MEAT — Mortality Engineered Auto-Trading

A multi-agent trading system on Alpaca built from Mortality Engineering — the theory in "I Built an AI That Dies If It Stops Making Money". Delivered as an installable DeepSeek Harness plugin: six agent-facing tools' worth of runtime, ten tools, a Web UI dashboard, and a paper/live trading population where the unprofitable die, the profitable reproduce, and every death leaves a record the living read.

Not investment advice. Paper mode is the default; live mode is gated by three independent operator actions. Read docs/ARCHITECTURE.md §9 before pointing it at money.

What is here

docs/ARCHITECTURE.md      theory -> trading mapping, six roles + stoplines, death rules, brakes
docs/CONTRACTS.md         module interfaces, data shapes, config defaults
docs/RESEARCH.md          per-agent isolated hourly research sidecar (news analyst) contract
docs/CONTROL-PANEL.md     operator control panel: research cadence, leaderboards, rewards, plans
docs/RUNBOOK.md           operating guide: phases, commands, emergency procedures
docs/VERIFICATION.md      what was verified, how to reproduce it, what was not
packages/automaton-trading/   the plugin bundle (zero runtime dependencies)
  index.js                Host plugin: 10 automaton_* tools + GET /automaton/state +
                          POST /automaton/control + Harness LLM bridge for research
  client.js               Web UI dashboard (sidebar panel; Settings -> Plugins tab)
  src/                    engine: treasurer, prospector, builder, executor, replicator,
                          external reaper, ledger, registry, brakes, journal, brokers, sim,
                          research (per-agent isolated news analyst)
  bin/automaton.mjs       operator CLI (probe, tick, run, simulate, journal, brakes, stop)
  test/                   node:test suites (mortality, market, plugin, integration, security,
                          research)
.env.example              Alpaca credential template (+ optional DEEPSEEK_API_KEY for CLI research)
automaton-data/           runtime state (gitignored): registry, ledger, obituaries, traces,
                          research/<agentId>.jsonl briefings

Quick start

cd D:\CounterStikers

# 1. credentials (paper) — this build's .env is already populated for the supplied paper account
Copy-Item .env.example .env   # only if .env is missing
notepad .env        # ALPACA_API_KEY_ID=...  ALPACA_API_SECRET_KEY=...

# 2. verify the account (no orders)
node packages/automaton-trading/bin/automaton.mjs probe
node packages/automaton-trading/bin/automaton.mjs config
node packages/automaton-trading/bin/automaton.mjs brakes

# 3. observe selection pressure without touching the broker
node packages/automaton-trading/bin/automaton.mjs simulate --cycles 200 --population 100 --seed 7

# 4. run the population (starts the external Reaper for paper/live)
node packages/automaton-trading/bin/automaton.mjs run

The account must be able to carry the population: population.initial × endowmentUsd must stay inside the 80% allocation cap of broker equity. This build ships automaton.config.json tuned for the supplied $2,000 paper account and the requested population: 30 agents × $20 endowment ($600 allocated, 30%), $1 death floor, 20% drawdown death, $10 child endowment, 40 hypothesis candidates per cycle so every agent can trade its own strategy. Edit that file to scale the economics if the account or population changes (keep the automaton-data/ mirror in sync — the Harness Host process runs with a different working directory).

The population is already seeded: 30 generation-0 agents, each with a distinct deterministic genome (unique symbol × strategy-family pair, unique parameters) and a compiled strategy spec. At the 2026-10-24 paper snapshot: 30 alive / 0 dead / cap 60, $20.00 each, reconciliation ok.

In the Harness chat: automaton_status, automaton_tick, automaton_start, automaton_obituaries, automaton_brakes, automaton_simulate, automaton_config, automaton_agents, automaton_stop, automaton_emergency_stop. Control panel: a sidebar entry (Automaton icon in the left sidebar) that opens the full dashboard in the central column — no floating window; the same dashboard is also a page in Settings → Plugins → MEAT — Mortality Engineered Auto-Trading.

