Skip to content

BACKLOG.md

Latest commit

 

History

History
255 lines (244 loc) · 15 KB

BACKLOG.md

File metadata and controls

255 lines (244 loc) · 15 KB
Raw

Autonet Backlog

Status banner (2026-07-10) — HISTORICAL RECORD. This file was "THE board" through substrate v4.1 (2026-07-09), but its head is now stale: the v4.1 rep/ATN mint split, supply-pegged β, and 3-field merkle leaf described below were superseded on 2026-07-10 by fees-only emission + money-only close (2-field leaf; REP claimed DAO-side 1:1; β dropped), shipped as v0.7.0. The active board has moved to session/internal tracking; this document is retained as a historical snapshot, not the live work list. For the current system read docs/tool_substrate.md (the Decision (2026-07-10) section) and docs/README.md. The body below is left intact.

Consolidated 2026-07-06; updated 2026-07-09 after substrate v4.1. This is THE board; update in place as items land or get ratified.

State: substrate v4.1 "gradient trust" BUILT 2026-07-09 (close + contract + surfaces; UNCOMMITTED, sim-validated in experiments/econ_attest/, spec = docs/tool_substrate.md Decision 2026-07-09). The vet GATE is retired (tools mint from first attested use); mint is rep/ATN-SPLIT under a supply-pegged β cap (zero-rep usage mints ATN, not reputation); drift weight = rep_share × credibility with NO ε floor + continuous reversal-aware credibility docking; vets survive as inspection reviews; recordTrainingForEpoch(amount, repAmount, epochIdHash, proof) with a 3-field merkle leaf + TrainingRecorded gains repAmount. PENDING (v4.1):

  • FLAG-DAY redeploy — v4.1 changes close output/CID + contract ABI + merkle leaf; needs a clean genesis on a fresh Substrate + all daemons on the v4.1 build before the next federated close (supersedes the v3 flag-day window below).
  • G1 — channel fee — RESOLVED 2026-07-10 (ATN-only + fee at settlement). Decision (user): service commerce is ATN-only; ServiceMarket.PaymentChannel.closeChannel now routes the provider payout through Substrate.payForService, so the 2.5% burn/recycle fires on the canonical rail (provider receives net; refund is fee-free). The token field is gone from the registry + channel. The non-ATN design call is moot. Built, not deployed (needs the v4.1 flag-day redeploy).
  • S₀ recalibration at launch (USER): BETA_S0=5000 (50 epochs × 100 pool) is the sim's engineering pick over a short 300-epoch horizon; retune to the real epoch cadence + pool size before mainnet.
  • v4.1 params seeded-not-blessed: BETA_MIN=0.05, CRED_DELTA=0.7 (δ EMERGED, others SEEDED), CRED_MASS_FLOOR=3.0, CRED_FLOOR=0.1, CRED_RECOVERY=0.10, pool 100. δ=0.7 and the necessity of a non-constant supply-pegged β emerged from sims; the rest are conventional, tunable post-launch.

Prior state: substrate v3 shipped and PUBLIC (master pushed 2026-07-08). Reviews replace debates (docs/tool_substrate.md, Decision 2026-07-08): mint = usage alone, per-axis reviews drift tool positions, work units left consensus, debate rail retired/dormant. Both economy E2Es re-verified green ON v3 against a live chain (tool economy 7/7 — mint = log1p(1) exactly; venture loop 8/8). Also landed: Secrets tab + vault audit trail, the unified Substrate visualization, Tools Local/Substrate catalog, review step wired into the agentic loop, register() idempotency fix, and the README-as-living-paper ("Autonet — The Recursive Principial Body") now serving both app whitepaper surfaces.

V3 OPS (time-sensitive):

  • FLAG-DAY (updated post-genesis 2026-07-08 evening): daemons need the CURRENT MASTER build (v0.6.0 unreleased — release pending user go), not v0.5.x — payForService rename + reputation-source voice
    • vault minter changed contract ABI and close inputs; v0.5.x builds cannot talk to the genesis Substrate. Restart onto master (or the next release) + re-register agents (now with owner-binding popups).
  • Deploy Firestore rules for substrate_viz (world-readable, client-unwritable) + run the substrate publisher next to a daemon (scripts/indexer/ is gitignored ops — lives outside the repo).
  • Chrome visual pass on the unified Substrate view (organic-feel knobs: organic_blob.dart constants, _kDescendFraction, _kOpennessSpeed).
  • Secrets broker E2E (vault setup + ATN_WORKER_ISOLATION=1) still unrun.

