The organism's long-term memory. Every researcher session APPENDS an entry;
nothing is ever rewritten or deleted. Read this file (newest-first is fine)
before proposing anything — dead ideas stay dead unless new evidence is
cited explicitly.
Format per entry:
## YYYY-MM-DD — <one-line headline> **Autopsy:** what yesterday's trading did right/wrong (genome by genome, with $). **Hypotheses tested:** idea → backtest verdict (net-of-cost numbers, not vibes). **Actions:** births (genome_id + gate evidence), mutations (param, before→after, why), kills proposed, no-ops. **Do-not-retry:** ideas confirmed dead this session, and why.
2026-08-15 — Organism born (seed population)
Autopsy: n/a — first entry. Prior system history: 36 books culled 2026-08-14
(all options + all unprofitable equity). Survivors became the seed gene pool.
Hypotheses tested: none this session — seeds carry prior validation:
- `g_tsla_orb15_holdeod`: 7yr TSLA 5-min battery winner (net Sharpe ~1.25 w/ QQQ
filter, OOS 1.12, positive every year). Born EMBRYO $2k.
- `g_tsla_macd_trend`: 26-strategy daily battery winner (net Sharpe 1.02/0.85 OOS).
Born EMBRYO $2k.
- `tsla_morning_pullback` keeps its dedicated timer (live +$3.3k/60 fills);
migrates into the runner after parity is proven.
Do-not-retry (inherited from repo history — cite new evidence to reopen):
- Mean-reversion / VWAP-fade / RSI-fade on TSLA intraday: LOSES net-of-cost
(2026-06-24 battery, all variants).
- "Wait for the pullback/FVG retrace" on TSLA breakouts: fights the edge.
- Naive ORB without liquidity/RVOL context on general stocks: zero-slippage
artifact (2026-06-05).
- Ross Cameron momentum-pullback on gappers: cost artifact, OOS PF 0.97 @5c
slippage (2026-06-15).
- Indicator-confluence day-trading (multi-indicator voting): PF 0.80 net (2026-06-08).
- Options strategies of any kind: NO free real-time options feed exists
(2026-08-14 verification) — do not propose until that changes.
- Fixed 2:1 targets on TSLA ORB: cuts Sharpe (1.33→1.15). Hold-to-close wins.
- Pre-FOMC drift, ETF pairs, BAB/low-vol tilt, VRP overlay: all dead in the
2026-06-04 105-agent research sweep (no net retail edge).
2026-08-15 — VWAP+EMA9 battery (owner-proposed): NO deployable entry edge
Hypotheses tested: the four canonical VWAP+EMA9 intraday playbooks on 5-min
RTH bars, TSLA (2021-2026) + SPY + QQQ, net 5bps/side, IS (…2024) vs OOS (2025…),
~30k trades examined. Engine sanity: the VWAP-fade control lost everywhere,
consistent with the 2026-06-24 battery. Full code + numbers:
lab/memory/experiments/vwap_ema9_battery.py
- EMA9×VWAP crossover: TSLA IS PF 1.23 (+74%) → OOS PF 0.65 (−41%). GROSS OOS
is also negative (−3.2 bps/trade) — the signal decayed, not just the costs.
SPY/QQQ: loses both segments. Classic IS-only artifact.
- Trend-pullback to EMA9 above/below VWAP (raw): PF 0.10–0.36 everywhere.
Fires ~6k times; death by frequency × costs.
- Steel-manned selective version (AM-only, ≤2/day, first-pullback, ≥0.2%
stretch): PF 0.29–0.53 net on all 3 symbols BOTH segments, and NEGATIVE
GROSS on TSLA (−4 bps/trade) — no cost model rescues a no-gross-edge signal.
- VWAP reclaim w/ EMA9 slope: PF 0.15–0.53 everywhere.
Actions: no birth. Battery archived to experiments/.
Do-not-retry: VWAP+EMA9 (any of: crossover, trend-pullback, reclaim, fade)
as an ENTRY signal on liquid large-caps at 5-min — no gross edge OOS, and
10bps round-trip exceeds the signal. The components remain valid as MANAGEMENT
tools (the live tsla_morning_pullback winner already trails with EMA9; VWAP-side
bias is admissible as a FILTER on a validated structural entry, e.g. ORB —
that combination is untested and is the only respectable follow-up).