Inside the dashboard: header stats, Start / Stop / Emergency controls, the operator Reaper row, and a clickable table of all agents. Agents have memorable strategy names instead of random ids (Momentum AAPL, MeanRev NVDA, Breakout TSLA …; the atm-xxxxxxxx id stays as the immutable key and secondary label). Each row shows both the inherited Genome strategy and the Trading (spec) actually compiled by the Builder, with a ≠ when they differ. Clicking a row opens a detail pane with current holdings (symbol, qty, avg/last price, market value, unrealized P&L, opened-at) and past activity (newest-first fills with side, qty, price, notional, realized P&L, time), plus strategy parameters, role costs and death info.

The plugin is already installed in this profile: @local/automaton-trading v1.0.0, enabled, application applied. Reconfigure via its row in packages/automaton-trading/cordis.patch.yml (stateDir, envFile, mode, autoStart).

The idea in one paragraph

In a normal agent economy, doing nothing is free, so agents idle. Give an agent a metabolism — an endowment that pays for inference, market data, and hosting, visible as runway hours rather than a balance — and doing nothing becomes the fastest way to die. The death condition lives outside the agent (a separate Reaper process that liquidates and writes the obituary, with grace = 0), so it cannot be argued away. Deaths are structured data; children inherit the parent genome plus exactly one mutation plus the obituary lessons. No role may both decide and act: the Treasurer never trades, the Prospector never sizes, the Builder never chooses the symbol, the Executor never sees account equity, the Replicator never computes surplus, and the Reaper is not reachable from the engine.

Where trading differs from the article: an agent earns return on capital, not service revenue, so risk-of-ruin deaths (drawdown / daily loss) fire long before pure cost-burn would. Both are implemented; the simulation's accelerated preset makes costs dominant so selection is observable in minutes (it is labelled as such in every result and is not the operating config).

