yanivil · Sep 8, 2026 · Sep 8, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -5,6 +5,17 @@ All notable changes to this project are documented here. Format follows
 
 ## [Unreleased]
 
+### Changed (docs: ten years on the index as it was)
+- Tuning page: the tuned profile replayed one calendar year at a time from
+  2016 to 2026 with point-in-time membership (runs 34246882614 to
+  34247722648): +0.23 R on 1629 traded signals, positive in 10 of 11 years,
+  drawdowns of 36 to 41 R inside four years (#108); Wolfe +0.45 R and the
+  best detector in the bad years, cups +0.03 R (#109); bear-regime and
+  high-VIX signals the robust context (#101, #99). The two-year hypotheses
+  on breakout age, reward:risk above 4, low VIX and the SMA band did not
+  survive (#98, #100 closed; #99, #101 rewritten; #102 closed as held). The
+  ledger and the README limitations state the ten-year numbers.
+
 ### Added (point-in-time universe for multi-year replays, #95)
 - `tools/universe_history.py`: index membership as of any past date from the
   git history of the pinned constituent dataset (a bare, blobless clone under

diff --git a/README.md b/README.md
@@ -154,7 +154,7 @@ docs/wiki/                           documentation, mirrored into the GitHub wik
 * Swing points are only recognised 5 bars after they print, so Inverse H&S and Wolfe confirmations can be reported up to 5 bars late.
 * Cup bases must be explained at least as well by a parabola as by a two-legged V; the rule is calibrated on reference shapes, not on market data (see the pattern catalog).
 * The spec profile requires breakout volume (1.4× for cups, 1.3× for H&S) and caps a cup at half its preceding advance, so it confirms far fewer setups than the legacy rules; breakouts without volume appear on the watchlist marked as such.
-* Signal quality is measured by replay, not proven live. `tools/backtest.py` (the `backtest` workflow) runs the scanner walk-forward over past sessions and `tools/evaluate_signals.py` scores the signals the nightly job actually committed. Every rule value was chosen on the year to 2026-09-04; a two-year replay that holds out the year before found the tuned profile at +0.30 R out of sample against +0.28 in the tuning year, on mostly bull markets and today's constituents, and the Wolfe detector negative out of sample. Replays can use the index membership as it was on each past day (`--constituents-asof`), which removes most of the survivorship bias; symbols that no longer trade still have no Yahoo history. The [tuning page](docs/wiki/03-Configuration-and-Tuning.md) states the check every further change must pass.
+* Signal quality is measured by replay, not proven live. `tools/backtest.py` (the `backtest` workflow) runs the scanner walk-forward over past sessions and `tools/evaluate_signals.py` scores the signals the nightly job actually committed. Every rule value was chosen on the year to 2026-09-04. Replayed on the index as it was in each year from 2016 to 2026 (`--constituents-asof`, which removes most of the survivorship bias; symbols that no longer trade still have no Yahoo history), the tuned profile made +0.23 R per trade, positive in ten of eleven years, with drawdowns of 36 to 41 R inside four of those years against an expected yearly gain near 35 R; the Cup & Handle detector was zero-expectancy over the same years. The [tuning page](docs/wiki/03-Configuration-and-Tuning.md) has the tables and the check every further change must pass.
 * The quality score is a gate, not a ranking: its buckets do not order outcomes in replay.
 * Yahoo Finance data is unofficial. Symbols with fewer than 60 bars are skipped and counted in `meta.errors`.
 * There is no persistent price cache: every run re-downloads two years of history for the whole universe.

diff --git a/docs/wiki/03-Configuration-and-Tuning.md b/docs/wiki/03-Configuration-and-Tuning.md
@@ -24,7 +24,7 @@ A third profile, **tuned**, is the spec with the four rules changed that the 202
 | spec | the engine specification as written | 22 | 53 % | +0.14 | 76 % | 10 at 0.00 R | 8 at +0.37 R | 4 at +0.03 R |
 | **tuned** | spec with the four relaxations above (Wolfe rhythm ±45 %) | **177** | **50 %** | **+0.34** | 72 % | 129 at +0.40 R | 28 at +0.13 R | 20 at +0.23 R |
 
-Year-long walk-forward replay, 250 sessions to 2026-09-04, horizon 60 bars, next-open fills, intraday stops (`backtest` workflow runs 33968691768, 33973275310 and 33985398488). Read with the usual caveats: today's constituents only, one year, one regime.
+Year-long walk-forward replay, 250 sessions to 2026-09-04, horizon 60 bars, next-open fills, intraday stops (`backtest` workflow runs 33968691768, 33973275310 and 33985398488). Read with the usual caveats: today's constituents only, one year, one regime. The ten-year replay of the tuned profile on the index as it was each year, +0.23 R per trade with drawdowns of up to 41 R inside a year, is in the out-of-sample section below.
 