2026-08-15 (later) — ORB×VWAP filter: redundant · Entry-style: genome MUTATED
Hypotheses tested: (1) VWAP-side bias as a filter on the validated ORB-15
entry — the one respectable follow-up from the VWAP+EMA9 battery. (2) The
live genome's close-through entry vs the June battery's stop-at-level fill.
Evidence: lab/memory/experiments/orb15_entry_style_backtest.json (+ code).
- VWAP filter: NO-OP. qqq+vwap produced the IDENTICAL trade list to qqq —
a close through the OR high is always above session VWAP already.
- Entry style: stop-at-level beats close-through in BOTH halves
(IS Sharpe 0.63→1.03; OOS PF 0.911→1.048, avg −0.037R→+0.023R). The
close-through variant this genome launched with was quietly giving back
the first leg of every breakout.
Actions: MUTATION (owner session): g_tsla_orb15_holdeod now arms a resting
STOP order at the OR level on the QQQ-agreeing side, cancels/re-arms on tape
flips (the Pine's OCO behavior). Runner gained a pending-order lifecycle
(submit_stop / order_status / cancel_order; fills convert to positions at the
actual fill price; unfilled orders expire 15:30). 39 tests green.
Do-not-retry: VWAP-side filter on ORB-15 (mathematically redundant at the
trigger). Close-through entries on level-break strategies when a resting stop
order is available — the fill model IS part of the edge.
2026-08-15 (later still) — Momentum rotation: NO BIRTH (loses to SPY)
Hypothesis tested: cross-sectional 12-1 momentum, long-only top decile,
monthly rebalance, on the SURVIVORSHIP-FREE panel (5,511 tickers incl. 1,541
delisted, 2021-05→2026-05), net of costs, one fixed config (N=1, no sweep).
Evidence: lab/memory/experiments/momentum_rotation_backtest.json
- Bias-free, research costs: IS Sharpe 0.78 (DSR 0.93) → OOS Sharpe 0.49
(DSR 0.73), OOS net +9.97%. Realistic costs: OOS 0.48, +9.64%.
- Benchmark, SAME OOS window: SPY buy-and-hold +22.2%, Sharpe 0.84.
- Survivorship isolation: dropping all 1,541 delisted names moved OOS Sharpe
0.494→0.498 — the bias is ~nil in this window/universe.
Actions: NO birth — positive absolute returns but roughly HALF the
risk-adjusted return of doing nothing (SPY). Fails the opportunity-cost bar
that retired low_vol_etf_tilt (2026-06-04).
Do-not-retry: plain 12-1 monthly momentum rotation on the broad US panel
as a genome, unless a variant demonstrably beats SPY OOS net-of-cost (regime
gating or vol-targeting would need their own honest N-trials accounting).
2026-08-15 — EPOCH RESET: fresh accounts, day zero
Owner deleted the three original paper accounts and created new ones, each
$100k: core PA3ESRQAV5PY · highrisk PA3UM5SV3FY3 · organism PA3Y2UAU2W8W
(all ACTIVE, crypto ACTIVE, shorting enabled). Keys rotated on Mac + VPS.
Database truncated after backup (/home/tradervex/backups/pre_reset_2026-08-15
.sql.gz) — the public track record starts HERE. Pre-reset history and every
lesson from it live in THIS journal and experiments/; the Do-not-retry list
carries forward. Breaker/HWM/killswitch/book state reset to $300k baseline;
stale strategy state archived to harness/state/archive_pre_reset_2026-08-15/.
Monday is day one of the clean record: stock engine rebuilds its book,
morning-pullback + ORB (stop-entry) + MACD genomes trade the organism account.
2026-08-15 (weekly session) — PEAD long-only: NO BIRTH, prior claim refuted
Autopsy: watermark unchanged — epoch reset happened earlier today, zero
trading days have elapsed since (Saturday night, market closed all weekend),
population still 2 EMBRYO cells / 0 live trades / 0 fills in `trades` since
the reset. Nothing to autopsy mechanically; no bug, no parameter data.
Hypotheses tested: population has headroom (2/12), so used the pre-Monday
window to test the one concretely-named fair-game idea in the researcher
contract: PEAD long-only (earnings-reaction pop → hold 3 days), using the
existing decision core at `harness/strategies/pead_live.py`. Built a FRESH
backtest this session — EDGAR 8-K/item-2.02 filings (via `company_tickers.json`
+ `submissions/CIK##########.json`, keyless) for earnings-reaction dates,
Alpaca daily bars already cached in `harness/data/` for prices, 108 large-cap
liquid names, chronological max_concurrent=8 portfolio sim, 5bps/side cost.
