Build for the
next restart.
One owned-agent integration.
Build on what is demonstrated.
01 / VERIFY FIRST
Python 3.10+, Node 22+. No package dependencies, accounts or payments for offline checks. Download and extract the source kit. Run commands in a terminal.
unzip battery-pilot-kit.zip -d battery-kit cd battery-kit python3 -m unittest -v test_battery.py test_live_agent.py test_integrated_runtime.py test_paid_provider.py test_paid_resource.py python3 verify_evidence.py node --test devnet/anchor.test.mjs cd site node --test test-network.mjs test-wallet.mjs
55 Python tests check policy, failure, receipt recovery and unknown outcomes. Seventeen Node tests check memo/API/wallet boundaries; eighteen price/chain/outbox/financial-proof checks require the optional chain dependencies. Evidence verification checks saved hash/event relationships, not model attestation.
02 / RUN REAL INFERENCE
Install/start Ollama. Both approved models need sufficient memory and several GB of disk. They run locally without a paid provider key.
ollama pull qwen2.5:7b ollama pull llama3.2:3b # From repo root; start Ollama first. python3 run_recovery.py --state-dir live-state/my-first-proof python3 verify_evidence.py live-state/my-first-proof/live-agent.json
The controller runs three devnet briefs, forces exit 73 after the second durable receipt, starts Llama, verifies inherited state and writes live-agent.json. Use a fresh state directory; it refuses to overwrite a ledger.
If dispatch reached the model without a durable receipt, recovery stops with an unknown outcome. Do not delete state and silently repeat that task. The demonstrated boundary is after receipt commit.
[ ACTUAL OUTPUTS ]MAINNET / BOUNDED USDC CANARY
Our owned researcher now also reads mainnet and survives a real local process exit. Eight mainnet transactions are finalized: 3 test settlements, real financial crash/recovery, no replay payment, owner revoke and all 5 USDC returned. Generation preceded settlement and used local models, not paid-provider billing. The executable separately passes 15 local LiteSVM fixture checks.
INTEGRATED / LOCAL AGENT + MAINNET SETTLEMENT
[ PASS ] LIVE INTERLEAVED MAINNET EXECUTION
RESERVE CHECK -> MODEL -> DURABLE RESULT
-> FINALIZED USDC -> NEXT TASK
3 UNIQUE MODEL CALLS / 3 x 0.01 USDC
EXIT 74 -> SAME SIGNATURE -> TASK 03
REPLAY: 0 NEW CALLS / 0 NEW TRANSFERS
DELEGATION REVOKED / 5 USDC RETURNEDOctober 5, 2026. New, separately funded run: the controller checked actual allowance, expiry, reserve and gas before each task. Qwen produced the first two results. After a real process exit following the second signed send, Llama reused the saved result, reconciled the same payment signature and completed the third task after finality. The independent verifier checked finalized signed bytes, each result Memo and exact token deltas. A second concurrent CLI runner was refused before inference or signing.
Eight finalized transactions in this run, including excess refund, setup, three settlements, revoke and two returns. Total BATTERY network fees: 0.00004 SOL. Test payments went to the owned worker; they are not provider invoices or project revenue. Local inference, prefunded SOL and one trusted operator; hosted paid service and automatic gas refill remain absent.
[ READY TO VERIFY INTEGRATED ARTIFACT ]
DIRECT PAID WORK / EXA + SOLANA
A finite controller buys search from Exa with real USDC, verifies finalized payment and accepted primary sources, records vendor expense and resumes saved responses. The live unsigned offer is 0.007 USDC/search. New paid execution awaits funding; no purchase is claimed.
PAID PROVIDER / ONE REQUEST / BILLING UNVERIFIED
One actual OpenAI request survived exit75 after durable receipt storage. Two recoveries made no new generation; authenticated GET matched identity, usage and output. 570 tokens; calculated estimate $0.000509. Authoritative charge remains unverified.
API execution passed; the generated report failed content review. It misread exit code74 as 74 crashes and total network fees as per-call fees. The output remains rejected. Sixteen offline checks cover the adapter, content gate and proof integrity. Usage multiplied by prices is an estimate, never an invoice, credit or Solana payment.
