Skip to content

Commit 9c40e64

Browse files
JohnCCarterclaude
andcommitted
docs(research): Fib selection-learning campaign CHECKPOINT (consolidation, no new claim)
Locks current state of the selection-learning line after five committed runs: Stage-2 headline → prominence-family sensitivity → k-sweep → W-gap (no_causal_gap) → Stage-1 (no_pivot_signal_above_prominence). KNOW (4h, frozen eight): a modest, OOS, live-available, buffer-stable, baseline-robust selection correlate carried by leg `cleanliness`, living in the leg gestalt (not the lone pivot, not detection coverage ~0.90), at low absolute agreement (AP ~0.057 vs ~0.83 ceiling) — not a reproduction. CRUX (open): is `cleanliness` a genuine selection signal or a detection/anchoring artifact? The pipeline is detector-conditioned; nothing run so far can resolve it. Loose end: set-level `exclusivity` (k*=3) was specced but never built. Frames next-step choice A (exclusivity / artifact diagnostic) / B (detector-independent anchor-probe) / C (pause + theory) — none started. Docs-only, no code/run/implementation, no new positive claim. Handoff pointer added for wiki-lint reachability. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
1 parent b515b08 commit 9c40e64

2 files changed

Lines changed: 95 additions & 0 deletions

File tree

docs/research_wiki/handoff.md

Lines changed: 7 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -126,6 +126,13 @@ legs/ranges* (labels = facit; **no edge/behaviour/backtest/PnL/Genesis/auto-fib
126126
[Results](reviews/btc-fib-selection-learning-stage1-results-20260624.md); summary + `cells/*.json`
127127
gitignored/regenerable. Re-run (deterministic, frozen data, **no `--refresh`**):
128128
`PYTHONUNBUFFERED=1 uv run --no-sync python -u -m fibengine.research.selection_learning_stage1 --stage1`.
129+
- **2026-06-24 Fib SELECTION-LEARNING campaign CHECKPOINT (docs-only, no new claim).** Consolidates
130+
the five committed runs (Stage-2 → prominence-family → k-sweep → W-gap `no_causal_gap` → Stage-1
131+
`no_pivot_signal_above_prominence`): what we KNOW (modest, OOS, live-available, buffer-stable,
132+
baseline-robust `cleanliness` lead in the leg gestalt; low absolute AP; 4h-only powered) and the
133+
single open CRUX (is `cleanliness` a genuine signal or a **detector/anchoring artifact**?). Frames
134+
the next-step choice A (exclusivity / artifact diagnostic) / B (detector-independent anchor-probe) /
135+
C (pause + theory) — **none started**. [Checkpoint](reviews/btc-fib-selection-learning-checkpoint-20260624.md).
129136

130137
**Next work requires a separate explicit GO. No W/gap, no Stage 1, no new sensitivity, and no Genesis
131138
may be started automatically.** Parked (test-only, separate GO): lock the facit-discipline refusal
Lines changed: 88 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,88 @@
1+
# BTC Fib Selection-Learning — CAMPAIGN CHECKPOINT (2026-06-24)
2+
3+
**Lean Fib Research. Research-only. Selection learning — NOT a behaviour/edge claim, no
4+
backtest/PnL, no Genesis, no auto-fib-as-truth, no label mutation.** This doc **locks what the
5+
selection-learning line has actually established** across its five committed runs, names the **single
6+
open crux**, and makes the **next-step choice (A / B / C) well-posed**. It starts **no** new track,
7+
authorises **no** code, run, or build, and adds **no** new positive claim. Decision needs a separate
8+
explicit GO.
9+
10+
## Scope of this line (unchanged since the 2026-06-17 prereg)
11+
12+
> *Can a model reproduce **how the human selects** swings/ranges* — labels = facit, **no
13+
> edge/behaviour/PnL/backtest/Genesis/auto-fib claim**. Stage-2 (leg/range gestalt) = headline target;
14+
> Stage-1 (per-pivot) = diagnostic floor. One primary cell `k=3`, coverage ceiling, **4h is the only
15+
> adequately powered TF** (1M/1w/1d underpowered throughout — **context, not refuted**).
16+
17+
## Verdict chain (committed, blind-locked rules applied verbatim)
18+
19+
| # | Study (date) | Locked verdict | Headline number (4h, powered) | Commit |
20+
|---|--------------|----------------|-------------------------------|--------|
21+
| 1 | Stage-2 headline (06-18) | modest single-feature lead | lift **+0.052**, CI [0.023, 0.120] vs magnitude | `ea6c2ea` |
22+
| 2 | Prominence-family sensitivity (06-18) | **`survives_prominence_family`** | lift survives vs summed (+0.043) AND max (+0.049) prominence | `ea6c2ea` |
23+
| 3 | k-sweep {0,3,6,12} (06-18) | **`k_stable_live_selection_signal`** | k=3/6/12 all survive; k=0 degenerate | `ea6c2ea` |
24+
| 4 | W-gap causal-availability (06-23) | **`no_causal_gap`** | gap(k=3) **−0.0045**, CI [−0.070, +0.031] incl. 0 | `b515b08``61c41d1` |
25+
| 5 | Stage-1 per-pivot (06-24) | **`no_pivot_signal_above_prominence`** | recall **0.902**; ranking lift +0.0228, CI [−0.035, +0.079] incl. 0 | `b515b08` |
26+
27+
## What we KNOW (positive, scoped to 4h + the frozen eight features)
28+
29+
1. **There is a real, modest selection correlate.** On the one powered cell the model out-ranks every
30+
§6 baseline OOS (magnitude + summed/max prominence), CI excludes 0, `p_one_sided(lift≤0)=0/2000`.
31+
2. **It is carried almost entirely by ONE leg-level feature — `cleanliness`** (net move ÷ path,
32+
standardized weight ~0.20, ≈3× the next term). Human-marked legs are **cleaner / more efficient**.
33+
It is **not** a multi-feature reproduction.
34+
3. **It is not hindsight** (W-gap `no_causal_gap`): a bounded 180-bar retrospective view buys no
35+
selection info the live view at `k=3` lacks — `cleanliness` is **live-available**.
36+
4. **It is not a buffer artifact** (`k_stable_live_selection_signal`): stable across k∈{3,6,12}.
37+
5. **It is not a coverage/detection failure** (Stage-1 recall ~0.90): the human's anchors **are** in
38+
the detector's pivot universe — the Stage-2 ceiling is **not** a detection problem.
39+
6. **It does not live in the lone pivot** (Stage-1 `no_pivot_signal_above_prominence`): per-pivot
40+
features add nothing over prominence. **The signal lives in the leg/range GESTALT**, and the one
41+
feature carrying it (`cleanliness`) is structurally a **leg-level** quantity.
42+
7. **Absolute agreement is LOW** — AP ~0.057–0.066 against a ~0.83 reachability ceiling. **The human
43+
is not "reproduced."**
44+
45+
> One-line state: *On 4h there is a modest, OOS, live-available, buffer-stable, baseline-robust
46+
> selection correlate, carried by leg `cleanliness` and living in the leg gestalt (not the lone pivot,
47+
> not detection coverage) — but at low absolute agreement, and on one powered TF only.*
48+
49+
## What we DON'T know — the single open CRUX
50+
51+
**Is the `cleanliness` lead a genuine human-selection signal, or a detection / anchoring artifact?**
52+
The entire pipeline is **conditioned on the detector's pivot universe**, and `cleanliness` is computed
53+
on **detector-defined legs**. If the detector preferentially surfaces or anchors clean/efficient legs,
54+
then "human legs are cleaner" is partly **mechanical** — baked in by candidate generation, not a fact
55+
about human choice. This question has been flagged at **every** stage (06-18 sensitivity, k-sweep,
56+
W-gap, Stage-1) and **deliberately left OPEN**; nothing run so far can resolve it, because every run
57+
lives inside the same detector frame.
58+
59+
**Secondary loose end:** set-level **`exclusivity`** (`k*=3`) was specced in the
60+
[§12 addendum](btc-fib-selection-learning-addendum-20260618.md) but the Stage-2 live whitelist actually
61+
built was `{magnitude, cleanliness, duration, prominence, structure_alignment}`**exclusivity was
62+
never implemented or run.** It is an unfinished feature, not a result.
63+
64+
## The next-step choice — A / B / C (all roads lead to the crux)
65+
66+
| | Track | What it does | Decides the crux? | Main risk |
67+
|--|-------|--------------|:-----------------:|-----------|
68+
| **A** | set-level exclusivity / `cleanliness`-artifact diagnostic | stays in the detector frame: (i) build the unbuilt `exclusivity` feature, does `cleanliness` survive its inclusion; (ii) a diagnostic separating "human prefers clean legs" from "detector surfaces clean legs" | **partly** | within-detector-frame circularity may be **irreducible** — a within-frame test may not fully break the circle |
69+
| **B** | detector-independent anchor-probe | removes the detector as candidate-generator and tests whether the anchor / `cleanliness` signal survives **without** the suspected source of circularity | **most directly** | requires **inventing a detector-independent frame** → new arbitrary choices (validity-over-convenience hazard); MUST be locked blind |
70+
| **C** | pause + write "current theory of human fib selection" | synthesis-only; consolidates KNOW/crux into a stated theory, defers the empirical crux | **no** (defers) | leaves the artifact question open; risks theorizing past the evidence |
71+
72+
- **A and B are two angles on the SAME crux** (cleanliness-as-artifact / detector-circularity); **C
73+
defers it.** Whichever empirical track is chosen, its verdict rule must be **locked blind before any
74+
build** (two-commit gate, as with W-gap and Stage-1), and must respect validity-over-convenience —
75+
no quietly-chosen control or frame.
76+
- **Tentative lean (not a decision — the GO is the user's next turn):** the crux *is* the artifact
77+
question, and only **B** attacks its suspected source head-on; **A**'s within-frame test may not
78+
break the circularity and **C** defers it. So **B, but only behind a tight blind lock** — and if a
79+
detector-independent frame cannot be defined without arbitrary convenience choices, **fall back to
80+
A**'s narrow cleanliness-artifact diagnostic. Recommend: **this checkpoint first, then A or B.**
81+
82+
## Non-claims (binding — carried from every prior lock)
83+
84+
Not a reproduction of human selection. **No edge / behaviour / PnL / backtest / strategy claim.** The
85+
`cleanliness`-as-artifact question is **OPEN** and this doc does **not** resolve it. Underpowered TFs
86+
are context, not refuted. No Genesis, no auto-fib-as-truth, no label/corpus mutation, no 1H, no ETH.
87+
This checkpoint is **descriptive consolidation only** — it adds no new positive claim and starts no
88+
track.

0 commit comments

Comments
 (0)