Code + artifact: `lab/memory/experiments/pead_earnings_drift_backtest.py`
+ `.json`.
- IS (2025-01-21…2025-10-31, n=69): PF gross 0.983 / net 0.918, avg ret/trade
net −0.13%. OOS (2025-11-01…2026-05-20, n=56): PF gross 1.037 / net 0.982,
avg ret/trade net −0.03%. Gross is flat-to-negative in BOTH windows —
this isn't a cost-erosion story, the raw signal has no edge before costs
either. Portfolio max DD 2.41% (fine, but moot given no edge).
- Gates (OOS, net): PF≥1.15 FAIL (0.982). ≥100 OOS trades FAIL (56, real
count, universe/window is just thin — not padded to hit the number).
MaxDD≤20% PASS (2.41%). Overall 1/3 — NO BIRTH.
- Data caveats worth remembering: XOM's ticker maps to a shell entity in
SEC's `company_tickers.json` (zero filings) — needed a manual CIK override
to the real ExxonMobil filer (34088) to get its 26 8-Ks; other symbols may
have similar shell/restructuring mismatches undetected here. TSLA's 8-K
2.02 tag also catches monthly/quarterly delivery-number releases, not just
GAAP earnings — roughly half its 14 triggers in-window are delivery prints,
a known imprecision of the free-EDGAR proxy vs. a real earnings calendar.
Sample is thin regardless (~16 months of cached bars, 1-2 earnings cycles
per name) — this result should be read as "no edge found in a fair, if
short, test," not a definitive multi-cycle refutation.
Actions: no birth, no mutation (nothing live to mutate), no bug fix
(nothing has traded yet). Population unchanged at 2 EMBRYO cells.
Do-not-retry: the repo's prior "PEAD long-side validated +0.79%/trade
(2026-06-05)" claim is REFUTED by this fresh, honest, net-of-cost run — do
not re-cite that number as evidence for anything going forward. Plain
earnings-reaction-pop-then-hold-3-days PEAD, entered via the free EDGAR
8-K/item-2.02 proxy at a 3% reaction threshold, on liquid large-caps: no
gross edge in either IS or OOS window tested. A respectable follow-up would
need either (a) much longer price history than what's currently cached
(only ~16mo in `harness/data/`) before concluding anything more permanently,
or (b) a cleaner earnings-only trigger that filters out non-GAAP 8-K/2.02
noise (e.g. TSLA's delivery prints) — untested, and not worth doing again
without a plan to fix the trigger precision or extend the price history.
2026-08-16 (weekly session) — Crypto Donchian trend battery: NO BIRTH
Autopsy: watermark unchanged since 2026-08-15 (epoch reset). Zero trading
days have elapsed — Saturday night reset, Sunday all-weekend market closure,
0 fills in `trades` since the reset, `organism_runner.log` shows `actions=0`
on every 5-min tick all day. `organism-select.timer` hasn't fired yet either
(next run Mon 2026-08-17 20:20 ET — first selection cycle since the reset).
Population unchanged: 2 EMBRYO cells (g_tsla_orb15_holdeod, g_tsla_macd_trend),
0 live trades each. Nothing to autopsy mechanically; no bug, no live param to
mutate.
Hypotheses tested: population has headroom (2/12) and carries zero crypto
genomes despite crypto majors being explicit fair game (rule #5). Tested
classic Turtle-style Donchian breakout trend-following, long-only spot, on
the sandbox's full crypto allowlist (BTC/USD, ETH/USD, SOL/USD, LTC/USD) —
daily bars, fresh Alpaca crypto fetch (keyless), 2021-01-01→2026-08-15
(SOL from its 2021-09 listing). Two canonical configs: System 1 (20-day
entry / 10-day exit) and System 2 (55/20). Entry = close breaks above prior
N-day high; exit = close breaks below prior M-day low (shift(1), no lookahead).
Equal-weight 25%/symbol when signaled long, cash (0% yield) otherwise. Cost
10bps/side (20bps round-trip, the contract's stated crypto minimum). IS
2021-2024, OOS 2025-2026(→08-15). Code + full numbers:
lab/memory/experiments/crypto_donchian_trend_backtest.py + `.json`.
