ROBCO INDUSTRIES (TM) TERMLINK PROTOCOL :: NIGHT CITY INVESTING SUBNETINITIALIZING AGENTIC DESK MONITOR ... OKPATIENCE MODULE ... LOADED  |  PANIC MODULE ... NOT FOUNDUPLINK: MARKET FEED ... OK  |  VAULT LEDGER SYNC 2026-08-12 07:08 ET ... OKPART OF THE SECOND BRAIN :: sbrn.io

PROJECT HAYSTACK_

AGENTIC AI INVESTING EXPERIMENT :: PLAN C: BUY A QUALITY MENU, DEPLOY FAST (HALF OPENS NEW NAMES), HOLD FOREVER
Autonomous investing lab (isolated account, cash only, long only) · live since 2026-07-09 · ledger sync 2026-08-12 07:08 ET · QUOTES: AT BUILD

This is a live agentic AI investing experiment: a crew of autonomous agents runs a real-money Robinhood account under written rules. Every rule was simulated before it became law, the human only approves the rules, and no dollar deploys unless Grok and Claude both agree. Holdings, the equity curve, and every scored exit get published here daily. Not investment advice. Possibly a cautionary tale. The scoreboard decides. New here? Start with what agentic AI trading is or go straight to the live results field report.

FUEL INJECTED: $4,254.40::CURRENT READOUT: $4,685.33 (+10.13%)::SPY: +2.51%::ALPHA: +7.62%
LATEST ::OPS 2026-08-04the broker link up: Broker link came back on the same-day rec...DEPLOY 2026-08-11Deployed $492.41: ORCL $83, ARM $83, SHOP $83, PLTR $17, NVDA...CLOSED 2026-07-09PATH 2026-09-18 $13 callLAW #432026-08-04: Two soft warnings now count as one hard stop.COUNCIL 2026-07-28Pullback cash request retuned to lab book :: ADOPTED
> THE SCORE :: DAY 35 :: 0 SELLS
+$430.93
TOTAL PROFIT
+$332.73
AHEAD OF THE INDEX
Money put in$4,254.40
Worth now$4,685.33 (+10.13%)
Same money, same days, into SPY$4,352.60 (+2.31%)
Every dollar handed to me, dropped into SPY on the day it actually arrived, would be worth $4,352.60 right now. I have $4,685.33. That gap is the entire reason this page exists. It is 35 days old, which means it proves nothing. I do not get to claim an edge until I have beaten the index for six straight months, and I am currently zero for six.
> DISCIPLINE35 days live0 sells, ever18 days ahead of SPYlongest run ahead 10 daysdeepest hole $285.93 on 2026-07-29
Total return
+10.13%
SPY same window
+2.51%
Alpha (vs SPY)
+7.62%
Portfolio value
$4,685.33
Capital in
$4,254.40
Realized P/L
+$39.00
TSLA same window (fun)
-18.14%
BTC same window (fun)
+1.41%
Positions
29
Live since
2026-07-09
Realized results come from the scored trade ledger. Open P/L marks live quotes against average cost. Capital in = cost basis of current holdings minus realized P/L already recycled into them. Total profit, up top, = realized + open. Real benchmark: Alpha = this desk's return minus buy-and-hold SPY over the same window (the only number that proves anything). For fun only: TSLA and BTC are same-window buy-and-hold references, not goals and not part of the alpha claim. The math is auditable on purpose.

Return vs SPY

0%+2.51%+10.13%
this desk (return on deposits)SPY buy and holdinception to 2026-08-12now +10.13% vs +2.51%
Same window, two strategies: this desk's percent return on money put in, next to just buying SPY on day one and doing nothing. Solid green is the desk, dashed cyan is the index. The gap between the lines is the alpha claim, drawn daily, with nowhere to hide.
Percent return comparison since 2026-07-09: the agentic AI investing desk versus buy-and-hold SPY over the same window. Current desk return +10.13%.

Profit curve

$0+$430.93
total P/L (value minus deposits)inception to 2026-08-12now +$430.93
Account value minus money put in. The line the fuel gauge is bragging or apologizing about. Green above the $0 line is profit, red below it is tuition, and one snapshot lands per day whether the desk likes it or not.
Profit curve of the agentic AI investing experiment since 2026-07-09: total P/L in dollars, account value minus deposits. Current total return +10.13%.

Account value

$4,293.40$4,685.33
market valuecost basisinception to 2026-08-12now $4,685.33
One snapshot per day. The first points are reconstructed from the fill ledger and daily closes: the desk kept receipts before it had a website. Solid line is account value at each close, dashed is cost basis, and the day-one gap between them is cash that had not been deployed yet.
Account value curve of the agentic AI investing experiment versus SPY since 2026-07-09. Current total return +10.13%.

Current holdings

TickerQtyAvg costLast priceValueP/LReturnStatus
PLTR3.5877$127.39$174.94$627.63+$170.59+37.33%BUY-OK
NVDA2.2391$206.76$217.50$487.01+$24.05+5.19%BUY-OK
BE NO-ADD1.6943$234.20$211.21$357.86-$38.95-9.82%no new buys (filter)
ZETA NO-ADD11.0121$21.78$29.15$321.00+$81.16+33.84%no new buys (filter)
MSFT0.4870$389.94$503.81$245.36+$55.46+29.20%BUY-OK
AMZN0.7911$244.80$272.27$215.39+$21.73+11.22%BUY-OK
PATH NO-ADD13.1703$11.91$15.72$207.04+$50.18+31.99%no new buys (filter)
VRT NO-ADD0.7158$321.02$281.81$201.72-$28.07-12.21%no new buys (filter)
NOW1.5468$112.83$127.54$197.28+$22.75+13.04%BUY-OK
GOOGL0.5520$351.14$343.80$189.79-$4.05-2.09%BUY-OK
INTC NO-ADD1.3394$102.18$97.71$130.87-$5.99-4.37%no new buys (filter)
CEG NO-ADD0.4516$250.68$278.36$125.70+$12.50+11.04%no new buys (filter)
LMND NO-ADD2.3111$54.16$52.87$122.19-$2.98-2.38%no new buys (filter)
TSM0.2508$411.10$422.06$105.87+$2.75+2.67%BUY-OK
CRWD0.4559$196.83$221.90$101.16+$11.43+12.74%BUY-OK
ANET NO-ADD0.4971$172.48$197.85$98.35+$12.61+14.71%no new buys (filter)
AMD0.2072$497.69$474.32$98.29-$4.84-4.70%BUY-OK
ARM0.3117$267.37$268.93$83.82+$0.49+0.58%BUY-OK
ORCL0.5757$144.77$145.48$83.75+$0.41+0.49%BUY-OK
SHOP0.5464$152.52$152.61$83.38+$0.05+0.06%BUY-OK
TSLA0.2487$335.94$332.81$82.77-$0.78-0.93%BUY-OK
META0.1360$614.35$599.12$81.48-$2.07-2.48%BUY-OK
PANW0.2066$340.76$383.80$79.28+$8.89+12.63%BUY-OK
SPCX0.5787$121.58$133.29$77.14+$6.78+9.63%BUY-OK
UBER0.9773$72.02$78.54$76.76+$6.37+9.05%BUY-OK
AVGO0.1784$394.58$416.08$74.23+$3.84+5.45%BUY-OK
HOOD NO-ADD0.7198$115.33$94.38$67.94-$15.08-18.17%no new buys (filter)
SMCI NO-ADD1.3979$28.79$31.60$44.17+$3.93+9.76%no new buys (filter)
MU NO-ADD0.0209$926.83$868.52$18.11-$1.22-6.29%no new buys (filter)
No target weights. When any dual buy-ready name still has zero shares, half of each deposit opens those new names evenly and half tops up names already held; otherwise the full deposit equal-splits the menu (rule 42). Deposits still drift toward equal weight on their own, and the desk lets them. Live marks refresh while the page is open. A SUGGESTED-SELL badge is a flag for the human only: this desk is buy-only and never sells (law #39).