TO FIX / HARDEN:

  • The v3 sybil trade — MITIGATION BUILT 2026-07-08 (balance-weighted voice, spec addendum in docs/tool_substrate.md): callers now collapse to HOUSEHOLDS (proven owner wallet) before log1p damping, and each household's usage/review credit scales by ε + household_ATN/supply (linear ⇒ balance-splitting is weight-neutral; zero-balance identities carry at most ε = 0.05 PROVISIONAL, needs blessing). Wired: close math (federated_reconcile.py), chain sourcing (voice_state.py → driver voice_source hook), fleet_voice WS + Owner page in atn_web (wallet dropdown). Tests: tests/test_voice_weights.py (10) + all close families green. Reads are PINNED to the previous epoch's anchor block (getAnchor(...).blockNumber) and served WITHOUT archive state: Substrate.sol ATN now carries IVotes-mechanism checkpoints (balanceOfAt/atnTotalSupplyAt, Trace208, no delegation layer — undelegated getPastVotes would read 0) and the agent/owner maps derive from event logs up to the snapshot. Chain tests tests/test_voice_snapshot.py (pin property verified); stale phase5_6/phase7_2 mint fixtures migrated to v3 tool events. REMAINING: (a) the ε floor is the residual sybil surface — bounded, but the dormant CON/bust rail stays the named backstop; (b) covert harm still has ONLY vetting as its dedicated defense — vet params (below) remain load-bearing. NOTE: close output changed + CONTRACT changed (checkpoints) — the shadownet redeploy now carries this too; same flag-day window as v3 (fleet is still one daemon).
  • world_persistence artifact_digest gap — FIXED 2026-07-08: digest mirrored into node.metadata (serializer carries it) + lift_artifact_digests rehydrates after every restore_world (world_persistence + both state_sync paths); roundtrip test pins it. NOTE: metadata now carries the digest → checkpoint content/world_cid changes — rides the open flag-day window.
  • Agentic-loop review guidance lives only in tool descriptions + _TOOL_CATEGORY_NOTES; agents with a custom system_prompt bypass the category notes (they still get the injected closing review turn).

USER-ONLY DECISIONS (nothing proceeds without these):

  1. Emission rate — shape RATIFIED + BUILT 2026-07-08: fee-recycled emission (docs/epoch_economics.md Decision) — pool = BASE floor + burned service fees recycled (wash-proof by conservation; treasury share to the DAO). Remaining user-only: bless the PROVISIONAL values — SERVICE_FEE_BPS (2.5%), FEE_TREASURY_BPS (50% of fee), BASE_EMISSION_PER_EPOCH (100 ATN). Fee params are on-chain constants (redeploy to change); BASE is consensus-flag-day.
  2. repPerReputation rate on ReputationMirror (default 1:1, timelock- governed) — the direction/quality exchange rate; flagged as the most consequential unbuilt decision; wants sims.
  3. Vet params blessing: sim says N=3 (code has 2.0); royalty×weight slash fix (rubber-stamper out-earns careful vetter in deep pools). Post-v3 these are LOAD-BEARING — vetting is the only dedicated defense against covert harm (see sybil trade above). Note the tool-economy sim memo is quarantined (models v2 mint); re-run the sweep against usage-only mint before citing absolute numbers.
  4. Channel challengeWindow (E2E uses 3600s).
  5. Push master to origin — DONE 2026-07-08 (no tag, no release). atn_web commits remain local-only.
  6. CLAUDE.md is gitignored — machine-onboarding contract doesn't travel with clones. Commit / genericize / keep local?
  7. Charter-anchor governor handoff — DONE 2026-07-08 (genesis v2): CharterAnchor governor = jurisdiction timelock; charter v1 ANCHORED on-chain via executed DAO proposal.
  8. Admin agent: merged but placeholder budget (100k tokens/day) — tune + enable; it starts spending on a 30m schedule once live.
  9. Microsoft Store: pay $19, fill the two msix identity placeholders; create WINGET_GH_TOKEN; first winget submission via wingetcreate new. (RELEASING.md in atn_web has the checklist.)
  10. ReputationMirror activation — DONE 2026-07-08 (genesis v2): setMinter(mirror) EXECUTED via DAO proposal; mirror can mint.

RATIFIED, NOT BUILT:

  • ATN = money, reputation = voice — BUILT + DEPLOYED 2026-07-08 (genesis): voice weights read soulbound reputation at the pinned block (reputationOfAt checkpoints); AutonetVault (renamed from ParityVault) live on shadownet as Substrate's immutable minter, buy-only, parity via jurisdiction Registry entries; vault mint moves balance, never voice. Genesis addresses in registry.json. Historical design entry follows:
  • ATN = money, reputation = voice (2026-07-08 evening) — the jurisdiction merge, attempt 6. Once ATN is buyable, balance-based voice = purchasable review power (pumped ratings would rig probe_tools discovery). Decision: switch the close's voice-weight source from ATN balance to SOULBOUND REPUTATION (earn-only, minted 1:1 with ATN, non-transferable) — same household collapse, same pinned-anchor reads (needs rep-at-block derivation: checkpoint or TrainingRecorded logs). ATN stays purely economic. ParityVault (buy-only v1): CoinOffering clone minting Substrate ATN via an IMMUTABLE minter address in Substrate's constructor (no-admin-keys preserved); payment tokens + rates = jurisdiction Registry entries (jurisdiction.parity.<token>), timelock-governed; ERC20 proceeds → DAO treasury. Rides the owed redeploy (payForService rename + clean genesis via sacrificial smoke instance). BLOCKED on user: make the trustless-contracts DAO factory deploy a jurisdiction WITH a coin-offering-shaped satellite slot; then the genesis event chains #7 (CharterAnchor governor handoff) + #10 (ReputationMirror activation) + the ParityVault minter grant.
  • Tokenized vault shares (2026-07-06): ERC20 per vault, book-value redemption (burn → unvested principal + accrued revenue), checkpointed halt votes, CoinOffering dividend pattern ported. Boats project re-conceived as the flagship venture agent.
  • Verifier rewards migration: mint-funded bootstrap → bps-of-raise fee once volume exists.
  • Capability-gap pricing: coverage-atlas gaps as mint multiplier (doctrine only, never implemented).

