docs(research): MTF nesting premise check — facit does NOT nest (different eras)

Cheap descriptive check (no lock) of the user's top-down "sniper" idea (model the
same swing decomposed 1M->1W->1D). Strict nesting on frozen facit: 1W-in-1M 5%,
1D-in-1W 46%. Date-range check is decisive: the TFs are different eras (1W mostly
2017-18, 1M 2020+) -> the corpus was never drawn as nested decompositions and
cannot test the idea; needs new deliberately-nested labels on ONE era (user will
redraw later). Secondary: labeling tool's same_candle_mtf_resolution resolves
1w->1d but not 1M->1w, so monthly anchors stay coarse -> extend RESOLUTION_TIMEFRAME
before the redraw. Matched-null crux recorded as REJECTED (A8/A11/A9). No code.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: JohnCCarter

docs/research_wiki/handoff.md

@@ -27,19 +27,19 @@ append-only trail lives in [log.md](log.md).
 
 **ETH/USD:** blocked until BTC protocol approved.
 
-## Next Step — learning curve DONE (marginally `saturated`); leverage = feature/crux, not labeling
-
-Learning-curve diagnostic **RUN** (Commit 2, 4h k=3): parity OK (`ap_full_facit`=Stage-2 0.0567,
-n_test_pos=65); verdict **`saturated` — but marginally** (means rise monotonically; only the last 5 %
-increment +0.0008 sits inside the f=0.95 ±0.0030 band). Per the locked asymmetric read this is
-**expected** for a ≈1-feature model → **the bottleneck is the feature side, not 4h data.** Grinding
-hundreds more 4h labels is **low-leverage for this model** (marginal gain within noise, ceiling ~0.057)
-— but the gently-rising curve does **not** license "never label". [Results](reviews/btc-fib-selection-learning-learning-curve-results-20260625.md).
-
-**GO-fork (user's call):** (a) **resolve the `cleanliness` crux** — matched-null (A8, NEW lock): is the
-one signal genuine or a detector artifact? the real lever; (b) **A′ decorrelated `exclusivity`** (NEW
-lock, low prior); (c) grow facit anyway via `--by-uncertainty` (now known low-leverage for this model).
-**No code/run/build until a separate explicit GO.**
+## Next Step — top-down "sniper" nesting (user will redraw, not today)
+
+Learning curve DONE → marginally `saturated`; lever is the feature side, not 4h data
+([results](reviews/btc-fib-selection-learning-learning-curve-results-20260625.md)). **Matched-null crux
+REJECTED** (not built): artifact-LOCK A8 gated it to `detector_artifact_supported` (the probe returned
+the opposite), A11 judged it asymmetric-weak, and it answers an A9 out-of-scope question.
+
+New direction (user): **top-down nesting** — model the SAME swing decomposed 1M→1W→1D (parent TF = the
+context that makes a leg "yours"); a few deliberately-nested labels, not hundreds of flat ones.
+**Premise check (frozen facit): does NOT nest** — the TFs are different eras (1W mostly 2017-18 vs 1M
+2020+; strict nest 1W-in-1M 5%, 1D-in-1W 46%). So it needs **new nested labels on ONE era → user
+redraws LATER.** **Prerequisite:** `same_candle_mtf_resolution` covers 1w→1d but not 1M→1w (monthly
+anchors stay coarse) — extend `RESOLUTION_TIMEFRAME` first. **No build until a separate explicit GO.**
 
 ## Recent Changes
 

docs/research_wiki/log.md

@@ -16,6 +16,23 @@ Types: `ingest`, `decision`, `review`, `question`, `maintenance`.
 > Pre-reset (2026-06-10 and earlier): [part 3](log-archive-pre-btc-reset-part3.md) →
 > [part 2](log-archive-pre-btc-reset-part2.md) → [part 1](log-archive-pre-btc-reset-part1.md)
 
+## [2026-06-25] review | Top-down MTF nesting premise check → facit does NOT nest (different eras)
+
+User reframed toward a top-down "sniper" idea: model the SAME swing decomposed 1M→1W→1D (parent TF as
+the relational context that makes a leg "yours"), a few deliberately-nested labels instead of hundreds
+of flat ones — operationalizing "context, not geometry" (`duration` & `magnitude` both ~0 weight;
+candle-count A→B median 5, range 1–272). Cheap premise check on frozen facit (descriptive, no lock):
+strict nesting (child swing within parent A→B window) = **1W-in-1M 5%** (1 of 21), **1D-in-1W 46%**
+(price 85%), 1D-in-1M 25%. Date-range check (advisor-prompted) is decisive: the TFs are **different
+eras** — 1W anchors mostly 2017–18, 1M anchors 2020+; they barely overlap. So the corpus was never
+drawn as nested decompositions and CANNOT test the idea; pursuing it needs **new deliberately-nested
+labels on ONE era** (user redraws LATER, not today), not re-analysis of these 97 HTF legs. Secondary:
+the labeling tool's `same_candle_mtf_resolution` (flag ON) resolves 1w→1d anchors to the exact day but
+has **no 1M→1w entry**, so monthly anchors stay coarse (whole month) — the user's felt imprecision;
+must extend `RESOLUTION_TIMEFRAME` before the redraw. Also recorded: **matched-null crux REJECTED**
+(artifact LOCK A8 trigger not met / A11 asymmetric-weak / A9 out-of-scope). Descriptive only, no claim,
+no code.
+
 ## [2026-06-25] review | Learning-curve diagnostic RUN → marginally `saturated` (4h k=3)
 
 Blind Commit-2 of the [learning-curve LOCK](reviews/btc-fib-selection-learning-learning-curve-lock-20260625.md)