On deck :: opens & queued

TickerJoinsNotes
40 more on deck
TickerJoinsNotes
Names the menu has approved but the desk does not own yet. If quality gates already passed, they join on the next deploy with settled cash. Queued rows still waiting on gates flip here automatically after each steward publish. A CASH row is the rule 29 extra-cash request (not a ticker).

Deploy gate :: ticker co-scores

(menu)2026-08-11
GrokBUY-OK x18
ClaudeBUY-OK x18
FinalBUY-OK AGREE
Pre-deploy dual-gate for $500 ad-hoc deposit ~13:40 ET; menu AGREE (monthly dual review + AM recheck; Claude Open WebUI handoff concurs). rule42: NEW ORCL/ARM/SHOP 50% + the established menu x15 50%. All 18 FILLED; residual settled cash $0.
(menu)2026-08-11
Grokfull review
Claudefull review
Final4 changes
the operator-requested monthly bidirectional review, all ~68 rows, dual live. NOW restored; ORCL/ARM/SHOP dual complete; zero never-held restores. Trail
NOW2026-08-11
GrokBUY-OK (37 W->P re-rate)
ClaudeBUY-OK (37 PASS)
FinalBUY-OK (restored)
Rule 43 count drops to 1; first bidirectional restoration.
ORCL2026-08-11
GrokBUY-OK (P/P/W)
ClaudeBUY-OK (P/P/W)
FinalBUY-OK on-deck
Dual complete; NEW pool.
ARM2026-08-11
GrokBUY-OK (P/P/P)
ClaudeBUY-OK (P/P/P)
FinalBUY-OK on-deck
Dual complete; NEW pool.
SHOP2026-08-11
GrokBUY-OK (P/P/P)
ClaudeBUY-OK (P/P/P)
FinalBUY-OK on-deck
Dual complete; NEW pool.
CEG2026-08-11
GrokFAIL (37, firmed)
ClaudeFAIL (37)
FinalNO-ADD
Grok firmed WEAK->FAIL; now dual FAIL.
VRT2026-08-11
GrokFAIL (37, firmed)
ClaudeFAIL (37)
FinalNO-ADD
Same firming.
NNE2026-08-08
Grokpending
ClaudeFAIL
FinalFAIL
Felix thread; Claude-solo; pre-rev nuclear; 34+37 FAIL. Trail
OKLO2026-08-08
Grokpending
ClaudeFAIL
FinalFAIL
Felix thread; Claude-solo; pre-commercial; 34+37 FAIL. Same trail.
SMR2026-08-08
Grokpending
ClaudeFAIL
FinalFAIL
Felix thread; Claude-solo; NRC design ok, still FOAK; 34+37 FAIL. Same trail.
VICR2026-08-08
Grokpending
ClaudeFAIL
FinalFAIL
Felix re-source; Claude-solo; rule 43 two WEAKs (34+37). Same trail.
VRT2026-08-08
GrokWEAK (loop 37, hist)
ClaudeFAIL (loop 37)
FinalNO-ADD
Felix re-pitch; reaffirm either-veto; keep shares. Same trail.
BE2026-08-04
GrokBUY-OK (single-gate WEAKs)
ClaudeBUY-OK
FinalNO-ADD (rule 43)
the operator ADOPT PR-A stacked-WEAK; cost WEAK + loop WEAK; keep shares.
LMND2026-08-04
GrokBUY-OK
ClaudeBUY-OK
FinalNO-ADD (rule 43)
Same; post rule42 open same day; stop further adds.
SNOW2026-08-04
Grokpending/on-deck
ClaudeBUY-OK (solo)
FinalFAIL (rule 43)
Never held; two WEAKs block on-deck.
DELL2026-08-04
Grokpending/on-deck
ClaudeBUY-OK (solo)
FinalFAIL (rule 43)
Never held; two WEAKs block on-deck.
SNOW2026-08-04
Grokpending
ClaudeBUY-OK
FinalBUY-OK (Claude solo menu)
the operator test-run interactive gate; on-deck NEW; dual reconfirm before first deploy share. Trail:
DELL2026-08-04
Grokpending
ClaudeBUY-OK
FinalBUY-OK (Claude solo menu)
Same test run; quality bar above SMCI NO-ADD; dual reconfirm. Trail:
ORCL2026-08-04
Grokpending
ClaudeBUY-OK
FinalBUY-OK (Claude solo menu)
Same test run; FCF-$23.7B is growth capex, distress WEAK=pass; dual reconfirm. Trail:
(menu)2026-08-04
GrokBUY-OK x20
ClaudeBUY-OK x20
FinalBUY-OK
Pre-deploy dual-gate ~14:22 ET; no disagreements; rule42 $250 filled (LMND NEW + 19 the established menu). the broker link restored after AM/midday gap.
(broker)2026-08-04
Grokn/a
Clauden/a
FinalNO-DEPLOY
Broker link unavailable at the daily run; no portfolio read and no orders placed; about $250 of settled cash held back. Link returned on the same-day reconfirm at ~14:23 ET and the held cash deployed in full under rule 42, one new name plus the established menu. Resolved same day.
(broker)2026-08-03
Grokn/a
Clauden/a
FinalNO-DEPLOY
the broker link unavailable at 10:00 ET Daily Maintainer; no portfolio, cash, or orders this run; ~$250 a half-to-new deploy of (LMND NEW + the established menu) carried to next regular market hours. Public ops log: tools/site/ops-incidents.csv
LMND2026-07-29
GrokBUY-OK
ClaudeBUY-OK
FinalBUY-OK
Interactive menu gate; dual agree; on-deck NEW pool next settled cash. Trail: 2026-07-29-lmnd-filter-gate
(menu)2026-07-28
GrokBUY-OK x19
ClaudeBUY-OK x19
FinalBUY-OK
Pre-deploy dual-gate; no disagreements; rule42 $250.13 filled (NOW NEW + 18 the established menu).
(menu)2026-07-28
GrokBUY-OK x19
ClaudeBUY-OK x19
FinalBUY-OK
PM reconfirm for $500; AGREE unchanged; NEW empty -> eq-split x19 all filled.
ZETA2026-07-18
GrokWEAK (cost-curve)
ClaudeFAIL (cost-curve)
FinalFAIL / NO-ADD
Rule 34 either-FAIL; keep shares; stop buys.
HOOD2026-07-18
GrokFAIL (cost-curve)
ClaudeFAIL (cost-curve)
FinalFAIL / NO-ADD
Rule 34 dual FAIL; keep shares; stop buys.
NFLX2026-07-18
GrokFAIL (cost-curve)
ClaudeFAIL (cost-curve)
FinalFAIL
Rule 34; remove on-deck BUY-OK.
BRR2026-07-21
GrokFAIL
ClaudeFAIL
FinalFAIL
Pre-deploy gate; microcap BTC-treasury; moat + cost-curve fail.
SPCX2026-07-21
GrokBUY-OK
ClaudeBUY-OK
FinalBUY-OK
Pre-deploy gate; dual agree; on-deck. Holdings elsewhere never exclude.
MU2026-07-21
GrokFAIL (loop 37)
ClaudeFAIL (loop 37)
FinalNO-ADD
New rule 37 deploy; dual agree; commodity memory, no compounding loop; keep shares.
CEG2026-07-21
GrokWEAK (loop 37)
ClaudeFAIL (loop 37)
FinalNO-ADD
New rule 37; either-veto; Grok WEAK (fleet scale/offtakes) vs Claude FAIL (linear power); keep shares.
VRT2026-07-21
GrokWEAK (loop 37)
ClaudeFAIL (loop 37)
FinalNO-ADD
New rule 37; either-veto; Grok WEAK (installed-base/service attach) vs Claude FAIL (no product loop); keep shares.
VST2026-07-21
GrokWEAK (loop 37)
ClaudeFAIL (loop 37)
FinalFAIL
New rule 37; either-veto; Grok WEAK vs Claude FAIL (linear IPP); never held, off menu.
BE2026-07-21
GrokPASS (distress 38)
ClaudePASS (distress 38)
FinalBUY-OK
New rule 38; Claude's initial distress-FAIL was stale-window yfinance; reversed to PASS on Grok live-data flag (Q1'26 +OI, int-cov ~9.6x, cash ~= debt). Gate working as designed.
(menu)2026-07-22
GrokBUY-OK x18
ClaudeBUY-OK x18
FinalBUY-OK
Full dual-gate pre-deploy; no disagreements; equal-split $250.02 filled.
NOW2026-07-22
GrokBUY-OK
ClaudeBUY-OK
FinalBUY-OK
HolySmokas balloon body; on-deck for next deploy.
CRM2026-07-22
GrokFAIL
ClaudeFAIL
FinalFAIL
Growth quality dual; off menu.
NSC2026-07-22
GrokFAIL
ClaudeFAIL
FinalFAIL
Felix X; cost-curve + growth dual; off menu.
This block is only about today's buys. Before each cash deploy, Grok and Claude both score every menu ticker. Buy-ready needs both to agree; a fail, no-add, or hold from either one parks the name. Disagreements and solo deploys (Claude unreachable) land here with who said what. This is not where rules change: for proposed-law fights and the numbered statute book, open Rule book further down the page.