Safety

• 18 hard denials enforced in code, before the broker client sees an order: no transfers/ withdrawals/ACH/wires/crypto/account changes (the methods do not exist), off-session trading, symbol/size/spread caps, daily order caps, daily-loss and drawdown deaths, live-mode arming, proof-required trades, replication gates, kill switch, allocation cap, one-way death — plus the v3.0 instrument gates D2/D15–D18: short sales, options structures (never naked short options), covered-call coverage, defined-risk spreads, margin debt and the option expiry guard.
• Paper is default. Multi-instrument trading (margin, short stock, options, spreads) is a paper-mode capability expansion under hard caps; live needs mode=live and allowLive=true and an operator LIVE_ARMED marker (automaton arm-live --confirm "ARM LIVE"). Keys are read only from .env; tool output and the status route are redacted.
• Per-agent isolated research sidecar: each living agent gets its own hourly LLM news briefing about its own book/symbols (Alpaca News API) with a private journal; the briefing can veto entries, shrink size or force exits, and the agent pays for it from its own ledger ($0.001/briefing). Shadow mode via research.influence:false. Contract: docs/RESEARCH.md. In the Harness the model comes from the Host llm service; a headless CLI run uses DEEPSEEK_API_KEY.
• The engine has no control surface over the Reaper. automaton_stop stops the loop only; the Reaper keeps liquidating condemned agents.
• Operator control panel (2026-10-29): the sidebar dashboard can set the research cadence (1h/3h/6h/24h, before open, after close, or both — applied live and persisted to automaton-data/settings.json), configure leaderboards and winner rewards (daily/weekly/ monthly ranking; 1st/2nd/3rd = 5/3/2% of the broker account cash by default, configurable, capped by treasury pool / allocation cap / max per period, paid once per period as virtual treasury→agent ledger transfers — no broker transfer exists), and inspect a richer agent detail page: strategy explanation, plan + next step, triggers, position stop/target/bars-held, ranking badges, P&L and a full transactions table. Contract: docs/CONTROL-PANEL.md.
• Fast intraday profile + two execution tiers: the paper population re-prices during trading hours only (engine.tradingHoursOnly: off-hours the loop does a cheap research/status pass and stops paying market-data/metabolism costs). Bots run in two cadence tiers: day-trading bots are analysed/trade every ~60 s on 1-minute IEX bars, mid/long-term bots hourly on 1-hour bars (cadence.day / cadence.swing); swing positions still get an exit-only risk pass every cycle. Bar-based hold rules are measured in each tier's timeframe. Trade-offs and revert instructions: docs/RUNBOOK.md §8a–8c.
• Per-bot agent brains + strategy categories: every bot is its own isolated agent — it reviews its strategy, book, market and research in a private context on its own cadence (day bots ~5 min, swing bots hourly), keeps its own review journal/policy, and writes decisions the Executor applies (trade/cautious/stand_down, size multiplier, exit pressure). The population is categorised by strategy type: Swing · mid-long (momentum / mean-reversion, 1-hour lane) and Day · intraday (breakout / gap-fade, 1-minute lane); genesis mixes them ~50/50 and existing populations rebalance once on first start. The panel groups agents by category and shows each bot's strategy type, estimated trading interval, and its latest agent thesis/plan. Contract: docs/BOT-AGENTS.md.
• Multi-instrument paper trading (v3.0): bots may use margin, short selling, long calls/puts, covered calls and defined-risk vertical spreads, actively chosen by the bot's agent or its family playbook (option premium sized to 35% of equity / $30 per order, max 2 contracts, DTE 2–14, auto-close at 1 DTE; short stock whole shares within the short cap; covered calls only with 100+ owned shares; naked short options are never allowed). Bots were raised to $50 each by reallocating the unallocated treasury pool (no cash minted; reconciliation conserved) and part of the population trades affordable underlyings (F/SOFI/BAC/INTC) so whole-share shorts and cheap option contracts are usable. New hard brakes D2/D15–D18 gate every order; death liquidates signed stock and option positions before settling the estate. Contract: docs/INSTRUMENTS.md; ops: docs/RUNBOOK.md §8d.
• Trade-activity optimization + diagnostics (v3.1): fixed the stale-runtime-config drift that silently kept the population at $20 bots / 5-minute cadence on 1-hour bars / no trading-hours gate despite the v3 settings (the workspace config is now authoritative and the state copy is migrated); bot agents are explicitly told that long stock is fractional and shown their dollar size budget and real broker capabilities, so a $500 share no longer triggers stand_down; reviews are forced due on restart so old decisions cannot block entries for an hour; strategies only compile specs for the bot's own family+symbol (unrelated-hypothesis fallback removed, cached mismatches cleared, 60 hypotheses/cycle for full universe coverage); and broker capability gating removes short/margin/covered-call/credit-spread permissions automatically while the paper account is below the margin minimum (current account ~$1,999.87 reports shorting/margin disabled; options level 3, long stock and long/debit options remain). The engine header now shows a Last cycle diagnostic (market state, executions, orders, errors, top skip reasons) and a restricted-account warning. Ops: docs/RUNBOOK.md §8e.

Status (this build)

