (historical; the maintained map is the root README + CLAUDE.md).
- VALIDATION_FINDINGS.md — VL-JEPA validation record (historical — do not edit).