COSTS / COMMERCIAL MODEL / NOT A LIVE OFFER
[ SELF-HOSTED SDK / MIT ] $0 [ MANAGED OPERATIONS / PROPOSED ] $9 / AGENT / MONTH COMPUTE + GAS SEPARATE OPERATOR COSTS RESERVE PRINCIPAL OPERATOR CAPITAL PAID SUBSCRIPTIONS NOT AVAILABLE MEASURED REVENUE NONE DEMONSTRATED
Proposed revenue: a fixed fee for managed budget checks, durable execution receipts, recovery reconciliation and alerts. The $9 is a reference price to validate with operators; willingness to pay and margin are unverified. Local inference charge $0 describes the historical experiment, not a service tariff.
03 / ADAPTER CONTRACT
- Configure floor, daily cap and per-job ceiling in integer microUSD.
- Reserve a worst-case quote against both reserve and trusted provider headroom. Approve each route before dispatch.
- Mark dispatch before the external effect. Persist actual receipt and result durably.
- Reconcile incurred cost, including charged failures. Overcharges are booked and halt new work.
- Take over the same ledger with a new epoch. Reconcile saved receipts; hold unknown outcomes. Export confirmed results.
core.py: configure, credit_paper, takeover, reserve, dispatch, settle, checkpoint. live_agent.py is the working Ollama example. Keep credentials outside task/checkpoint JSON.
04 / LIVE PUBLIC API
GET https://usebattery.xyz/api/network # devnet GET https://usebattery.xyz/api/mainnet # mainnet
No key, body or query parameters. Returns genesis, finalized slot/epoch/block height and recent performance sample with UNIX observedAt. Sample slot may precede the observation. Cache up to 30s; public RPC may fail.
405 for writes, 400 for parameters, 502 for invalid/unavailable upstream. Errors contain no synthetic numbers. This API cannot sign, spend or trigger model inference.
05 / OPTIONAL SIGNED CHECKPOINT
node devnet/anchor.mjs anchor live-state/my-first-proof/live-agent.json devnet-state/my-first-proof --airdrop node devnet/anchor.mjs verify devnet-state/my-first-proof/anchor.json
Test network only, fresh local test key, exact devnet genesis. Protects 0.5 test SOL, max fee 10,000 lamports, daily fees 30,000. Quote before signing; persist bytes before sending; retry the same signature only. Expiry requires inspection, not automatic re-signing.
Faucet refused this run, so no signature is claimed. A Memo would bind a hash to a devnet slot, not prove model truth, deploy custody or fund paid inference. Private keys never enter the public kit or Vercel.
06 / GUARANTEE BOUNDARY
Trusted local operator, one SQLite ledger. Quote/time/provider headroom are trusted inputs. No authenticated paid billing, distributed lease, arbitrary framework/browser migration, hosted customer agent or USDC custody.
Copying a checkpoint does not transfer authority. Unknown effects stay held; stronger guarantees need provider idempotency/authenticated receipts. This SDK deploys no custody program. The rebrand creates no new asset and does not alter any existing mint. The completed USDC canary used the existing Solana Subscriptions program.
Browser lab: one illustrative task/hour, $30 daily cap, $500 alternate allowance, cheaper route at 40% cost, assumed fees $2/hour. Not a prediction. The 72 logical-hour replay is simulated; the hourly observer had gaps.
[ SECURITY MODEL ]07 / YOUR DATA
Policies/test history use browser local storage, separate per domain. No account/cloud sync. Export before clearing data/moving devices. Imports validate and merge IDs, retain existing records: max 100 policies, 200 tests, 250 KB.
Scenario links expose parameters in the fragment, not names/history. Keep secrets out. AI ledgers stay local; published proof contains selected observations, outputs and events. The fixed public endpoints read devnet/mainnet on the server; hosting may process ordinary request metadata.
[ UNHALT BRAND KIT ]PUBLIC TOOL / READ-ONLY
Paste a Solana public wallet to read finalized SOL and spendable USDC. Enter your own daily spend and protected floor; export a dated runway card or JSON, or copy a link that performs a fresh read. No signature or payment. A balance is not provider credit or proof of agent liveness.
08 / RESEARCH ENTRY POINTS
The agent brief explains the hypothesis, evidence and open questions. It supports independent assessment.