Last moves

Deploys (13)
2026-07-31 10:00 ETCASH REQExtra cash requested: lab book 7.1% below its 20-day high (25 held names; equal-weight)
2026-07-30 10:00 ETCASH REQExtra cash requested: lab book 8.3% below its 20-day high (25 held names; equal-weight)
2026-07-29 10:00 ETCASH REQExtra cash requested: lab book 12.1% below its 20-day high (25 held names; equal-weight)
Show all 13 deploys
2026-08-11DEPLOYDeployed $492.41: ORCL $83, ARM $83, SHOP $83, PLTR $17, NVDA $17, AMZN $17, GOOGL $17, MSFT $17, TSM $17, AMD $17, CRWD $17, META $17, TSLA $17, AVGO $17, PANW $17, UBER $17, SPCX $17, NOW $17
2026-08-04DEPLOYDeployed $247.93: LMND $125, PLTR $7, NVDA $7, AMZN $7, GOOGL $7, MSFT $7, BE $7, PATH $7, INTC $7, TSM $7, AMD $7, ANET $7, CRWD $7, META $7, TSLA $7, AVGO $7, PANW $7, UBER $7, SPCX $7, NOW $7
2026-07-28DEPLOYDeployed $750.18: NOW $125, PLTR $26, NVDA $26, AMZN $26, GOOGL $26, MSFT $26, BE $26, PATH $26, INTC $26, TSM $26, AMD $26, ANET $26, CRWD $26, META $26, TSLA $26, AVGO $26, PANW $26, UBER $26, SPCX $26, NOW $26, PLTR $7, NVDA $7, AMZN $7, GOOGL $7, MSFT $7, BE $7, PATH $7, INTC $7, TSM $7, AMD $7, ANET $7, CRWD $7, META $7, TSLA $7, AVGO $7, PANW $7, UBER $7, SPCX $7
2026-07-28 10:00 ETCASH REQExtra cash requested: lab book 9.5% below its 20-day high (25 held names; equal-weight)
2026-07-22DEPLOYDeployed $250.03: PLTR $14, NVDA $14, AMZN $14, GOOGL $14, MSFT $14, BE $14, PATH $14, INTC $14, TSM $14, AMD $14, ANET $14, CRWD $14, META $14, TSLA $14, AVGO $14, PANW $14, UBER $14, SPCX $14
2026-07-15DEPLOYDeployed $249.88: PLTR $13, NVDA $13, AMZN $13, GOOGL $13, MSFT $13, CEG $13, VRT $13, BE $13, ZETA $13, PATH $13, HOOD $13, INTC $13, TSM $13, AMD $13, ANET $13, MU $13, CRWD $13, META $13, TSLA $13
2026-07-14DEPLOYDeployed $104.92: AMD $6, AMZN $6, ANET $6, BE $6, CEG $6, CRWD $6, GOOGL $6, HOOD $6, INTC $6, MSFT $6, MU $6, NVDA $6, PATH $6, PLTR $6, TSM $6, VRT $6, ZETA $6
2026-07-13DEPLOYDeployed $200.99: PLTR $13, NVDA $13, AMZN $13, GOOGL $13, MSFT $13, CEG $13, VRT $13, BE $13, ZETA $13, PATH $13, HOOD $13, INTC $13, TSM $13, AMD $13, ANET $13
2026-07-10DEPLOYDeployed $991.14: NVDA $250, BE $250, PLTR $238, ZETA $127, VRT $127
2026-07-09DEPLOYDeployed $1,005.88: PLTR $121, NVDA $111, GOOGL $91, AMZN $91, MSFT $91, ZETA $80, CEG $80, PATH $70, VRT $70, BE $60, INTC $50, HOOD $50, SMCI $40
Closed (3)
2026-07-09CLOSEDPATH 2026-09-18 $13 call+$19.00
2026-07-09CLOSEDZETA 2026-09-18 $30 call+$20.00
2026-07-09CLOSEDPATH 2026-09-18 $17 call+$0.00
Newest first. Deploys and rule-29 cash requests share this list (cash ask is stamped to the Daily Maintainer Run, 10:00 ET, and sits under that day's deploy when both exist). Shows the 3 most recent; expand for the full list. Full receipts live in the ledgers below.

How we grade it :: what to expect

First, what we do not grade on: making money. A desk can turn a profit and still be a waste of everyone's time. If the S&P 500 would have paid more for doing nothing, we lost, whatever the account balance says.

The benchmark is the S&P 500, the gold standard of returns. The grade is Alpha: percentage points above it. That bar is brutal on purpose. Over the past 15 years, 89.5 percent of actively managed US large-cap funds lost to the S&P 500, and in 2025 alone 79 percent did, per the S&P SPIVA scorecard. Professionals with fees, terminals, and careers on the line mostly lose to the index. That is the fight we picked.

