research(experiments): the 08-13 restatement restated on v0.13.3 (#737, #738)

[eaitbrahim]

Closes #737. Closes #738.

The verdict

0 of 240 configurations clear n ≥ 100 ∧ profit_factor > 1.0 at the taker rate this account actually pays, priced per product. Zero at the maker rate. 960 trials — five shipped signal families × 24 assets × 3 fees × 2 slippage regimes, plus the transfer check and the frequency axis. No argmax in Arm A, zero free parameters.

The 08-13 null holds across a population 2.7× larger, under a strictly more expensive cost model, on a corpus 23 days longer.

The finding, which is not that number

The engine never moved the 08-13 numbers. 565 new candles did.

The deployment DB had grown since 08-13, so a naive re-run would have absorbed that and reported it as an engine effect. A truncated-corpus control separates them:

cell	08-13 printed	today's engine, 08-13 corpus	today's engine, full corpus	bars added
turtle XRP-USD	157 / 1.223	157 / 1.223 — identical	157 / 1.223	+0
turtle PAXG-USDT	238 / 1.145	238 / 1.145 — identical	238 / 1.145	+0
turtle FET-USD	269 / 1.206	269 / 1.206 — identical	272 / 1.207	+565
turtle CRV-USD	260 / 1.017	260 / 1.018	267 / 1.023	+565
turtle ZEC-USD	268 / 1.442	271 / 1.411	274 / 1.700	+565
pullback ZEC-USD	170 / 1.044	171 / 1.025	174 / 1.695	+565

Plus a fourth identity from Arm A: rsi_meanrev at its shipped oversold=20 reproduces §4's anchor to four decimals — median gross 1.1251 at median n=42.

#442's gap-through-stop exit fill, its ratchet policy, and #523's re-derived cap are all inert for these rules. Recorded because the honest expectation was the opposite, and because a conservative-only correction that turns out to be a no-op is worth knowing before the next one is deferred on the same reasoning.

The deletion test, run in reverse on real data

§2 of 08-13 argued ZEC's edge was tail-carried, evidencing it by deleting three trades. Adding three moved pullback_continuation on ZEC-USD from gross 1.025 → 1.695. A profit factor a fortnight of ordinary data moves by +0.67 on 174 trades is measuring three trades, not an edge.

And per-product pricing deletes the episode anyway. ZEC is the fourth-thinnest name in the universe — 107.3bp, 21.5× the floor, on $1.08M/day. Priced at its own liquidity its gross falls to 0.889: a loss before any fee.

Corrections of record against 08-13

None of these rewrite it — it keeps its numbers and gains a forward pointer (#247).

1. §6's scope claim is false. Three families were measured; five ship.
2. §3's "all seven die at the maker rate" was an artifact of the flat floor. Priced per product, six of the seven are dead at zero fee. So is the WLD/TON observation (WLD 1.061 → 0.626 at 120.9bp; TON 0.774 → 0.317 at the cap).
3. §3.2's 0.034 transfer gap re-measures at 0.117–0.136. Stably unprofitable holds; not overfit is weaker than stated.
4. The 08-20 note's break-even row is superseded. Priced at ψ=5bp, the floor 0 of 24 assets reach. Re-derived at the measured median ψ=52.3bp with the note's own arithmetic: p_be 20.52% against a reconstructed 14.9% win rate — 5.62 points underwater inside the allowance. Rail 14 is still worth 14.3 points of break-even, more than any rule change in this directory; it is the boundary between decisively and clearly negative, not between negative and break-even.

Independent reproduction

Arm A re-derives the 09-01 per-product restatement from scratch, four days later on a longer corpus: median PF 0.3120 flat → 0.2192 per-product, delta −0.0928 against its −0.090, every per-family figure within 0.005.

The one cell above 1.0

Arm B's turtle_breakout on ZEC-USD, per-product taker: PF 1.0275 at n=96 — four trades below the admission floor, on the asset already known to be regime-bound and tail-carried. Reported because it exists, not because it means anything.

Also here (#738)

docs/experiments/README.md was missing six records and had no 2026-09 section, while its own preamble calls it a complete index. Verified mechanically: every link resolves, every .md/.py has an entry.

Verification

• keel trials verify: chain intact at 94 rows; ledger diff is 1 insertion, 0 deletions
• tests/test_fee_reality_block.py 16 passed · tests/research/test_ledger.py 12 passed
• ruff check clean on both drivers
• The committed control driver was re-run from a clean invocation and reproduces its table exactly

Not changed

Nothing promoted, no rule row added, no config, allowlist or shipped parameter touched.

🤖 Generated with Claude Code

https://claude.ai/code/session_01EXz13qp1UM3pBa6BqsRqvC