Doc 2175 - ZOL & ZOE Sovereign-Agent Progression: Required Deliverables + Architect's Progress Report
Status: DESIGN / ASSESSMENT ONLY. Nothing in this doc activates any capability. No economic, on-chain, or public-autonomy action is enabled by it.
Date: 2026-08-01
Trigger: Brandon's 5-stage "ZOL & ZOE Sovereign-Agent Progression" proposal (staged, cautious, $0-default). This doc is the "Required Deliverables" Brandon asked for before activation, plus the Architect's Progress Report.
Grounding: A read-only file-by-file audit of origin/main (2026-08-01) + related repos. Every "EXISTS" traces to a file:line; every "PROPOSED-ONLY" is explicitly no-code. Per Brandon's rule: no PR is cited as proof unless its diff contains the code.
Related: [[project_brandon_organism_directives]], [[project_brandon_two_plane_architecture]] (Heart), [[project_dreamnet_trust_layer]], [[project_dreamnet_spore_sdk]], doc 2170 (diagnostic), 2171 (reaudit), 2169 (daily-note), 2174 (bidirectional relay).
The organism can ACT (receipts on real tool actions) and REMEMBER (the afferent digest routes receipts to memory). It cannot yet be VERIFIED, CERTIFIED, TRUSTED, or ECONOMICALLY CONTROLLED. It is a reliable operational loop with observability - not yet a sovereign system. ~25% of the 5-stage vision is production code, ~50% is built-but-flags-OFF, ~25% is research-only.
Stage 0 (the readiness gate) is NOT fully passed - so per Brandon's own rule, no later stage may proceed. The single highest-leverage move is to CLOSE STAGE 0 (turn the built organism ON), not to grow new limbs.
| Stage | Score | Shipped (file:line on origin/main) | Gap |
|---|---|---|---|
| 0+5 Proprioception | 7/10 | receipts.ts (emits on repo-improver/error-remediation/hermes), receipt-envelope.ts (dreamnet.receipt.v1 + sha256, PR #2515), afferent-digest.ts (nightly receipts->memory, #2774), packages/heart-fleet/* (leases+liveness+outbox+reconcile, 5 props), cost-governance.ts ($10/day cap + alerts) |
No claims table; memory retrieval ~88% orphaned; no reputation layer |
| 1 Identity/Certification | 2/10 | src/lib/ens/resolve.ts + subnames.ts (ENS resolve + subname request, NOT wired to routing) | Zero ToolGym/University/competency-claim code |
| 2 Economics | 3/10 | cost-governance.ts (LLM budget cap live); approval_class field in schema | No ERC-4337/CDP code; approval gates are POST-facto (audit), not pre-dispatch; no on-chain spend policy |
| 3 Temporal (WarpKeeper) | 0/10 | (Sparkz operational, but zero history) | No snapshots/state-diff/rollback/0xSplits code anywhere |
| 4 Farcaster agency | 6/10 | src/lib/publish/farcaster.ts (Neynar posting + rate-limit error handling); ZOL (@zolbot) + ZOE both post | No mass-reply guard, no PRE-approval gate, no unified rate limiter |
Flags (all default OFF, verified): ZOE_HEART_FLEET_CANARY, ZOE_REPO_IMPROVER_LEASES, ZOE_OUTBOX_DEMO. So the Heart lease/fencing/outbox is coded + tested but governs nothing in prod today.
