[HISTORICAL RECORD. Label added 2026-09-22 for publication.]
This is our own original protocol for this test, as filed in our lab on 2026-07-14 (commit id withheld because the repository is private).
Below the line it is shown unchanged, except one internal reference, redacted and marked
[internal note reference removed].

Timing: the protocol and both result files first appear together in that same commit. Its
"pre-registration" wording is therefore our own record, not an independently sealed lock. Only the
daily clock was pre-registered; the hourly clock was added later.

Corrections to this text:
- His "2,000" is his career count of bots he backtested, not the width of the search in his video, which the video does not state.
- We ran the search method his video demonstrates on a coin-flip signal: not his exact procedure (the risk skeleton follows his only in shape) and not his bot.
- The words it quotes as his ("verified identical on TradingView") paraphrase his video; they are not his exact words.

Full report: i05c-noise-search.html
----------------------------------------------------------------------------------------------------
# I05c — "The AI that data-mines 2,000 strategies" — Noise-Survives-the-Search receipt
## Pre-registration (write the bar BEFORE the run — GT B04 doctrine)

**Specimen:** YouTube `BPFwaD0CgZ8` (David / David Tech). Claim: ChatGPT + an MCP server
searched ~2,000 backtested configs and surfaced a gold **1H** strategy at **1,669% net /
28% maxDD / 31% WR / 456 trades / PF 1.89**, "verified identical on TradingView," then hooked
to Telegram for a 30-day forward test.

**Thesis under test (ours):** the headline is a **multiple-comparisons artifact**, not an edge.
A search that backtests ~2,000 configs and reports the single best will surface a spectacular
result *even when the entry signal carries zero information* — because the winner is the extreme
right tail of a luck distribution, not a discovered edge. His forward test does not rescue it:
forward-testing the survivor of a 2,000-wide search never pays down the data-mining.

**Why this stands in for his gold strat:** his literal instrument (spot gold 1H, Asia-session
filter) is not in our survivorship-free store (US equity + ETF only; gold = GLD, market-hours,
no Asia session). The lesson is about the **search**, not the instrument — so we reproduce the
*procedure* on real data with a provably-null signal. If a null signal + his procedure
manufactures his headline, the headline is not evidence of an edge.

### Design (provably-null signal on real data)
- **Data:** `SPY` daily OHLC, survivorship-free store, 2004-01-02 .. 2024-12-31 (real prices;
  relatable; upward drift = the *friendly* case, which only makes the point stronger).
- **Split:** IN-SAMPLE = 2004..2014 (the search happens here). OUT-OF-SAMPLE = 2015..2024
  (used only to validate the winner; the search never sees it).
- **Risk skeleton = his, faithfully:** long-only, one position at a time, ATR(14) stop at
  `k·ATR`, target at `R=3.8 × risk` (his R:R 3.8), fixed-fractional sizing (risk `f=2%` of
  equity per trade → compounds like a TradingView % boilerplate). **Gaps honored at the open**
  (a gap through the stop loses MORE than 1R — no flattering fills).
- **Entry SIGNAL = the null:** `Bernoulli(p_entry)`, seeded per config, **independent of price
  by construction.** Stands in for "some indicator combo with some parameters" — the point is it
  has no predictive content.
- **Search grid (~2,048 configs, his ~2,000):** `seed ∈ 128 values × p_entry ∈
  {0.02,0.04,0.06,0.10} × k_atr ∈ {1.0,1.5,2.0,2.5}`. Rank by IN-SAMPLE net return; **report the
  single best** (his exact procedure).
- **Fair null = buy-and-hold** (`[internal note reference removed]`). Because SPY
  rose in both windows, random longs earn positive *beta*; we therefore judge the winner by
  **alpha over B&H**, not raw return, so market exposure is not mistaken for the search "working."

### Pre-registered predictions (what proves the thesis)
1. **Best-of-2,048 IS net% is spectacular** — same order of magnitude as his 1,669% (at minimum
   >> the median config), with WR ~20–35% and PF > 1.5: a config that *looks* like a real edge,
   and a large positive **IS alpha over B&H**.
2. **The winner's OUT-OF-SAMPLE alpha over B&H collapses to ~0 or negative** — because random
   entries carry no signal into unseen data. **THE KILL.** (Raw OOS return may stay positive from
   beta; the *alpha* is what must vanish.)
3. **IS rank does not predict OOS** — across the top-N IS configs, rank-correlation(IS net, OOS
   net) ≈ 0; and the share of all configs that beat B&H OOS is ~coin-flip.
4. **Friction bites** — with a realistic per-trade cost the median config's alpha is negative and
   the winner shrinks.

### Anti-hypocrisy guard (we are not David)
Report the **full distribution** + the winner's **OOS** + **friction** regardless of outcome. No
cherry-picking our own demo. Seeded RNG ⇒ reproducible. If prediction 2 FAILS (winner holds its
alpha OOS), that signals a **bug** — a null signal cannot have OOS edge — investigate, do not
narrate.