The Forge

Paul Vilevac — bleenq intelligence Working note v1.0 — August 12, 2026. Companion to “The Securitization of Intelligence” (working paper v1.9.3). Drafted with the lab’s research session; every number below traces to the lab’s registered study ledger, worklog, and git history (private).

Disclaimers#

Nothing here is investment advice. The trading described is paper trading on simulated accounts, run under a self-imposed risk budget, for research purposes. Realized figures are simulated fills against real market data and carry all the usual caveats. This note exists to document a method, not a track record.

The claim#

The main paper stakes its credibility on a measurement culture: pre-registered claims, kill conditions written before the data can argue, misses logged at the same prominence as hits. This note documents where that culture was industrialized: a deliberately small momentum day-trading laboratory (“the popper lab”) — a fleet of mechanized trader-personas running on paper accounts against live microcap tape since July 2026.

The lab was never going to manage the real book. Microcap poppers cannot absorb serious capital, and the ledger (reported honestly below) shows the day-trading edge itself is marginal. That was, in the end, the point: the lab’s product is not P&L. It is doctrine, instruments, and calibrated humility, produced at experimental tempo, at tuition prices — $50–150 of simulated risk per trade, dozens of complete experiment cycles in six weeks. The big paper’s method is the lab’s method, scaled up in horizon and capital. This note is the evidence.

What the lab is#

• A fleet of mechanized personas — codified versions of published trading methodologies (Ross Cameron’s first-pullback, Tim Bohen’s dip-and-rip, an SMB scalp, a Larry Williams volatility breakout, one house-original “ride the monster” model) — each running as an independent bot with its own paper account, risk tier, and ledger, on live SIP data.
• A registration regime: every study, threshold change, and arming decision is pre-registered in a numbered ledger (34 entries as of this writing) with the metric, dataset, and adopt/kill rule fixed before the run. Config changes ship only with the evidence linked.
• A forward truth ledger: every fire from every model — including alert-only models that never trade — is scored for forward R nightly, so signal quality is measured independently of execution luck.
• A risk layer built on the assumption that the broker will surprise you — because it did, repeatedly (below).

The doctrine the forge produced#

Each line below was measured into existence — most of them against the operator’s initial intuition.

1. The edge lives in selection, not in the pattern — and not in the exit. The lab’s single most-replicated result. Five exit-management mechanisms have now been tested against the dumb bracket (fixed stop, fixed target) on this book: breakeven ratchets, profit partials, canary exits, trend-break exits (all July 8), and — most recently — a sophisticated multi-timeframe VWAP-gated virtual trailing stop adapted from a system the author architected at Zecco circa 2008 (study #34, August 12: 675 trades, full 1-minute resolution, two overlay variants; both lost to the plain bracket on both cost footings). Five candidates, five burials. A machine-learning autopsy agrees from the other direction: given 226,911 in-trade states, a gradient-boosted exit model spent its capacity re-learning entry-time constants — there was nothing in-trade worth knowing. The one survivor is a narrow conditional guard (an over-extension veto), which sharpens the rule rather than breaking it: exits stay dumb; conditioning belongs at entry.

2. Where the entry edge is, it is the anti-crowd. The lab’s one strongly positive machine-learned result: an entry-quality score trained on 53,124 labeled decision points across three years, validated walk-forward (+0.93R/entry top-decile uplift, 15 of 15 held-out months positive), which distills to a depth-3 decision tree retaining ~104% of the model’s edge — that is, the edge is readable: low cumulative volume so far is good; loud is bad; mid-volume is good only in the morning. The score is strongly anti-correlated with volume (ρ ≈ −0.8) — it is not a momentum proxy; it is a crowding detector. It survived a transfer test onto two live personas’ actual fires (+0.3 to +0.65R uplift, every month positive) and refused to transfer to liquid-universe breakouts — a negative result that maps the edge’s domain boundary precisely. The founding thesis of the whole popper method (“buy the first pullback before the crowd”) was independently re-derived by the machine from the data.

3. Regime is a confirmable configuration, not a forecast — and gates must be mechanical. A mean-reversion sleeve (Connors RSI-2) failed its breadth gate in July and was parked. In August, with the melt-up configuration confirming, it was given a path back — but the path is a pre-registered forward evaluation on its live alert stream with six gates fixed in advance, reviewed on a calendar date, explicitly framed as a regime trade whose entry filter self-extinguishes when uptrends break. The main paper’s Branch-B posture (“trade the configuration by rules; let tells flip you”) is this pattern at macro scale.

4. The broker is part of the system under test. Three incidents, each converted into a standing mechanism within twenty-four hours:

• The naked-stop incident (July 30–Aug 6): tight stops submitted at fill-time below market are silently canceled by the broker; positions rode unprotected into forced liquidations while a circuit breaker, blind to out-of-band closes, re-armed the offender daily. → A stop-guard sweep (any position without live protection is emergency-closed and alarmed) plus honest breaker accounting.
• The friendly-fire incident (Aug 7): the guard’s first live day, it shot seven healthy positions because the broker’s order-listing semantics hide a bracket’s stop leg once the entry fills. → Protection harvesting from the flat order ledger; the fix’s test suite now encodes the broker’s actual semantics, which the original tests had modeled wrong.
• The queued-order incident (Aug 11): a pre-market signal with a one-cent stop sized a 3,749-share position whose day orders queued to the open — a malfunction trade canceled by hand nine minutes before it went live. → Entry guards: no submission outside regular hours; no stop tighter than the backtests’ own label floor. (Bonus lesson, Aug 12: the broker cancels a bracket’s sibling stop when a target partially fills — the guard caught the orphaned remainder three seconds later, its first legitimate save.)

The transferable principle: every safety mechanism was built from a specific observed betrayal, then encoded as a test. The macro program’s spiral-tells checklist and pre-registered re-entry triggers are the same engineering at portfolio scale.

5. Measurement infrastructure is a first-class deliverable. The lab’s instrumentation caught its own bugs: a bucket-tagging audit proved live instrumentation clean (98% live-vs-replay agreement) before concluding a divergent model was genuinely rotten (it was; it was disarmed); a model-attribution flaw that silently swallowed a shadow experiment’s data stream was found because the experiment’s expected cadence was itself a registered prediction. You can only detect a dead data stream if you pre-registered what alive looks like.

The honest ledger#

Through August 12: the fleet’s realized paper P&L is approximately flat-to-negative — the flagship microcap persona’s cumulative era stands slightly negative after giving back an early run; two personas were disarmed by evidence, one retired, one killed for entry mechanics. The profitable legs of the wider paper book are precisely the ones the main paper’s barbell predicts: the liquid volatility-breakout model and the multi-day sleeves (a defensive ETF rotation; an overnight vol-parity allocator reproducing its designer’s Sharpe within noise). The day-trading lab, judged as a business, is a wash. Judged as a research instrument, it produced: one validated ML edge with a readable distillation, five buried exit mechanisms, a domain map of where the edge stops, a hardened execution safety layer, and a working registration culture — in six weeks, for the price of a used laptop’s depreciation.

What transfers to the bigger work, specifically#

The one-sentence version#

The popper lab is where “measure, register, falsify, and let the rules trade” stopped being a philosophy and became muscle memory — and the strongest argument for strictly limiting its strategies’ access to live capital is the lab’s own finding: its value concentrated in the method and the instruments, exactly as a forge’s value is the blades, not the fire.

Sources: the lab’s registered study ledger (studies #1–34), worklog, backtest results record, and repository git history — all held privately; the registration discipline (pre-declared levers, dated entries, misses retained) is the point the note documents, and the artifacts back every figure should verification ever be required.