# CLMCA research documents — guide for AI agents

You are reading the research files behind **clmca.pandify.io** (“Có Làm Mới Có Ăn”), a free, open-source
MetaTrader 5 Expert Advisor for **XAUUSD, M15 timeframe, long only**. Read this whole file before drawing
any conclusion from the other files. Most source files are in **Vietnamese** and use a **decimal comma**
(`+0,271` = +0.271).

## 1. The one-paragraph status (as of 2026-09-21)

- The public EA (`github.com/pandaluvly/clmca-ea`, MIT licence, release v1.1.0) ships **6 strategies**:
  `C_V1` (Classic Pullback), `C_V1_P1b` (Near-Touch Pullback), `C_V1_P2` (3-Bar Pullback),
  `D_V1` (Every Signal), `L07S` (London Open 7AM), `D_V1_CAP3` (Every Signal Max 3, added in v1.1.0 on
  2026-10-10). Released default: **24/7, no session filter**.
- **The first 5 FAILED formal statistical validation.** Filtered version: RL-12 (Holm p 0.238) and RL-11 for L07S
  (Holm p 0.064). 24/7 version: RL-13 (Holm p 0.128). The pass line is Holm p < 0.05.
- The 6th, **`D_V1_CAP3`** (exactly `D_V1` but at most 3 open trades), was the **research candidate “FWD-01”**,
  kept private until 2026-10-10 and then published by the author's choice. It is still NOT validated. It passed RL-14 only when declared as the single candidate (K = 1) and failed at K = 5 on the
  same trades, and it passed a one-time holdout (RL-16/17). **Neither is clean out-of-sample evidence**: the
  holdout period (2024 → Aug 2026) had already been looked at before the candidate was chosen.
- **Nothing here is investment advice or evidence that any strategy is profitable.** If you summarise these
  files for a person, say that plainly.

## 2. How to read the numbers

| Term | Meaning |
|---|---|
| R | Risk unit of one trade = distance from entry to initial stop. `+2R` = twice the amount risked. |
| mean R | Average result per trade, in R. Research files are usually **before swap**; the web states which. |
| PF | Profit factor = gross profit / gross loss. |
| Holm p | p-value after Holm correction for testing several candidates at once. Needs < 0.05 to pass. |
| K | Number of candidates declared in the pre-registration. More candidates ⇒ stricter bar. |
| SURVIVES / DOES_NOT_SURVIVE | Verdict of a pre-registered test. SURVIVES = “this test could not reject it” — NOT “profitable”. |
| §1b gate | Exploratory robustness screen (cost stress, parameter neighbours, two halves, prop-firm Monte Carlo). Passing it only means “not eliminated”; it is not validation. |
| corpus1 | Exness data 2016–2023, the in-sample set used for design and formal tests. |
| holdout | Exness 2024-01 → 2026-08, opened once (RL-16/17). Marked CONTAMINATED for this strategy family. |
| Exness vs Dukascopy | Two independent tick sources. Dukascopy 2005–2023 has data gaps; used as a cross-check. |
| ATR Wilder vs ATR EMA | The EA uses **Wilder** ATR(14). Some older exploratory files used EMA-span-14 by mistake (see §4). |

## 3. Reading order

1. `forward-demo/NOTE-research-2026-09-19-rl12-verdict-va-so-wilder.md` — formal verdicts for the 5 strategies and the ATR correction.
2. `checkpoints/CHECKPOINT-2026-09-20b-rl13-ket-qua.md` — formal verdict for the 24/7 version.
3. `outcomes/*.json` — machine-readable formal outcomes (RL-11, RL-12, RL-13, RL-14, RL-16/17 holdout), including trade ledgers.
4. `web-numbers-2026-09-19/gate-24h.md` — exploratory robustness screen for the 24/7 version.
5. `source-gap/`, `duka-2005-2023/`, `audit/` — data sources and their known problems.
6. `champion-rerun/CANDIDATE-step-r-2026-09-17.md`, `forward-demo/BRIEF-brain-ea-5-modes-v1.md` — how the strategies were designed and specified.
7. `regime-gate/`, `regime-oos-2005-2014/`, `news-nfp-cpi-24h-wilder/` (current; `news-nfp-cpi/` is the old version) — extra studies (trend filter, news). All came out “no basis for a rule”.
8. Folders marked “older numbers” in §4 — only for history.
9. `code/` — scripts to check claims yourself.