- System 1 (20/10): IS n=112, gross PF 3.29 / net PF 3.20, portfolio
+705.71% (the 2021-2024 bull carried it hard). OOS n=49, gross PF 0.82 /
net PF 0.78, avg net ret/trade −1.00%, portfolio −13.55%, max DD −37.7%.
Net PF < 1 OOS — fails the ≥1.15 gate outright.
- System 2 (55/20): IS n=48, PF 2.79/2.73, +120.7%. OOS n=22, PF 0.33/0.32,
avg net ret/trade −6.26%, portfolio −31.4%, max DD −32.1%. Worse than
System 1 — the slower system's few big whipsaws hurt more when the trend
the whole method depends on isn't there.
- Benchmark: BTC buy-and-hold same OOS window: −33.28%. Both systems lose
LESS in absolute drawdown than holding BTC naked (partly just being out of
the market more), but that's cold comfort — net PF < 1 means the strategy
itself has negative expectancy, not just "beats a falling asset."
- Read as regime, not bug: 2021-2024 was a historic multi-leg crypto bull —
textbook trend-following terrain. 2025→2026-08 has been a grinding
BTC-down / range-bound regime — the whipsaw environment trend-following is
known to bleed in. This is the IS-only-artifact pattern already burned once
this week (VWAP+EMA9 2026-08-15): great in-sample number, sample-dependent,
no OOS edge.
Actions: no birth. Gates (OOS, net, best of the two): PF≥1.15 FAIL (0.78
best case). ≥100 trades: pooled IS+OOS trade count clears it, but the OOS-only
segment that matters for the PF number is thin (49, 22) — moot since PF
already fails. MaxDD≤20%: FAIL (−37.7%, −32.1%). 0/3 on the segment that
counts. Population still 2 EMBRYO cells, 0 live trades — nothing to mutate.
Do-not-retry: plain Donchian/Turtle breakout trend-following (any of
20/10, 55/20) as a standalone long-only entry on the crypto majors basket
(BTC/ETH/SOL/LTC) at daily granularity — strongly regime-dependent, decisively
negative net-of-cost in the 2025-2026 chop/down window tested here. A
respectable follow-up would need an explicit regime gate (e.g. only trade
breakouts while a longer BTC trend filter — 200sma or similar — agrees) tested
as its own honest IS/OOS split, not a data-mined patch on this result — untested,
not worth attempting again without that plan.
2026-08-19 — Icarus fade (owner-proposed): reversal is REAL, the short expression is UNDEPLOYABLE
Hypothesis tested (owner, re: MRNA): "stocks that go up too fast too
quickly dip on profit-taking — fade them." Survivorship-free close panel
(5,511 tickers incl. 1,541 delisted, 2021-05→2026-05), events = +50%/10d
(E10_50, n=2,949) and +30%/5d (E5_30, n=6,518), 20-bar cooldown, entries at
next close, IS ..2024-12 / OOS 2025-. Shorts net of 10bps/side + 25% APR
borrow; long-continuation control; first-red-day trigger variant. Code +
full numbers: lab/memory/experiments/icarus_fade_backtest.py + .json.
- The direction is right. After a +50%/10d run, the average name bleeds:
gross short +3.5%/10d, +8.1%/20d IS; +2.1%/+4.0% OOS. The long control
LOSES in both segments (PF 0.48–0.93) — buying the spike is the worst seat
at the table. Median short outcome positive nearly everywhere.
- The tradeable edge is thin and tail-poisoned. Best config
(E10_50 naive h20) OOS net PF 1.161 / +1.8%/event — a hair over the 1.15
gate — but single events lost −199%…−558% net (E5_30: −2,808%!). Close-only
stops cannot cap an overnight squeeze; a handful of GME-style rips erase
years of median wins. The faster E5_30 trigger is NET NEGATIVE OOS
(PF 0.74–0.93). first_red (wait for the first −2% close) forfeits the
early reversion without dodging the squeezes: OOS PF 0.94–1.10.
- Costs are understated by design: flat 25% APR borrow; real HTB fees on
exactly the squeeziest names run 100–300%+ and land on the worst events.
The marginal 1.16 PF does not survive realistic borrow skew.
- MRNA context on test day: +210%/10d, closed AT its run high, 4.3× volume —
3–6× beyond the tested threshold and still FRONT-SIDE (no distribution
bar yet). Exactly the profile of the events that produced the tail losses.