| Layer | Owner (authority) | Depends on | Sovereignty note |
|---|---|---|---|
| Identity (ENS/DID) | Zaal (retains, per proposal) | ENS resolver; a registrar for *.dreamnet.eth (unverified) | Do not claim zol/zoe.dreamnet.eth issued until resolution + ownership verified |
| Memory | Zaal (dotfiles repo is source of truth) | Supabase (receipts), the memory store | One brain, many lanes |
| Execution authority | ZOE orchestrator; Heart leases (when ON) | packages/heart-fleet, agent_runs table | Currently ungoverned (flags off) |
| Economic authority | Zaal, explicit, per-action | (none live) - CDP/4337 not integrated | $0 default; human-gated; human-executed. Claude designs, never executes |
| Federation | ZAO (first conformant node) | DreamNet Spore SDK (spec) | ZAO impl is independent; SDK = spec, not a dependency for survival |
- Runaway spend (Stage 2): an agent with a funded account drains it. Mitigation:
$0default;$50/dayceiling only after explicit promotion; per-action limit; contract+function allowlist; human approval above threshold; pre-dispatch approval gate (does not exist yet - build before any funding); simulate-before-submit; idempotency keys; emergency pause. - Key compromise (Stage 2): agent key leaks. Mitigation: keys in
~/.zao/private(chmod 600, never committed - secret-hygiene.md); rotation procedure; separate keys per agent (ZOL != ZOE); on-chain allowlist limits blast radius even if key leaks. - Impersonation / engagement farming (Stage 4): agent posts as a human, mass-replies, fake engagement. Mitigation: no impersonation, no mass reply/follow, per-channel daily cap (build - none exists), conversation-quality threshold, approval-class for sensitive comms, receipts on meaningful actions.
- Fabricated competency (Stage 1): an agent claims authority it did not earn. Mitigation: competency claims must carry test traces + evidence + verification + expiry; authority increases only when claims REPEATEDLY survive verification (Brandon's rule).
- Vacuous receipts (Stage 0/5): a receipt emitted for a no-op, inflating apparent throughput. Mitigation: receipts assert the real effect (silent-failure-guard.md), the day-7/done-condition promotion signals are behavioural not self-reported.
- Split-brain execution (Stage 0): two instances double-act. Mitigation: Heart leases/fencing (built - flip the flag to make it real).
simulate -> verify -> testnet -> limited guarded production -> review -> broader authority
- Phase A (now): CLOSE STAGE 0. Deploy the merged proprioceptive layer; verify receipts flow + the digest runs; flip
ZOE_REPO_IMPROVER_LEASES=true+ verify a lease fences a real path; do the memory curation pass; define the emergency-pause + owners. NO economic/certification/public-autonomy work until this passes. - Phase B: Stage 4 guardrails (cheap, high-safety-value). Build the mass-reply guard + pre-approval gate + unified rate limiter for the ALREADY-LIVE Farcaster agents. This hardens what's already public. No new spend.
- Phase C: Stage 1 pilot (design + sim). One ZOL tool (
post_farcaster) through a ToolGym-style competency harness; produce a competency claim with traces. Identityzol.dreamnet.eth= verify resolution/ownership, do NOT claim issued. - Phase D: Stage 2 scaffolding (SIM/TESTNET ONLY). Design the ERC-4337/CDP account with
$0default + allowlists + pre-dispatch approval gate + simulation harness + reconciliation. Testnet only. NO mainnet funding, NO real value. Human-gated at every promotion. - Phase E: Stage 3 (observation only). WarpKeeper = reliable temporal observation of Sparkz (snapshots/diffs) FIRST. Automated keeper actions stay disabled until contracts identified + conditions specified + financial side-effects simulated + emergency shutdown exists.
- Code: every change is a PR; deploy has auto-rollback (zoe-autodeploy verify-on-fresh-checkout + boot-health + rollback-on-error). Flags default OFF = merging changes nothing until flipped; rollback = unset the flag.
- Economic: emergency pause (build) halts all agent txns;
$0default means the floor is safe-by-default; testnet-only until reviewed. - Public: operator pause on the Farcaster agents; receipts + memory make every action auditable + reversible-in-record.
- Stage 0 closeout: ~$0 (deploy + flag flips + a curation pass; existing infra).
- Stage 1 pilot: ~$0 (a harness + one tool; local/OpenRouter for eval).
- Stage 2 sim/testnet: ~$0 real value (testnet gas is free/faucet); CDP/AgentKit may have platform costs - verify before committing.
- Stage 3 observation: minimal (snapshot writes to Supabase).
- Stage 4 guardrails: ~$0 (code).
- Real money only enters at Stage 2 "limited guarded production" - Zaal's explicit, separate decision, not covered here.
- Stage 0 PASS = receipts from >=3 loops in the last 24h (not 1); the digest ran + wrote a memory record;
ZOE_REPO_IMPROVER_LEASES=truein prod + a logged lease acquire/skip; a retrieval test answers a prior decision without re-prompting; an emergency-pause + owner registry exist. - Stage 1 PASS = one competency claim exists with test-trace + evidence + verification + expiry; identity resolution verified (or explicitly "not issued").
- Stage 2 PASS = a testnet txn simulated -> verified -> submitted -> reconciled, with the pre-dispatch approval gate blocking an over-threshold action;
$0mainnet. - Stage 3 PASS = N weekly Sparkz snapshots with a working state-diff + replay; zero automated money movement.
- Stage 4 PASS = the mass-reply guard + per-channel cap + pre-approval gate live; zero impersonation; receipts on meaningful actions.
Readiness gates passed: Stage 0 partially (proprioceptive code merged: receipts, envelope, afferent digest, Heart package). Not passed: leases-govern-real-execution (flags OFF), retrieve-without-reprompting (memory 88% orphaned), emergency-pause + economic owner (do not exist).
Safe to activate now: the Stage-0 closeout (deploy + flip the Heart flag + curation - low risk, reversible), and the Stage-4 guardrails (mass-reply guard etc. - pure hardening of what's already live). Nothing else.
Design-only (must NOT activate): all economic/on-chain (Stage 2), certification issuance (Stage 1), WarpKeeper keeper-actions (Stage 3), any new public-autonomy without the guardrails.
Evidence produced: the grounded current-state map (25% prod / 50% flags-off / 25% no-code), the per-stage scores, the anti-fabrication trace (all EXISTS -> file:line).
New failure modes discovered: (1) approval gates are POST-facto audit, not PRE-dispatch - a real gap before any spend. (2) The already-live Farcaster agents have NO mass-reply/impersonation guard - a live risk independent of the sovereign vision. (3) The Heart is 100% tested / 0% live - fencing exists but governs nothing, so split-brain is currently only prevented by the one-instance convention, not the lease layer.
Single highest-leverage next milestone: CLOSE STAGE 0 - flip ZOE_REPO_IMPROVER_LEASES=true, deploy + verify receipts flow + the digest runs, do the memory curation pass. It converts the whole 3-week build into live capability and unblocks everything after it, at ~$0 and low risk.
Three ranked alternatives (if not Stage 0 first):
- Stage-4 guardrails (mass-reply/pre-approval) - hardens a LIVE public surface; high safety ROI, low effort. (Do this alongside Stage 0.)
- Stage-1 ToolGym pilot (one tool) - proves the certification loop cheaply; but premature before Stage 0 records reliably.
- Stage-2 sim scaffolding - highest future value but highest risk-of-drift; do NOT start before Stage 0 + the pre-dispatch gate.
1-3 iteration roadmap: (1) Close Stage 0 + Stage-4 guardrails. (2) Stage-1 pilot + verify identity resolution. (3) Stage-2 sim/testnet scaffolding with the pre-dispatch gate. Re-run the readiness gate between each.
Self-critique: This assessment is a snapshot; "merged" != "running" - the Stage-0 "7/10" assumes the merged code deploys cleanly, which is unverified until the VPS runs it. The scores are judgment, not a metric. The cost estimate is order-of-magnitude. The audit read code, not runtime - a merged file can still fail at boot (the esbuild lesson).
Next logical progression (without waiting for a prompt): Close Stage 0 - and it is deploy-and-verify work, all in-lane and reversible, not a new build. The economic and public-autonomy stages stay design-only and human-gated until Stage 0 passes its acceptance criteria. This doc activates nothing.