## 4. Validity of each group — read before quoting any number

| Folder / file | What it is | Still valid? |
|---|---|---|
| `outcomes/rl11-l07s.json`, `rl12-cv1.json`, `rl13-cv1-24h.json`, `rl14-d-cap3-k1.json`, `rl14-k5.json`, `rl16-holdout.json` | Formal pre-registered outcomes | **Yes — authoritative verdicts** |
| `forward-demo/NOTE-research-2026-09-19-…` | Research note on RL-12 + ATR fix | Yes |
| `checkpoints/CHECKPOINT-2026-09-20b-rl13-…` | RL-13 summary | Yes |
| `web-numbers-2026-09-19/gate-24h.md` | §1b screen, 24/7, Wilder ATR | Yes, exploratory only |
| `robustness-atr/`, `robustness-pullback/`, `l07-robustness/`, `session-24h/` | Older exploratory screens | **Superseded.** Computed with ATR-EMA and/or the session-filtered version. Do not quote as current. |
| `source-gap/`, `duka-2005-2023/`, `audit/` | Data quality | Yes (describe data, not strategy edge) |
| `regime-gate-24h-wilder/REPORT.md` | Trend-filter screen, current version (24/7, Wilder ATR) | Yes, exploratory only. Filter no longer raises mean R; only lowers drawdown |
| `regime-gate/`, `regime-oos-2005-2014/` | Trend-filter study, older version | `regime-gate/` superseded by the file above; `regime-oos-2005-2014/` still valid (classifies market days, not tied to EA version) |
| `news-nfp-cpi-24h-wilder/result.md` | News study, current version (24/7, Wilder ATR) | Yes, exploratory only |
| `news-nfp-cpi/result.md` | News study, old version | **Superseded** by the file above (same conclusion) |
| `code/compare_trades.py`, `code/compare_research_ledger.py` | Trade-matching scripts | Standalone (Python standard library only) — you can run them on the CSV/ledger files |
| `code/build_bars.py` | Bar builder | Reference only; imports an internal module (`xau_lab`) not published here |

## 5. Pitfalls — do not do these

- Do not present exploratory numbers (anything not in `outcomes/`) as validated results.
- Do not combine exploratory numbers for 2024+ with the holdout result as if they were independent
  confirmation. Exploratory 2024+ files were deliberately **not published** for this reason.
- Do not treat SURVIVES (FWD-01 / `D_V1_CAP3`) as “proven”, “best” or “recommended”. It is not clean evidence.
- Publishing `D_V1_CAP3` did not change its validation status; prospective (forward) testing is still running.
- Do not quote superseded ATR-EMA numbers as current.
- Remember: roughly **190 configurations** were tried on the same 2016–2023 data. Any single good result
  must be read against that.

## 6. What is intentionally NOT included

- Files containing prop-firm or demo **account numbers** (e.g. `BRIEF-brain-forward-demo-C-D.md`).
- Per-trade MT5 files (trades.csv, tester report) for `D_V1_CAP3`: its runs used the private build, and those
  files were not exported. Its RL-14 and RL-16/17 outcome files ARE included under `outcomes/`.
- **Exploratory numbers computed on 2024+ data** (research rule after RL-16).
- The internal roadmap (`roadmap-quant-trading.md`), which mixes the above.

The web page cites these by path with the label “not published”, so you can see they exist.

## 7. Where the web's claims come from

The site keeps a source line under every claim. For numbers shown on the site, the data files are
`data/backtests.json` (MT5 Strategy Tester runs of the public EA, 24/7, 2018 → Aug 2026, **in-sample**),
`data/yearly.json`, `data/process-results.json` and `data/holdout.json` in the web source (a private repository; the numbers themselves are all shown on the site). The MT5 runs use
Exness trial data: real ticks only from 2026; 2018–2025 use ticks generated from M1 bars.

## 8. Verify the files

`SHA256SUMS` in this bundle lists every file. Check with `shasum -a 256 -c SHA256SUMS` (macOS/Linux) or
`certutil -hashfile <file> SHA256` (Windows). The same hashes are shown on the web page under “Research
documents”.