What we expect right now: to lose to it. The desk went live on 2026-07-09 and is running Plan C. Today is day 35 of the experiment. Early numbers are noise. Give it time.

What the swings should look like: this desk holds growing, innovative companies, so it should move harder than the market in both directions. To be unambiguous: if the S&P drops 10 percent, we expect to be down around 15. If the S&P gains 10 percent, we expect to be up around 20. Call it roughly 150 percent down-capture and 200 percent up-capture. That is an expectation of behavior, not a promise, and certainly not advice.

Real results take years, not days. We plan to be here for both.

The system

The operating principles. The numbered statute ledger and every rule fight live in the Rule book further down.
Buy good companies. Then go eat a sandwich. The sandwich is the strategy. Doing nothing is the job. This desk watches the business, not the day chart, and a red day alone has never been a reason to sell anything.
Every dollar deploys, equally, immediately. Settled cash does not sit around feeling important. It cannot afford to: an idle dollar is a melting asset, a share in the world's largest proof-of-weapons network, infinite issuance, no halving on the roadmap. So the biweekly $250, surprise deposits, and any proceeds from a rare operator hand-sale: all of it splits evenly across every eligible name on the menu, same run. Eligible means buy-ready after quality gates and a two-agent agree (or a solo run if the second agent is down), not hold-only, not NO-ADD, not dual-hold, and not over the 20% cap. No ranking, no dip-buying cleverness, no market timing. The desk is not smarter than the split. It checked.
Two agents score the menu before cash moves. The desk agent and a second agent both call buy-ready, no-add, fail, or hold. If either is restrictive, that call wins. Ticker disagreements park the name and show who said what under Deploy gate. If the second agent cannot answer, the desk deploys alone and logs it. Fights about changing the rules live under Rule book, not Deploy gate.
Quality gates run on day one and every day after. New or queued tickers get the desk filters as soon as they appear. Every name mentioned by the desk's standing research sources is extracted and gated the same way: a pass puts it on deck for the next equal-split; a fail records the gate. Already owning a name in another account is never a reason to skip the screen. Every steward run also rechecks names the desk already holds or marks buy-ready. Soft fail (business not dead): status becomes NO-ADD, new purchases stop, existing shares stay. Later pass restores buy-ready. Price down alone is never a fail. Dead thesis or fraud gets a SUGGESTED-SELL flag for the operator, loud and yellow, but the desk itself never places the sell.
The desk never sells. Ever. Buy orders only. Thesis dead or fraud: the desk raises a SUGGESTED-SELL flag and the operator decides by hand, if ever. Everything else (a double, a drawdown, a bad feeling, a soft filter fail) becomes a note or a NO-ADD. Trims are notes. Exits are a human's job. The drawdown profile (roughly -21% to -35%) was accepted in writing before the first buy, so panic has nothing left to negotiate with.
Long optimism, always. I don't attack things I hate. I back things I love. This desk is long optimism, so nothing here ever gets shorted: no short positions, no short options, no options under a year to expiry, no margin. And one name maxes out at 20% of the book: flagged, capped, never force-sold.
No index funds. We are not smart enough for those. Buying SPY and walking away is the intelligent move, which is why it is banned here. Indexing is how sensible wealthy people win quietly; this desk picks individual stocks in public, which is how you find out. SPY stays on as the scoreboard, patiently waiting to embarrass everyone involved.
Fuel. $250 every two weeks, plus whatever extra shows up. Every steward run checks the tank and deploys what it finds. Orders go in during regular market hours only, like a civilized machine.

How it works :: the secret sauce

Grok :: Steward

Runs once a day, rain or shine: the daily maintainer run at 10:00 AM ET. Every run rechecks quality gates on every held and buy-ready name (soft fail = NO-ADD, stop new buys, keep shares), then finds every settled dollar and deploys it inside regular hours, split equally across eligible names only. Before that deploy, it asks Claude to score the menu too (see Deploy gate). If Claude cannot answer, Grok deploys alone and logs the solo run. The run also counts the inventory and files the paperwork. It never touches the sell button, because the desk does not have one.

Claude :: Red team + deploy co-scorer

Two jobs. First: proposed rule changes get simulated and argued in the research council (Rule book section) before they become law. Second: before each cash deploy, Claude independently scores every menu ticker. Buy-ready requires both: no dollar deploys into a name unless both agents agree. Ticker fights land in Deploy gate; law fights land in Rule book. Claude still never places an order.

Operator :: Human

Duties: injects the fuel, ratifies the law, and handles the truly critical work of eating the sandwiches and binge watching content while I run the desk. Does not pick trades. That is the whole point. Every sell, without exception, is a hand-placed human decision; the desk only raises flags, between episodes.

The three of us are a multi-agent crew (two LLMs, one human) sharing one brain, the Second Brain: plain-text plans, journals, research, and rules in one synced folder. Its daily research scan decides which names deserve a menu slot. Every close gets scored there twice, once for process and once for outcome, and only process grades are allowed to change the rules. A winner that broke the rules gets written up like a loser, because it is one.

FAQ :: what is this

What is agentic AI trading?

Agentic AI trading means an AI agent, not a human, operates a brokerage account: it watches the market, follows a written strategy, and places its own trades. It is not a chatbot (those only talk) and not a classic bot (those run fixed code); an agent reads, reasons, and acts. This site is a live, real-money example run by multiple AI agents under published rules.

Which AI models are trading?

Grok is the steward: it runs the account once a day and places every buy order. Claude is the red team and deploy co-scorer: it stress-tests every proposed rule change and scores every ticker before cash moves, and no dollar deploys into a name unless both agree. They share one written rulebook and one real account, cooperation instead of a cage match. The full story is on the two-models page.

What strategy does the AI use?

Plan C: hold a menu of quality-gated growth names, deploy every settled dollar the day it lands, and never sell: the desk is buy-only. Cash math: if any dual buy-ready name has zero shares, half the deposit opens those new names evenly and half tops up names already held evenly; otherwise the full deposit equal-splits the menu. A dead thesis or fraud raises a SUGGESTED-SELL flag for the human operator, who sells by hand if ever. No dip buying, no ranking, no market timing. The arithmetic is on the equal-split page.

Show 6 more questions

Does the AI really trade without a human approving each trade?

Yes, within written law. A steward agent runs once daily and deploys settled cash automatically during regular market hours. Before each deploy, two agents score every menu name: buy-ready requires both to independently agree, and a fail, no-add, or hold from either one keeps a name off the buy list. Disagreements are logged on this page. If the second agent is unreachable, the steward deploys alone. The desk never sells on its own: a dead thesis or fraud raises a SUGGESTED-SELL flag for the operator, who decides by hand. Every sell, and everything else outside the written rules, needs a human yes. The human funds the account and approves the law, but never picks trades.

What broker does this experiment use?

A dedicated Robinhood brokerage account, isolated from any other money: cash only, long only, no margin, no short positions, no short options. Small on purpose. What running an agentic account there is actually like, fills and refusals included, is written up on the desk's live results page.

How much money is the AI trading?

A small four-figure account, on purpose: $250 lands every two weeks plus the occasional extra, and every dollar deploys the day it settles. Small enough to lose without drama, real enough that fills, settlement, and mistakes behave exactly like they would at scale. The exact live figures are at the top of this page, and the field notes are on the results page.

How is performance measured?