ENGINEERING FOLLOW-UPS (flagged by builders, none blocking):

  • Jurisdiction governance timing: the genesis deploy's VOTING_PERIOD constant (50400, "~1 week in blocks") lands on-chain as 3,024,000s (~35 days — timestamp clock, units mismatch). Make votingDelay/ votingPeriod/timelock minDelay env-configurable in deploy-autonet-jurisdiction.js and use short values on shadownet. RESOLVED same evening: durations made env-configurable (seconds, no hidden x60), jurisdiction REDEPLOYED with 0/120s/0 (genesis v2, all addresses rotated - registry.json is truth), and the genesis proposal (mirror minter grant + charter v1 anchor) went propose->vote->queue->execute LIVE in ~3 minutes. Mirror IS a RepToken minter; charter v1 IS anchored. Etherlink log-range cap requires PROPOSAL_ID/FROM_BLOCK on the proposal scripts.
  • Owner binding is UNWIRED in the live path (2026-07-08): the daemon registers agents via plain registerAgent (ownerless, agentOwner=0x0); registerAgentBound/rotateOwner exist only in the contract + tests + E2E script. Balance-weighted voice therefore falls back to household=agent-id for every real agent — the human-wallet household model is built but unreachable from the product. Wire: frontend collects the EIP-712 OwnerBinding signature from the connected wallet during registration (fold into the existing popup flow), daemon submits registerAgentBound; Owner page gets a "rebind fleet to this wallet" surface (rotateOwner) — also the key-loss recovery UI.
  • Rename payForInference → payForService (next redeploy): it settles services/venture revenue generally; "inference" is residue from the decentralized-inference framing and invites conflation with the sponsor/dependent provider pipe (which never touches ATN).
  • TrialRunner: transport seam → wire to live WS service-request client; OnChainService needs a generic contract-call so attestTrial submits directly (currently returns calldata).
  • Orchestrator drawer: add the 120s stall timeout the chat surface got (silent-stall symmetry).
  • Adopted tools: OS-level isolated runner (vault track) as the wall behind tool_guard's audit-hook tripwire; macOS isolation untested.
  • Indexer: mirror the new events — ToolRegistered, VentureCreated, CharterAnchored, service registry — into Firestore collections (tools/services collections are documented as future, not implemented).
  • MCP server quirk: any Claude Code session with the atn MCP configured silently stands up a full runtime on :7700 — should attach to an existing daemon when one is running, self-host only otherwise.
  • Registrations carry-over GC (someday): the consensus map grows monotonically; mechanical GC = drop digests with zero usage for N epochs, re-registration re-admits. Not a design item.
  • Shadownet redeploy — DONE 2026-07-08 (user-gated, user said go): Substrate 0x112F617De881ef0e94B010d9756aa8aef2794208 (checkpointed ATN + owner binding + endpoint surface), CharterAnchor 0x2B239bBE...c47C (governor = deployer pending #7 handoff; charter v1 anchored, hash 5756ed3a...), ServiceRegistry 0x080346d2...6400. PaymentChannel NOT deployed (challengeWindow = decision #4). registry.json + Firestore settings/network updated, indexer@shadownet restarted, testnet smoke green (anchor + dual mint verified on-chain), live read_voice_state verified (snapshot-pinned, ε floor pre-mint). REMAINING user-side: restart daemon(s) onto this build, re-register agents (rpb_reconcile_registrations flips them local-unregistered).
  • atn_web: Flutter SDK one-line patch (VS 2026 generator) dies on flutter upgrade; CMAKE_POLICY_VERSION_MINIMUM=3.5 needed per local windows build until firebase fixes its CMakeLists.

EXPERIMENTS:

  • Phase 9 (equilibration at depth) — still unrun, still the pre-committed FINAL test of the equilibrated kernel. (Post-v3 the live path doesn't read equilibration, but the pre-commitment stands: run it or formally retire it via a dated decision.)
  • Phase 11 (proposed, unregistered): minimum-model-tier gap bare vs substrate-assisted, claim = gap widens as tool corpus grows — the decentralization mandate made falsifiable. Phase-10 harness reuses.

THE ACTUAL CRITICAL PATH (non-code): one real customer / demand beachhead — a paid service category with real buyers where crypto settlement is an advantage (hardware/compute rental for agents = best named candidate). Show one other human the venture-loop E2E; fund one tiny real venture. The v3 flywheel (reviews → ranking → usage → mint) cannot spin with one participant. The network begins the moment it isn't just its author.