 ### Calibrating the review rules (2026-09-06)
 
@@ -63,17 +63,39 @@ Read all of it with four caveats. Both years were mostly bull markets (70 % and
 
 | Adopted value | Chosen on | Confirmed on | Unseen window |
 |---|---|---|---|
-| tuned profile, four relaxations | 63 sessions to 2026-09-04, run 33968787516 | 250 sessions to 2026-09-04, run 33973275310 | held: +0.30 R, run 34186512341 |
-| `WW_TIME_SYM_TOL` 0.45 | 250 sessions to 2026-09-04, run 33985398488 | none | failed: Wolfe −0.16 R on 12 traded rows (#102) |
+| tuned profile, four relaxations | 63 sessions to 2026-09-04, run 33968787516 | 250 sessions to 2026-09-04, run 33973275310 | held: +0.30 R on the unseen year (run 34186512341); +0.23 R over 2016-2026 on the point-in-time index, positive in 10 of 11 years (the eleven runs below) |
+| `WW_TIME_SYM_TOL` 0.45 | 250 sessions to 2026-09-04, run 33985398488 | none | failed on 12 unseen-year rows, then held over ten years: Wolfe +0.45 R on 228 signals, positive in 8 of 11 (#102); the 0.30 alternative has not been replayed over the same years |
 | `MIN_REWARD_RISK` 1.0, `MAX_WAIT_BARS` 60 | 250 sessions to 2026-09-04, runs 34023518929 to 34023527178 | none on their own | untested alone; the profile that includes them held |
 | `CUP_TRIGGER` handle_high, kept | 250 sessions, run 33985391832 against 33973275310 | none | untested |
 | `WATCH_PROXIMITY` 5 % | the 2026-09-04 report, 10 of 17 rows lost in a day | none | a reporting choice, not a replay question |
 