• 10 test suites: mortality 70, market 43, plugin 43, independent integration 34, security 53, research 13, rankings 16, control 17, bot-agent 9 and instruments 7 — 305 tests, 305 pass, 0 fail. Run each file directly, e.g. node test/mortality.test.js; node --test hits a sandbox spawn EPERM.
• Control-panel engine verified in test/control.test.js (research cadence windows + live reconfigure + settings-overlay persistence/rollback; strategy plan/transactions; daily/weekly/ monthly leaderboards; period-boundary rewards with cap scaling, once-per-period and live reward reconfigure) and in test/rankings.test.js (16 tests). Host control actions and every dashboard mapping are covered by the plugin suite (42 tests). Live UI verification follows the next Host restart.
• Research sidecar verified end-to-end: test/research.test.js (isolation, schema clamp, malformed-reply containment, per-agent charging, hourly scheduling, stale expiry) plus scripts/research-smoke.mjs (10/10: real createAutomaton composition, per-agent journals, cadence, executor receives briefings). The executor side is covered by market tests (veto/size/exit/stale-off). Live Harness-LLM verification happens after the next app restart.
• Independent verification found 9 real defects across two rounds (Reaper kills not persisted → resurrectable zombies; dry-run brake always NO_PROOF; unreachable daily-loss death; generation truthiness; stale counters; sub-cent costs rounded away; sim mode never executing death; replication gate ignoring the drawdown rule and thus letting a birth kill the parent; day P&L not persisted). All fixed and re-verified; scripts/verify-fixes.mjs runs 18 checks covering them.
• 30-agent population seeded for the $2,000 paper account: 30 × $20 = $600 allocated (30%), $1 death floor, 20% drawdown death, $10 child endowment, cap 60. Each agent has a distinct deterministic genome (10 symbols × momentum/meanrev/breakout, randomized lookback/entryZ/exitZ/stop/target/hold/size/session) and a persisted compiled spec — 22 distinct proven symbol×family pairs, 30 distinct compiled strategy parameter sets.
• Sidebar control panel is live: verified through the Client slot inspector — sidebar.panellist has an active automaton-trading occupant (order 30) and the keyed main slot has the automaton-trading dashboard panel; the old shell.overlay floating entry is gone. The dashboard shows strategy agent names, genome-vs-spec, and per-agent drill-down with holdings and past activity (30 unique names, tested). Operator controls go through a loopback-only POST /automaton/control (start | stop | emergency-stop | reaper-stop, confirm required where destructive).
• Reaper orphan incident (hardened): a Reaper left over from the previous app generation kept saving a stale registry snapshot every 30 s (wiping agent names). A new Reaper now reloads registry.json/ledger.json before every sweep and writes a heartbeat (reaper.heartbeat.json, v2); engine.start() refuses to reuse a Reaper without a matching heartbeat; the panel shows the Reaper row plus an operator-only two-step Stop Reaper (POST /automaton/control {"action":"reaper-stop","confirm":"STOP"}, implemented in the Harness Host). The engine still exposes no way to stop the Reaper and no agent tool exists for it. If the Host cannot signal that process, run once in your own terminal: Get-CimInstance Win32_Process | Where-Object { $_.CommandLine -like '*reaper-process.js*' } | ForEach-Object { Stop-Process -Id $_.ProcessId -Force }.
• Harness restart required to load the new Host routes (reaper-stop, REAPER_UNAVAILABLE handling) and the fixed engine modules. The client half (sidebar + main dashboard) is already hot-loaded and its slots are active. After the restart the Reaper row can stop the stale orphan, then Start spawns a fresh heartbeat-v2 Reaper: 1) Stop Reaper (if the stale banner shows), 2) Start.
• Plugin installed and live: 10 automaton_* tools registered (verified returning the 30-agent snapshot), dashboard tab registered, kill-switch path verified end-to-end (armed → file → status → cleared).
• Simulated population (100 agents, 200 cycles, seed 7, accelerated preset): 17–25 dead by drawdown, 75–83 alive; reproduction requires a parent that has actually grown, so it is verified deterministically in the test suites and fires in paper/live only on real surplus.
• Paper account connectivity is verified against Alpaca (2026-10-24): probe reads the real account (equity $2,000, ACTIVE, market closed). A full 30-agent cycle consumed 1,172 real hourly bars for 10 symbols, produced 22 dated-proof hypotheses, 30 compiled specs and 30 executions, placed 0 orders, and the broker independently reports 0 orders / 0 positions. The D3 wall denies with MARKET_CLOSED using the broker's own clock (next open Monday 09:30 ET).
• External Reaper verified against the paper account: engine.start() spawns it detached (pid alive), it swept the real account (checked=2 killed=0), and engine.stop() left it running by design; it was then stopped as the operator.
• Paper verification surfaced 3 more real defects, now fixed: the data API returns no bars without an explicit window and pages multi-symbol requests one symbol at a time (broker now requests a bounded window per symbol); D3 only checked the weekday time window, so a Saturday signal could have been submitted (the broker clock is now forwarded into the brake chain); weekend IEX spreads are gross so the paper spread ceiling is 100 bps.
• Harness restart required to load the new Host routes and the fixed engine modules (the plugin's module graph is cached); CLI and test runs already use the fixed code.
• Still unverified: actual paper fills and live trading. The market was closed for the whole build, so no order could legitimately fill; live mode was never armed. Control-panel rendering is verified structurally (active slot occupant) but not visually.
• The paper keys were supplied in chat and live only in the gitignored .env; they are not in the package, the profile config, tool output, or the status route. Rotate them in the Alpaca dashboard if you consider the chat exposure a risk.