Actions: NO BIRTH. Verdict: mean-reversion after parabolic runs exists
gross, but a naked-short genome fails on tails + borrow. The professional
expression of this edge is DEFINED-RISK (long puts / put spreads, max loss =
premium) — unavailable here: no options feed (2026-08-14 verification).
Do-not-retry: naive or first-red-day SHORT fading of parabolic run-ups
on the broad panel as a genome — OOS PF ≤1.16 with unbounded single-event
tails. Reopen ONLY with (a) a real options feed for defined-risk expression,
or (b) a portfolio-level sim with hard per-event notional caps + realistic
per-name borrow-fee data showing the tails are survivable — neither exists
today.
2026-08-19 (later) — Icarus, options expressions: PUTS DEAD on real data; spreads inconclusive
Correction first: the repo DOES hold options history — harness/data/options,
2.4GB, 353 liquid underlyings, per-contract DAILY + 5-min TRADE bars,
expiries 2024-05→2026-05 (Alpaca paid-era downloads), + SPY-only NBBO/OI.
"We'd need to buy history" (earlier today) was wrong for a first pass.
Put-buying pilot (lab/memory/experiments/icarus_put_pilot.py): 85 of
1,331 post-2024-06 parabolic events overlap the options universe; 28 had a
tradeable put print (~45d, 95%-strike, 10% spread pads). Result: mean
−49.3%/trade on premium, median −88.5%, win 21%, PF 0.27 — negative in
2024, 2025 AND 2026. The underlying reversal (+2-4% net short, prior
battery) is REAL but SMALLER than what the options market charges at
parabolic tops: entry IV is at panic highs, then crushes. Removing the
spread pads entirely still leaves ≈−30%/trade — the pads are not the story.
DEAD. Do-not-retry: long puts (or any long-premium structure) on
parabolic-run-up fades — the vol is priced worse than the move.
Call-credit-spread pilot (icarus_callspread_pilot.py): the vol-SELLING
mirror — n=9 only (both legs rarely print same-day), and results carry
crossed-print artifacts (ret-on-risk +890%/−170% are impossible for a
defined-risk spread → stale trade prints, not real fills). INCONCLUSIVE,
not dead. Two-leg structures cannot be priced from trade bars; they need
historical NBBO QUOTES, which exist here only for SPY. Cheapest honest
paths if ever revisited: a one-off CBOE DataShop/ORATS EOD-quote slice for
the ~130 event names, or Polygon's quote tier (~$199/mo — NOT the $29
aggregates tier, which would not fix this).
Standing verdict on the whole Icarus family: underlying short = marginal
net with fatal tails; long puts = dead; credit spreads = untestable without
quote data. NO BIRTH anywhere. The free nightly scanner + shadow book
(scripts/icarus_scan_run.py, live since tonight) remains the only active
piece; its pre-registered gate (≥60 shadow trades, net PF ≥1.3) is now the
bar for spending anything on quote data for the spread variant.
2026-08-19 (owner decisions) — researcher 2x/week; reserve stays cash; e2e validation
Owner: "proceed with recommendations, validate end to end."
- Researcher cadence → Sun + Wed 23:30 UTC (was weekly). Owner has
pushed twice for more AI-strategy flow; sessions are Haiku on the metered
key, so doubling is cheap. Timer deployed + verified — the FIRST Wednesday
session fires tonight. Prereqs re-verified: claude CLI 2.1.233 at
/home/tradervex/.local/bin/claude, ANTHROPIC key present, wrapper intact.
- Cash reserve ($100k) stays idle. Trigger to revisit, pre-registered:
when the Organism's allocated cells exceed ~60% of its own $100k account
(currently $7,750 = 8%), reserve deployment becomes a real conversation.
Capital is earned by results, not granted by boredom.
End-to-end validation (all green): full suite 4,802 passed / 0 failed;
VPS on latest commit; 25+ timers armed with sane next-runs; icarus unit
force-run OK (walk-back + idempotency verified); TSLA fractionable=true on
the organism account (MACD sizes 0.23 sh ≈ $80 — the fix is deliverable);
site 9/9 pages 200, sleeves sum exactly ($99,657 + $100,000 + $97,361 =
$297,018); health panel's two 24h failures both explained (mp SIGABRT
09:38 ET — consequence defused by the reconcile guard; health-alerter blip
19:05 UTC = it probed the API mid-deploy-restart, self-recovered).