-**Hypotheses under test**, each needing the rule above and most the multi-year replay of #95: #98 breakout age at first report, #99 the SMA50-to-SMA200 band, #100 reward:risk of 4 or more, #101 VIX below 15, #102 Wolfe out of sample. Deferred: #97, a per-signal probability model, premature at about 120 resolved trades a year.
+### Ten years on the index as it was (2026-09-08)
+
+Eleven runs, one calendar year each from 2016 to 2026 (to 2026-09-04), tuned profile, point-in-time membership (`asof=true`, see [Testing and Contributing](04-Testing-and-Contributing.md)), `bars=500`, `horizon=60`, `grid=false`: runs 34246882614, 34246967266, 34247049007, 34247134013, 34247217621, 34247300839, 34247387000, 34247471926, 34247553575, 34247636362 and 34247722648 in that order. Coverage is the share of that year's members with Yahoo history; the rest, delisted or renamed since, is the survivorship bias that remains.
+
+| Year | Coverage | Signals | Hit rate | Mean R | 95 % CI, month blocks | Max drawdown |
+|---|---|---|---|---|---|---|
+| 2016 | 77 % | 111 | 45 % | +0.39 | [−0.11, +0.76] | −17 R |
+| 2017 | 80 % | 136 | 40 % | +0.60 | [+0.16, +1.07] | −15 R |
+| 2018 | 80 % | 153 | 26 % | −0.14 | [−0.52, +0.15] | −41 R |
+| 2019 | 84 % | 240 | 32 % | +0.26 | [−0.22, +0.73] | −37 R |
+| 2020 | 84 % | 124 | 32 % | +0.15 | [−0.17, +0.56] | −17 R |
+| 2021 | 88 % | 157 | 35 % | +0.16 | [−0.09, +0.45] | −19 R |
+| 2022 | 91 % | 176 | 29 % | +0.03 | [−0.50, +0.43] | −38 R |
+| 2023 | 94 % | 169 | 43 % | +0.40 | [−0.24, +0.92] | −36 R |
+| 2024 | 95 % | 133 | 35 % | +0.16 | [−0.25, +0.58] | −19 R |
+| 2025 | 97 % | 173 | 38 % | +0.27 | [−0.14, +0.60] | −15 R |
+| 2026 to 09-04 | 97 % | 118 | 44 % | +0.24 | [+0.01, +0.45] | −10 R |
+
+Pooled, weighting each year by its traded signals: 1629 traded signals, +0.23 R, 36 % hit rate, positive in 10 of 11 years. Per pattern: inverse H&S +0.23 R on 1096 signals, positive in 9 years and negative in the two bear years 2018 and 2022; Bullish Wolfe Wave +0.45 R on 228, positive in 8 years and the best detector in 2018, 2019 and 2022; Cup & Handle +0.03 R on 305 with an 18 % hit rate, positive in 5 years, 2017 alone supplying the profit (#109). By SPY regime at the scan day: bull +0.16 R on 1190, neutral +0.21 on 213, bear +0.59 on 226 (positive in 5 of the 6 years with bear sessions). VIX above 25 ran +0.50 on 142 in 8 of 8 years; a stock whose SMA50 sat more than 5 % under its SMA200 ran +0.35 on 499 against +0.15 to +0.20 elsewhere (#99, #101). Score buckets from 60-69 to 90-100 ran +0.27, +0.22, +0.22 and +0.19: the hit rate rises with the score and the mean R does not, so the score gates and does not rank (#82). Drawdowns inside the year reached 36 to 41 R in 2018, 2019, 2022 and 2023 against an expected yearly gain near 35 R (#108).
+
+What the ten years overturned from the two-year window, each a lesson in what one exploratory pass over two bull years produces: fresh breakouts are not better (age 0 at first report +0.10 R against +0.42 for age 4 to 8; #98 closed), reward:risk above 4 is not bad (+0.45 R with a 16 % hit rate; #100 closed), a VIX below 15 is not bad (+0.21 R pooled, and 2017 at a VIX near 10 was the best year; #101 rewritten), the SMA band U-shape is one-sided (#99 rewritten), and the Wolfe verdict on 12 out-of-sample trades was noise (#102 closed as held). What survived: the tuned profile's expectancy, the score's role as a gate, and the reversal-after-washout context.
+
+**Open questions:** #99 and #101 (the deep-down-trend and bear-regime contexts, information first, no rule), #108 (the drawdowns), #109 (cups), and #97 (a per-signal probability model, whose first revisit condition the ten years now meet).
 
 ### Market context (2026-09-08)
 
-The report header and every backtest row carry the SPY regime (close and SMA50 against the SMA200), the VIX and the breadth of the universe, and the backtest summaries add a per-regime slice with VIX and breadth among the feature buckets. In the two-year local replay, tuned signals scanned in a non-bull SPY regime ran +0.68 R on 75 rows against +0.17 on 237 bull-regime rows, and rows scanned at a VIX below 15 were negative in both years (−0.55 on 16, −0.19 on 19). The non-bull rows all came from March to June 2025, one episode, so nothing gates on the context; issues #93 and #101 track the evidence and a rule would need the multi-year replay of #95.
+The report header and every backtest row carry the SPY regime (close and SMA50 against the SMA200), the VIX and the breadth of the universe, and the backtest summaries add a per-regime slice with VIX and breadth among the feature buckets. Over the ten years on the point-in-time index (the section above), signals scanned in a SPY bear regime ran +0.59 R on 226 against +0.16 on 1190 in bull regimes, positive in five of the six years with bear sessions, and signals scanned at a VIX above 25 ran +0.50 R in eight of eight years; a VIX below 15, negative in the two-year window, ran +0.21 R pooled. Nothing gates on the context; #101 tracks how the report should present it.
 