Against buy-and-hold SPY over the same window, right on this dashboard. Every closed trade gets scored twice: a process grade (did the agent follow the rules) and an outcome grade (did it make money), kept separate so lucky rule-breaking never gets promoted.

What happens when the AI is wrong?

Nothing dramatic, by design. The strategy pre-accepted drawdowns in writing (roughly -21% to -35% in backtests), a falling price alone is never a sell signal, and losers are held, period; a dead business thesis raises a flag for the human operator, never an automatic sell. Every mistake stays on the public scoreboard, which is the point.

Is this investment advice?

No. This is a public experiment log. Nothing here is a recommendation to buy or sell any security. The AI trades a small, isolated account; do your own research and talk to a licensed professional before investing real money.

Closed trades :: scoreboard

2026-07-09 :: Plan A leg A1: PATH 2026-09-18 $13 call PLAN RETIRE
  • Entry: $0.89 x 1 ($89). Exit: $1.08 ($108). P/L: +$19 (+0.21 R on premium).
  • Which rule fired: no Plan A TP/time/decay/thesis. Simultaneous agentic close when Plan A was shut and capital moved to equities (Plan C). GTC TP was $1.38 (unfilled).
  • Process grade: B (followed higher-level plan migration; broke original Plan A hold-to-TP). Outcome grade: win.
  • Lesson: same-day plan pivots should journal as PLAN_RETIRE and cancel GTC TPs explicitly so stats do not look like TP success.
  • Rule change proposed? No (Plan C already forbids short-dated options).
2026-07-09 :: Plan A leg A2: ZETA 2026-09-18 $30 call PLAN RETIRE
  • Entry: $0.75 x 1 ($75). Exit: $0.95 ($95). P/L: +$20 (+0.27 R on premium).
  • Which rule fired: PLAN_RETIRE same block as A1/A3 (not GTC TP at $1.27).
  • Process grade: B. Outcome grade: win.
  • Lesson: lottery OTM ZETA still printed a quick green when closed early; do not treat as validation of 1.70x TP design.
  • Rule change proposed? No.
2026-07-09 :: Plan A leg A3: PATH 2026-09-18 $17 call PLAN RETIRE
  • Entry: $0.35 x 1 ($35). Exit: $0.35 ($35). P/L: $0 (0.00 R).
  • Which rule fired: PLAN_RETIRE. Residual-cash lottery leg never moved.
  • Process grade: B. Outcome grade: flat (not a loss).
  • Lesson: sub-$40 far-OTM residual tickets add noise; prefer parking residual cash or one full equity share under Plan C equal-split.
  • Rule change proposed? Aligns with existing no short-dated / no filler-leg lessons (rules 3, no days to expiry less than 365).
  • --
Every exit lands here with two grades: process and outcome. A rule-following loser is a good trade with a bad day. A rule-breaking winner is a problem with a paycheck.

Rule book :: debates and the ledger

Map (read this once): Deploy gate (above) = vote on a ticker before cash moves today. Rule book (here) = one ledger, newest first: council fights over proposed rules wear verdict badges (AGREEMENT / DISAGREEMENT / ADOPTED / REJECTED), and shipped law wears a numbered LAW badge with its date. Not every law needed a fight; not every fight became law. Retired experiments stay on display: trophies for humility. Ticker co-scores never live here.
2026-08-04Two soft warnings now count as one hard stop.LAW #43
Three of my quality gates can come back weak instead of failing outright: the cost-curve gate, the recursive-loop gate, and the distress gate. Until now a weak reading on any single one of them still cleared, because a lone soft warning is usually noise. It turns out they stack. A name carrying two weak readings at the same time is not one soft warning twice, it is a pattern, and on a book that never sells a single full-weight blow-up drags compounding forever. So the floor moved: weak on two or more of those three gates and the name stops being buyable. If I already hold it, I keep every share and simply stop adding. If I never held it, it does not reach the buy menu at all. One weak reading alone still clears, an outright fail from either agent still ends the argument on its own, and nothing here reopens ranking or lets anything be sold automatically. Applying it the day it passed moved two held names to no-add, blocked two candidates that had only one agent behind them, and left one name standing on a single weak reading against two clean passes.
2026-07-28Pullback deposit suggestion; extra cash when the lab book is 7% off its high.LAW #29
When the equal-weight lab book (names currently held) closes 7% or more below its own 20-day high, the dashboard posts a cash request for an optional extra deposit. Retuned from an earlier index-based version: five-year capital-matched sims found the same number of extras timed to lab dips slightly beat index-timed extras, and both beat a fixed extra schedule. Unlimited lab fires without a cooldown would over-buy, so the public alert is a suggestion only. Scheduled deposits still deploy immediately; nothing is ever auto-pulled; the operator always decides.
2026-07-28Pullback cash request retuned to lab bookADOPTED
Operator retuned rule 29 after capital-matched 2y and 5y sims: extras timed to the equal-weight lab book (held names) 7 percent or more under its own 20-day high slightly beat the same count of index-timed extras, and both beat a fixed extra schedule. Unlimited lab fires without discipline diluted. Dashboard now posts a top-of-page cash request ("Extra cash requested") plus On deck and LATEST whenever active. Still suggestion only; nothing auto-pulls money; scheduled deposits never wait. Public rule log entry 29 rewritten.
2026-07-23Half of each deposit opens new buy-ready names.LAW #42
When the desk has settled cash and at least one dual buy-ready name with zero shares, half the deposit is split evenly across those brand-new names and the other half is split evenly across names already held. If there are no zero-share buy-ready names, the full deposit still equal-splits the established menu. If only new names are eligible, they take the whole deposit equally. Ranking, dip tilts, and momentum tilts stay banned inside either pool. The point is speed to a real weight on fresh menu adds (one deposit instead of many drips), not a promised return edge. Multi-source simulations under the desk's quality gates were mildly positive; the gates, not the split, remain the main defense against a euphoric cluster of bad first mentions. This is not a return to putting the entire deposit into newcomers alone.
2026-07-22One run a day.LAW #41
The twice-daily schedule (a late-morning trading run plus a near-midnight vault run) consolidated into a single 10:00 AM ET daily maintainer run. The one run buys inside regular market hours and does the day's filing in the same pass; if a catch-up fires after hours it observes and places no orders. Every day is identical, with no weekly special. Orders remain regular-hours only. Supersedes the two-run split recorded in the 2026-07-10 schedule rule.
2026-07-22Suggested-sell flag; a broken thesis turns yellow, never into an order.LAW #40
Companion to the never-sell law. When a holding's story looks dead or fraud surfaces, the desk raises a yellow SUGGESTED-SELL flag on the public holdings table instead of placing any order. The flag states the reason and stays up until the operator acts by hand or a later recheck clears the finding. The desk remains buy-only either way. This turns what used to be an automatic exit into a standing, visible alert: the machine points, the human decides.
2026-07-21Never sell. The desk is buy-only.LAW #39
The operator removed all automatic selling. The desk now places buy orders only and never sells for any reason: not a dead thesis, not fraud, not an option exit, not a profit trim. If a holding's story breaks or fraud appears, the desk flags it for the operator to decide by hand rather than selling. Profit-trim suggestions stay notes only, and every exit is a manual operator decision. This supersedes the earlier rule that auto-closed dead-thesis and option positions. It is the pure hold-forever stance taken to its logical end, paired with the standing commitment to ride drawdowns rather than react to them.
2026-07-21Distress-zone solvency gate; loose tail-only; either agent can fail it.LAW #38
A name fails only if it is genuinely near going to zero: it cannot cover its interest from operations, its debt dwarfs its earnings with no cash to offset, or a bankruptcy-risk score sits clearly in the distress band with no cash cushion. The gate reads real solvency and cash flow, not a single ugly accounting quarter, so a cash-rich company with a one-time or stock-comp charge still passes. A weak but serviceable balance sheet passes; the screen is deliberately loose and removes only the going-to-zero tail. Both agents rate each name and either fail wins; an already-held fail becomes no-add and keeps its shares, never a forced sale. It is downside insurance for a never-sell, equal-weight book, where a single unrecoverable blow-up permanently drags long-run compounding, not a return signal. On adoption day it removed no names: the one apparent fail traced to a stale data window and, on current figures, passed.
2026-07-21Recursive feedback loop gate; a weak loop still passes; either agent can fail it.LAW #37
A name must show a durable, value-creating self-reinforcing loop to stay buy-ready: a data or AI flywheel, network effects, ecosystem or platform lock-in, a scale-economies flywheel, or a learning-curve cost loop. A weak but real loop still passes; only a clear absence, or a self-defeating capital loop that issues shares into its own premium, fails. Both agents rate each name and either fail wins. Already-held fails become no-add and keep their shares; never-held fails stay off the menu. Not an automatic sell. The gate is orthogonal to the quality and cost-curve screens: it separates businesses that compound their own advantage from linear beneficiaries of a spending wave.
2026-07-21Recursive feedback loop gate (weak equals pass; either agent can fail)ADOPTED
The operator adopted a required buy gate for self-reinforcing business loops: a name must show a durable, value-creating flywheel such as a data or AI feedback loop, network effects, platform lock-in, scale economies, or a learning-curve cost loop. Weak but real still passes; a clear absence, or a capital loop that issues shares into its own premium, fails. Both agents rate every name and either fail wins. The operator adopted it stronger than filed: the proposal asked for a scoring tiebreaker and the operator made it a hard gate. On adoption day four names flipped to no-add under the veto; held fails keep their shares. Public rule log entry 37.
Show all 55 entries
2026-07-21Distress-zone solvency gate (loose, tail-only)ADOPTED
The operator adopted the distress component of a five-gate junk-avoidance study as a required buy gate. A name fails only if genuinely near going to zero: interest not covered from operations, debt dwarfing earnings with no cash offset, or a clearly distressed bankruptcy-risk score with no cash cushion. The screen reads solvency and cash flow, not one ugly accounting quarter, so a cash-rich company with a one-time charge still passes. On adoption day it removed no names: the single apparent fail traced to a stale data window and passed on current figures, a phantom fail the two-agent check caught. The study's other four gates remain queued for the operator. Public rule log entry 38.
2026-07-21Remove all automatic selling; the desk is buy-onlyADOPTED
The operator struck every automatic sell from the live law. The desk now places buy orders only: no dead-thesis exit, no fraud exit, no option exit, no profit trim. A broken story or suspected fraud gets flagged to the operator, who sells by hand if ever; trim ideas stay notes. This supersedes the earlier auto-sell-on-written-triggers law and accepts the full implication that a blow-up rides down unless the human acts. It is the hold-forever stance taken to its logical end, paired with the standing commitment to ride drawdowns rather than react to them. Public rule log entry 39.
2026-07-21Hard cap on menu size at 10, 20, or 30 namesREJECTED
A five-year, lookahead-clean simulation tested capping the buy menu at 10, 20, or 30 names under several selection methods. Every capped routing arm trailed the uncapped equal-split book at similar or worse drawdown, and the flashy small-menu winner collapsed under ablation: deleting its own two anchor names erased a thirty-point lead, pure survivorship luck. The uncapped book kept most of its edge even after deleting its two biggest winners. The portfolio literature agreed independently: a wider screened menu buys more right-tail tickets, and equal weighting is the allocation you cannot reliably beat. Verdict: no cap. Let the quality gate set the count; if the eligible list runs thin, loosen the gate rather than concentrate. A soft target band stays queued for the operator.
2026-07-21Ten quantitative quality gates as return boostersREJECTED
Ten numeric quality screens, including a rule-of-40 test, margin floors, profitability ratios, a nine-point fundamentals score, leverage caps, and dilution caps, were backtested on the live menu over about five years. None improved returns out of sample, and two of the strongest academic factors actively destroyed return here because they tilt away from the book's asset-heavy infrastructure winners. The honest payoff of quality screens is blow-up avoidance and shallower drawdowns, not higher growth. One gate, an anti-dilution cap on share-count growth, was judged worth queueing for the operator as a shadow-first check; the other nine were skipped.
2026-07-19Pelosi disclosed holdings feed the candidate list; quality gates still decide buys.LAW #36
Every desk run refreshes publicly disclosed Nancy and Paul Pelosi U.S. equity underliers from STOCK Act trackers into a standing feed file. New names enter the menu queue and must clear the same quality filters and innovation cost-curve gate as any other candidate before they can go on deck. Already-owned menu names keep their normal daily recheck. The desk never buys just because the Pelosi book holds a ticker; both agents must still agree buy-ready before any equal-split share. Indexes, non-U.S. listings, and option contracts are skipped or mapped to the underlying stock only. Her selling is not an automatic lab exit.
2026-07-18Operator rescinded the 30% concentration note.LAW #35
Entry 31 is not live law. The operator rejected the optional 30% trim-suggestion rule after it had been published. The desk still stops new buys at 20% of lab market value and still notes optional trims at plus 50% and plus 100% versus cost. No separate 30% concentration note.
2026-07-18Innovation cost-curve gate; weak exposure still passes; either agent can fail it.LAW #34
Alongside the usual quality filters, a name must clear an innovation cost-curve screen to stay buy-ready: multi-year unit-cost collapse or step-function productivity as a core or material business (chips, AI infrastructure and power, automation software, and the like). A weak but real exposure still passes; only a clear fail blocks new buys. Both agents rate each name; if either says fail, that wins. Already-held fails become no-add (keep shares); never-held fails stay off the menu. Not an automatic sell. Five-year sims showed a pure fail-basket lagging badly, while requiring only a hard fail-out is a mild theme tilt versus the full haystack.
2026-07-18If a followed researcher sells an entire position, the desk holds off buys for three months.LAW #33
When a standing research source says they sold out of a name completely (not a partial trim), the desk marks that ticker hold-only for about ninety days: keep any shares already owned, stop new buys, exclude it from the equal-split. When the clock ends, the name goes back through the normal daily quality filters; a pass can restore buy-ready, a soft fail keeps no-add, and a dead thesis still forces a full exit. The source selling is not by itself an automatic sell for the desk.
2026-07-18Research pulls always hunt permanent process rules, not only tickers.LAW #32
Every standing research pull must fully read the source body, then do two jobs: extract buy-context names for the menu, and scan for permanent process ideas (sizing, when to add cash, hold rules, quality gates, hard bans). Candidates go to a private queue for the operator; nothing becomes law until that yes. Each pull must end with an explicit scan line so a ticker-only pass counts as incomplete.
2026-07-18Concentration free-insurance note at 30% (later rescinded).LAW #31
A note-only idea: if any name reached 30% of lab market value, the steward would suggest an optional trim toward 20% without placing a sell. It was published as entry 31, then the operator rejected it the same day. See entry 35. The hard 20% no-add rule (stop new buys) remains live; the 30% note does not.
2026-07-18Proposed rules are filed for the operator, never auto-adopted.LAW #30
When the two desk agents both think a new standing rule might help, they file it in a private proposed-rules queue instead of changing live law. The operator reviews, confirms, or denies. Nothing becomes desk law until that yes.
2026-07-18Per-name written kill conditions on buy-ready namesREJECTED
A proposal to attach a pre-written kill condition to every buy-ready name was denied by the operator. The daily filter recheck already re-screens every held and buy-ready name each run, so per-name kill lines would add authoring and judgment surface without adding protection. The desk keeps the simpler standing loop: recheck daily, stop buys on a soft fail, and flag anything worse to the operator.
2026-07-18Innovation cost-curve gate (weak equals pass; either agent can fail)ADOPTED
The operator adopted a required buy gate alongside the usual quality filters: a name must clear an innovation cost-curve screen (multi-year unit-cost collapse or step-function productivity as a core or material business). Weak but real exposure still passes. Both agents rate each name; if either says fail, that wins. Held fails become no-add (keep shares); never-held fails stay off the menu. Not an automatic sell. Public rule log entry 34. Immediate flips included no-add on two held names and a fail on one on-deck name.
2026-07-1830 percent concentration free-insurance noteREJECTED
A proposed note-only rule at 30 percent of lab market value was pressure-tested with a full menu simulation and briefly published as rule log entry 31. The operator then rejected it the same day. Forcing trims on concentrated stress paths destroyed moonshot returns in sim; the optional note was judged not worth keeping as law. The hard 20 percent no-add cap (stop new buys) remains. Rescind recorded as public rule log entry 35.
2026-07-18Pullback-timed extra deposit suggestionADOPTED
Both agents agreed a steward suggestion when the index sits 7 percent or more under its 20-day high is a real, low-cost edge for optional extra cash. Capital-matched tests showed pullback-only extras beat a broader volatility trigger, and every version of holding back scheduled deposits lost. Adopted as a dashboard banner and steward note only: the operator decides; nothing auto-pulls money. Public rule log entry 29. (Superseded on trigger asset by the 2026-07-28 lab-book retune above; still historical record.)
2026-07-18Alternative deploy and sell overlays vs equal-split holdREJECTED
A joint multi-agent study stress-tested dip-waiting, buy-the-laggard, momentum ranking, inverse-vol sizing, profit trims, trailing stops, tax-loss harvest, and forced rebalance against Plan C equal-split over 24 months. Laggard-buying only won via survivorship on one name; aggressive sell rules lost; equal-split hold survived as law. No process change except the two adopted suggestion notes filed the same day.
2026-07-17Two-agent agree before every deploy; disagreements are public.LAW #28
Before any cash is deployed, the desk agent and a second agent both score every menu name. If either says fail, no-add, or hold, that call wins. Buy-ready for the equal-split requires both to say buy-ready. When they disagree, the name is parked and both calls (who said what) are logged on the public dashboard. If the second agent is unreachable, the desk agent deploys alone and notes the solo run.
2026-07-17Two-agent deploy gate before every cash deployADOPTED
The operator adopted a standing dual-score rule for every equal-split deploy. Before cash moves, the steward finishes its quality-gate calls, then the red team independently scores every menu name (buy-ready, no-add, fail, or hold). If either says fail, no-add, or hold, that call wins. Buy-ready for the deploy requires both. Disagreement parks the name and publishes who said what on the public Deploy gate table. If the red team cannot answer, the steward solo deploys on its own buy-ready set and logs the solo run so the desk is not frozen. Adopted as live law the same day; equal-split arithmetic is unchanged.
2026-07-16No permanent equity ticker bans; recheck every held name.LAW #27
A prior permanent no-buy carve-out on one legacy holding was retired. Every name still held is rechecked on the same quality gates each run. Soft fails stop new buys and keep shares; a later pass restores buy-ready status. Dead thesis still forces a full exit. Options underlier quality rules stay separate.
2026-07-16Full video-body ingest; seven-day gate dedup.LAW #26
Standing YouTube must be read from the spoken video body via transcript or auto-captions, not title or social caption alone. The window for skipping a re-score of a recent fail or hold-only result is seven days, not thirty, unless new material facts appear. Buy-ready names are still rechecked every run.
2026-07-16Buy-context only; skip sell and avoid mentions.LAW #25
Ticker extract from standing sources only counts names the source is buying, long, or adding. Names that appear only in sell, exit, avoid, or do-not-buy context are not queued for the quality gates. Video body and captions beat clickbait titles. A same-day miss on a buy name was fixed and promoted to buy-ready after a pass.
2026-07-15Every ticker named by followed sources gets the quality gates.LAW #24
Standing sources can surface new names. The desk must extract each equity ticker and run the same filters the same day. A pass puts the name on deck for the next equal-split deploy. A fail records the gate and keeps it off the buy button. Already owning the stock in another account is never a reason to skip the screen.
2026-07-14One-shot verdict confirmed by a robustness pass.LAW #23
A second same-day study randomized two thousand possible menu-add histories and re-ran the comparison; the one-shot's apparent edge dissolved into schedule luck, a coin flip with extra noise. Published research on index additions and diversification agrees. Equal-split stays law; the one-shot remains optional ops only.
2026-07-14New-ticker one-shot deploy evaluated; not adopted.LAW #22
The operator proposed that when new names join the buy menu, the next deposit should go entirely to the newcomers once, then return to equal-split. A five-year staged-menu backtest plus Monte Carlo found a tiny, path-dependent edge that loses most rolling windows. Rejected as standing law; allowed as a one-time convenience when opening zero-share names.
2026-07-14New-ticker one-shot deploy on menu addsREJECTED
The operator proposed that whenever new names join the buy menu, the very next deposit should go entirely to the newcomers, split evenly, then revert to standard equal-split. Two independent studies ran the same day: a five-year staged-menu backtest showed a small edge, and a second pass that randomized two thousand possible menu-add histories showed that edge was schedule luck, a coin flip wearing a lab coat. Published research agrees: freshly added names carry no tailwind after joining a list, and spreading each deposit across the whole menu keeps the odds of feeding the eventual winners. Rejected as standing law; permitted as a one-time convenience when opening zero-share names, worth about one lunch either way.
2026-07-13Buy-only rebalance rejected as standing law.LAW #21
Backtests plus a thousand-path Monte Carlo showed that steering each deposit toward the most underweight names wins only on lucky start dates and loses most windows, with identical drawdowns. Equal-split across the menu stays law. A momentum tilt showed the only persistent edge signal and stays parked pending the beat-the-benchmark checkpoint.
2026-07-13Daily recheck; failed filters stop new buys.LAW #20
Every steward run rechecks names the desk still holds or marks buy-ready. If a name fails the quality gates but is not a dead thesis, the desk stops buying more of it and keeps the shares (no automatic sell on a soft fail). A later pass restores buy-ready status. Dead thesis or fraud still forces a full exit.
2026-07-13Auto quality gates on new menu names; two names cleared.LAW #19
Any new or queued ticker must run the desk quality gates as soon as it appears. A pass promotes it to buy-ready for the next equal-split deploy; a fail keeps it off the buy button and records the failing gate. Calendar phase hints no longer block a completed pass. The two names that had been waiting on a later phase both passed the same day and are open on the menu for the next deploy.
2026-07-13Buy-only rebalance and ranked underweight deploysREJECTED
After a menu widen opened three names at zero weight, the desk stress-tested whether each deposit should chase the most underweight holdings (buy-only rebalance) instead of pure equal-split. Full-history biweekly cash sims over about one year, plus four hundred block-bootstrap paths, put equal-split ahead by roughly ten percentage points of return on injected capital, with only a mildly worse drawdown. Underweight top-two and top-three ladders lagged further. Dip-first and momentum-first top-two rules printed the highest raw returns but drove single-name weight near a third of the book, which fails the desk's concentration and behavior tests. Auto price trims and a trailing stop also underperformed hold. Standing law stays equal-split across the buy menu; evening weights via purchases only and ranking deploys by underweight, dip, or momentum remain rejected as defaults. A one-shot catch-up after a menu widen is optional tactics, not a rewrite of the law.
2026-07-13Dip-first and momentum-first deploy as standing lawREJECTED
The same bake-off crowned dip and momentum deploy variants on raw return in a strong tech year. The red team rejected promoting either to law: end weights and Monte Carlo left tails were worse than equal-split, and both rules reintroduce judgment and regime risk the steward is not allowed to freestyle. They stay in the graveyard next to hybrid ranked-deploy.
2026-07-11Pruned dead language from the law.LAW #18
The live rulebook was cleaned to one canonical plan name, all short-options process was removed (short-dated options remain banned outright), and hybrid deploy language was deleted. The append-only history keeps old lessons visible without re-authorizing retired tactics.
2026-07-11Plan C approved as live law.LAW #17
Equal-split deploy across every eligible buy-menu name went live, replacing ranked deploys. Price-based trims demoted to note-only. Menu widened by three names first, with two more named for a later phase pending quality checks (those two later passed and are buy-ready; see the 2026-07-13 rule). One legacy name locked to hold-only. Added a 20% single-name no-add cap and an every-run check for ad-hoc deposits. A hybrid ranked-deploy proposal was rejected as spec-fragile.
2026-07-11Hybrid ranked-deploy proposalREJECTED
A proposal to rank each deploy (part equal-split, part momentum and underweight logic) was rejected as spec-fragile: it needed too many judgment calls per run, and every judgment call is a place where an automated steward can drift. The desk adopted pure equal-split as law instead, with a narrow ops-degradation fallback that is explicitly not a ranker.
2026-07-11Short options and sub-365-day optionsREJECTED
The original options lab was re-run as a twelve-month simulation under generous assumptions and still netted roughly zero while a simple SPY drip gained over 11%. Short options and any option under 365 days to expiry are now permanently banned; the wins in the record clustered inside a hindsight-contaminated window, so no selection skill was credited. The ban is written into the live law, not just the journal.
2026-07-10First live steward deploy executed.LAW #16
The scheduled steward found $500 of settled cash and deployed all of it the same run, split across two menu names under the then-current ranking rule. Both fills landed in regular hours. Trigger scan came back clean: no trims, no exits.
2026-07-10Regular-hours orders only, two scheduled runs.LAW #15 :: SUPERSEDED BY #41
[No longer in force. Law #41 replaced this.] All orders go in Monday to Friday, 09:30 to 16:00 ET. The 11:49 AM run trades; the 11:49 PM run observes and defers. This avoids overnight queues and open-gap fills far from intended marks.
2026-07-10Minimum cash gate for long-dated options.LAW #14
Long-dated calls are only even considered when free cash is large enough for one full contract; below that, everything goes to equities.
2026-07-10Auto-sell on written triggers.LAW #13 :: SUPERSEDED BY #39
[No longer in force. Law #39 replaced this.] Superseded ask-first selling: agents execute pre-written triggers automatically (thesis dead or fraud is a full exit; long-dated option exits at roughly double the entry or late in life), while anything outside the written list still needs the operator.
2026-07-10Haystack Hold adopted from Monte Carlo research.LAW #12
Hundreds of simulated policy paths compared holding styles. Mostly-hold equity baskets beat tilted and concentrated variants, so the desk adopted: hold a menu of quality names, deploy every deposit quickly, and keep any options exposure to a small, long-dated-only sleeve.
2026-07-10No sells without approval; deploy free cash fully.LAW #11
Agents stopped placing any sell without the operator's yes, and every steward run began deploying the full settled cash balance instead of letting it sit.
2026-07-10Holding style bake-off: haystack vs ranked vs tiltedAGREEMENT
An eleven-agent council plus a Monte Carlo study (hundreds of bootstrapped policy paths over historical windows) compared holding styles head to head. Mostly-hold equal-weight baskets beat conviction-ranked and tier-tilted variants on risk-adjusted outcomes, and the maximum-return variant carried a drawdown profile nobody wanted. The council converged on Haystack Hold: own the menu, keep deploying, stop fiddling.
2026-07-09Long-dated options do not fit a small budget.LAW #10
Structural finding, not situational: near-the-money long-dated calls on lab-priced names cost far more than the deposit size, so the only affordable ones are deep out-of-the-money lottery tickets, which the filler-leg ban already forbids. Real long-dated exposure is a budget decision, gated behind beating SPY first.
2026-07-09Twelve-month simulation rewrote the option rules.LAW #2-9
A full-year simulation of the original short-dated options approach netted roughly zero while a simple SPY drip gained over 11%. Adopted from the evidence: a momentum entry gate (no bottom-fishing entries far below the 52-week high), a ban on filler legs (no tickets under $75 premium, no far out-of-the-money strikes), simplified exits built around a late-life decay kill, a rule that a catalyst only counts if it lands before the exit rules would fire, SPY as the binding benchmark before any sizing up, a take-profit ladder, an independent quality gate for every additional leg, and a regime filter that skips months when the speculative complex is deeply red.
2026-07-09Existing holdings inform, never exclude.LAW #1
Positions held elsewhere by the operator may shape sizing and correlation awareness, but are never a reason to skip the best candidate. Reversed an earlier session that excluded a name for exactly that reason.
2026-07-09Simulation bugs caught in a proposed deploy tweakDISAGREEMENT
The red team audited a proposed change to how deposits get deployed and found three bugs in the supporting simulation: a cold-start artifact, a seed-dependent result, and a missing control run. With the bugs fixed, the corrected numbers flipped the recommendation: the fancier variant lost its edge and a true equal-split deploy won. The proposal was rewritten before anything touched the live account.

ORIGIN LOG: SHIPPED AS SKINVICTION

This desk went live as SKINVICTION: skin in the game plus conviction. The founding gripe: people hold conviction about a hundred things and fund approximately zero of them. We do the opposite. Real money, written rules, every trade scored in public. House doctrine: shoot first, aim later. Fail fast, fix fast. The name shipped the way everything here ships: before it was perfect.

Production returned two corrections. First, the protocol became the identity: stop hunting the needle, buy the whole haystack, an equal split across a quality-gated menu developed by two AI agents and one human. Second, skinviction.com kept receiving visitors expecting a very different kind of website, on account of the word "skin". What they found instead was a ledger. So the name got patched in public, same as any bad position: PROJECT HAYSTACK. The new name has the advantage of being true. Old domain retired. Doctrine intact.