Tomorrow's expected prints: PEAD exits CSCO+AMAT ~09:40 ET (books at real
fills for the first time), mp/ORB trade the open, icarus scan processes
8/19 tonight at 22:00 ET, researcher session 19:30 ET.
2026-08-19 (second session, Wed) — ORB15 first trade; morning_pullback selection pressure — no-op
Autopsy: Market: SPY +0.44%, QQQ +0.08%, BTC +6.46%, ETH +9.54%, VIX 15.84. TSLA +2.8% from open to close; favorable for breakouts.
Genomes:
- `g_tsla_orb15_holdeod` (EMBRYO): 1 trade: shorted 5 TSLA at 337.55 (OR low break), exited at 342.28 (stop), realized loss −$23.62. Stop-order mechanics sound (confirmed: resting stop-sell at 337.64 fired correctly, exit at stop-buy 341.14 filled at 342.28 due to normal slippage in a rally). Single trade is noise; insufficient data to judge the 2026-08-15 stop-entry mutation.
- `tsla_morning_pullback` (PAPER): 3 fills: sell 144 TSLA @ 343.15 (+$450), sell 145 @ 347.39 (+$1,068.65), buy 296 @ 348.37 (−$943.14) = net +$575.51. Profitable today but in a neg_streak: SELECTION punished −5.58% (2026-08-17), 0% (2026-08-18), capital cut from $10k → $3.75k over 2 days. Fitness −0.0314/−0.0316. This is a legacy genome (pre-organism seed); its parameter space is not fully documented here.
- `g_tsla_macd_trend` (EMBRYO): 0 trades. Still waiting for a setup.
Population: 3 cells live, 2 of 12 allowed. All genomes firing/idle as expected.
Hypotheses tested: none this session. Considered new births (crypto Donchian with regime gate; ORB with VIX filter), but each would need honest backtest evidence (PF ≥1.15 OOS, ≥100 trades, maxDD ≤20%) on cached data. The Do-not-retry list from 2026-08-15 through 2026-08-19 covers every major hypothesis tested; the next respectable follow-ups lack validated backtests and would cost 30–60 min to build fairly.
Actions: none. No births (insufficient hypothesis evidence). No mutations (ORB15 data too thin; morning_pullback parameter space unclear). No kills. Population stable.
Do-not-retry: nothing new. The inherited list remains the guardrail (mean-reversion TSLA intraday, VWAP+EMA9, naive ORB without liquidity context, Donchian crypto, Icarus shorts + puts, PEAD, momentum rotation).
Session tone: skeptical quant discipline. A no-op with a clear autopsy beats a forced mediocre birth. ORB15's stop-order mutation is sound; it needs more trading data before we know if it's a winner. Morning_pullback is being correctly pruned by SELECTION's 10-day fitness window — let it prove parity or get retired. Next session: watch for ORB15 and MACD accumulation, and monitor whether morning_pullback stabilizes or continues to decay.
2026-08-19 (owner-proposed) — Weekly $5/$10 reversal grid: NO BIRTH (loses to SPY money-weighted, badly)
Hypothesis (owner, exact spec): buy $5 of any stock down ≥5% on the week;
sell $10 of any holding up ≥10% on the week; all stocks simultaneously.
Long-only interpretation (can only sell what you hold). Survivorship-free
panel, signals Friday close, fills next session close, 25bps/side.
Code: lab/memory/experiments/weekly_reversal_grid_backtest.py + .json.
Artifact hunt worth remembering: v1 said +$1.04M — fantasy. The panel's
per-day $5/$1M filter makes 3,320 names drop out for a median 296 days
(exactly their collapse window); liquidating on dropout exited dip-buys at
pre-collapse prices. v2 holds through dropouts (eats the reappearance
price) and liquidates only true delistings under recovery scenarios.
LESSON: this panel is built for TOP-of-market strategies; for buy-losers
strategies its filter censors the loss tail — rebuild unfiltered or bound.
v2 results (5yr, ~125k buys / ~40k sells, peak deployed ~$204-255k):
- delist recovery 100%: PnL +$62k vs SPY-same-flows +$152k
- delist recovery 50%: PnL +$33k vs SPY +$163k
- delist recovery 0%: PnL +$3k vs SPY +$173k (IS segment NEGATIVE −$8.6k)
Max DD ~$47-53k against those gains; ends holding 2,925 positions.
And v2 is STILL optimistic (no averaging-down inside dark valleys; stale
marks flatter DD).
Why it fails (the intuition): sells fired only 40k times vs 125k buys —
weekly −5% names mostly keep deteriorating, so the +10% harvest can't keep
up and the book becomes a warehouse of chronic decliners. Weekly reversal
exists as a cross-sectional RANKING edge (gross), not as an absolute
"every dipper is on sale" edge. The spec underperforms doing nothing
(same dollars into SPY) by 2.5-50x depending on delisting kindness.
Actions: NO BIRTH. Implementation was feasible (fractional $5 orders,
free grouped-daily signal, ~1k orders/wk within Alpaca limits) — feasibility
was never the problem; the edge is.
Do-not-retry: fixed-dollar threshold grids that buy absolute weekly
losers market-wide, long-only, any thresholds — dominated by SPY same-flows
in every delisting scenario tested. A respectable follow-up would be the
academic CROSS-SECTIONAL weekly reversal (rank-based, liquid universe,
long-short, quintile spreads) on an UNFILTERED panel — different strategy,
untested here, needs the unfiltered panel build (~4h of free API calls).
2026-08-20 — LEDGER CORRECTION: mp's crash didn't just orphan a short — it INVERTED the book
Owner asked where the Organism's ~$3k went; the site's ledger couldn't
answer because its morning_pullback rows were FICTION. Reconciling against
Alpaca's own fill feed (the definitive record) revealed:
- 8/18: mp shorted 296 @335.50 (never recorded — crash era), crashed
through its flatten window, and the NEXT morning its wiped state met a
long signal: its "buy 289" actually COVERED the forgotten short
(realized −$1,302.35), and its two "profitable long exits" (+$450,
+$1,068 as booked) were actually sell_short fills OPENING a new 296
short — which the 14:50 orphan-reconcile then flattened for −$942.24.
- 8/17's booked −$557.55 was really −$365.80 (signal-priced entries).
True realized, per broker fills: mp −$2,610.39 · ORB −$23.62 · PEAD
+$364.14 → matches account equity to the cent with the live position.
The 4 fictional rows were REPLACED with 7 broker-truth rows, every one
carrying a visible correction note — the correction itself is part of
the public record, not a quiet edit.
Why it cannot recur (all live before this entry): run_once reconciles
state vs broker EVERY tick (an inverted book gets flattened within a
minute, not traded on), entries+exits book ACTUAL fills (commit dd8650f
for PEAD, 8/20 for mp), and the twice-daily invariant auditor alarms on
any broker/state divergence. Today's live test: mp is short 284 @342.25
with state, broker, and site all agreeing.
2026-08-20 (second correction) — the fill-poll never worked; PEAD booked MAY prices
Owner caught it from the numbers: "today +$170 but the ledger sums +$751?"
The account was TRUE; the ledger was fiction — again, one layer deeper:
- The actual-fill poll added 8/19-8/20 compared order status via
`str(status).endswith("filled")` — but Alpaca's enum prints
"OrderStatus.FILLED", so the check NEVER matched and every runner fell
back silently.
- mp's fallback (live 5-min close) was merely imprecise. PEAD's fallback
read the DAILY-BARS CACHE — last refreshed in MAY — and booked CSCO's
exit at its May close (122.57 vs the real 110.25 fill) and AMAT at
522.12 vs 492.49: fantasy +$364 on a real -$408 day.
Corrected to broker-activity truth (recipe notes on every row):
PEAD today -$407.65 (CSCO -125.05, AMAT -282.60) · mp today +$287.13
(entry VWAP 342.106; covers 339.32/+395.62 and 342.87/-108.49).
Cross-check: research equity today +$286.65 ≈ mp realized ✓; core
-$45.42 = PEAD's close-to-close ✓ — pennies.
Fixes shipped: (1) status check is case-insensitive; (2) PEAD's fallback
is a LIVE latest-trade quote, last resort = entry price (a flat record
beats fantasy); (3) equity snapshots FREEZE each date's row at the 16:10
ET close — after-hours repricing no longer leaks into the yesterday
baseline (core's polluted 8/19 row reset to broker last_equity); and the
one that ends this class: (4) the invariant auditor now runs
fills_reconcile — every ledger row of the last 2 days must match a
broker FILL activity within 0.5%, and it runs twice every trading day.
It passed post-correction; it would have flagged CSCO@122.57 pre-market.