 **The nightly scan runs `--profile tuned`**, chosen on this evidence on 2026-09-05; the default in code stays `spec`. Change the flag in `.github/workflows/daily-scan.yml` to switch. The cup entry stays at the handle peak: a rim-B entry replayed at 19 cup signals and +0.17 R against 20 and +0.23 R. The grid also shows a close-based stop (exit on the first close at or below the stop) lifting tuned's mean R to +0.35 at a 52 % hit rate; the report's stop level is unchanged, that is an execution choice.
 

diff --git a/docs/wiki/04-Testing-and-Contributing.md b/docs/wiki/04-Testing-and-Contributing.md
@@ -87,7 +87,7 @@ Two more tables follow each summary. **Excursions by horizon** gives, overall an
 
 **Out-of-sample check.** `--split YYYY-MM-DD` prints every table three times: all sessions, the sessions before the date, and the sessions from it. A rule chosen on one window is judged on the other; how the windows are used is the protocol in [Configuration and Tuning](03-Configuration-and-Tuning.md). `--bars 500` makes each scan see only its last 500 bars, which is what the nightly's `2y` download gives it, so a `--period 5y --days 500` replay reproduces two years of nightly runs rather than scans with ever-longer histories. Caveats: today's constituents only (survivorship bias), the last `horizon` sessions are still open, and both replayed years to 2026-09 were mostly bull markets.
 
-**Point-in-time universe.** By default a replay scans today's constituents, so symbols that left the index during the window are missing and symbols that joined are present before they joined: survivorship bias, which flatters bullish patterns (the dataset's 2016 snapshot holds 179 symbols that are not members today). `--constituents-asof` (workflow input `asof=true`) reconstructs the membership as of each scan day from the git history of the constituent dataset the scanner already pins: `tools/universe_history.py` makes a bare, blobless clone of `datasets/s-and-p-500-companies` under `.cache/` (under 1 MB, refreshed by a fetch on later runs), reads `data/constituents.csv` at the last commit on or before each day, downloads the union of members over the window, and scans a symbol only on the days it was a member. The report opens with a coverage table: members per year and how many of them have Yahoo history. Two limits remain. The dataset records a change some days after the index does, with about weekly commits since 2020 but one a year in 2017-2019. And a symbol that no longer trades, or was renamed, has no Yahoo history, so the replay still cannot trade it; the coverage table is the measure of the survivorship bias that is left (a transient Yahoo miss counts the same way, so compare the share across runs before reading much into one). The first such replay, 2022 on 531 members of which 481 had history, ran the tuned profile at +0.03 R on 175 signals with a −38 R drawdown: the bear year the two-year replay never contained. `--end YYYY-MM-DD` (input `end`) replays the `--days` sessions on or before a date, so a past year is one dispatch, and `grid=false` skips the variants and the other profile's pass when only the primary profile matters. The download must reach the window: `period=10y` for anything from 2016 on.
+**Point-in-time universe.** By default a replay scans today's constituents, so symbols that left the index during the window are missing and symbols that joined are present before they joined: survivorship bias, which flatters bullish patterns (the dataset's 2016 snapshot holds 179 symbols that are not members today). `--constituents-asof` (workflow input `asof=true`) reconstructs the membership as of each scan day from the git history of the constituent dataset the scanner already pins: `tools/universe_history.py` makes a bare, blobless clone of `datasets/s-and-p-500-companies` under `.cache/` (under 1 MB, refreshed by a fetch on later runs), reads `data/constituents.csv` at the last commit on or before each day, downloads the union of members over the window, and scans a symbol only on the days it was a member. The report opens with a coverage table: members per year and how many of them have Yahoo history. Two limits remain. The dataset records a change some days after the index does, with about weekly commits since 2020 but one a year in 2017-2019. And a symbol that no longer trades, or was renamed, has no Yahoo history, so the replay still cannot trade it; the coverage table is the measure of the survivorship bias that is left (a transient Yahoo miss counts the same way, so compare the share across runs before reading much into one). Eleven such runs, one per calendar year from 2016 to 2026, are the multi-year evidence on the [tuning page](03-Configuration-and-Tuning.md): the tuned profile at +0.23 R per trade, positive in ten of eleven years, with drawdowns of up to 41 R inside a year. `--end YYYY-MM-DD` (input `end`) replays the `--days` sessions on or before a date, so a past year is one dispatch, and `grid=false` skips the variants and the other profile's pass when only the primary profile matters. The download must reach the window: `period=10y` for anything from 2016 on.
 
 A runner replays about 5 seconds per session per profile over the full index, and the workflow runs `--grid` unless `grid=false`, which adds the other profile's pass: