docs(sp13): v2 spec + plan after critical review
Spec v2 supersedes v1 with five major fixes from the critical review: - Drop alpha-vs-benchmark reward (mathematically tautological — Long trades had alpha = -costs always against always-long benchmark) - Split Phase 0 into 0a (Hold-only, clean test of user hypothesis) + 0b (aux amplification probe, only if 0a partial), avoiding the v1 confounded experiment - Bound direction-skill bonus relative to |alpha| (cap_ratio × |alpha|) to prevent reward gaming on small-alpha correct-direction losers; spec adds 8-quadrant worked-example matrix verifying the no-negative-EV invariant - Replace 5-epoch absolute-threshold gate with 10-epoch trajectory criterion to avoid false-negatives from aux head underconvergence - Aux_w controller adds dual-EMA stagnation detector (decays toward base when no improvement) — prevents permanent destabilization of Q-head in data-limited case Plan v2 mirrors spec changes: - Phase 0a (Hold elimination + dir_acc instrumentation, ~300 LOC, 1 atomic commit) - Phase 0b (aux_w controller replacement, conditional, ~80 LOC) - Layer B (aux head regression -> binary classification, ~120 LOC) - Layer C (skill bonus + luck discount with calibrated bounds, ~150 LOC) - Layer D (30-epoch validation + 3 new pearls) ISV slot allocation [372..380): drops slot 371 (BENCHMARK_PNL_CUMULATIVE), adds 374 (AUX_DIR_ACC_LONG_EMA — stagnation), 379 (SKILL_BONUS_CAP_RATIO). Plan integrates Explore-agent touch-list (50-70 sites) with corrected enum ordering: Short=0, Long=1, Flat=2 (preserves codebase Short-first convention, avoids 30+ stale-comment churn). Adds direction-bias signal handling at experience_kernels.cu:5159 (Hold's 0.5 softening gate vanishes -> [1, 1, 0]). Three new pearls planned for Layer D close-out: - pearl_redefine_success_for_predictive_skill - pearl_skill_bonus_must_be_alpha_bounded (calibration lesson) - pearl_reward_quadrant_audit_required (meta-pearl) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
jgrusewski
File diff suppressed because it is too large
Load Diff
@@ -0,0 +1,386 @@
|
||||
# SP13 — Redefine Success for Predictive Skill (v2)
|
||||
|
||||
**Date**: 2026-05-04 (v2 — supersedes v1 in-place)
|
||||
**Branch**: continues `sp11-reward-as-controlled-subsystem` (single coherent architectural lineage)
|
||||
**Builds on**: SP11 (cap stability) + SP12 (event-driven density) + asymmetric loss aversion
|
||||
**Anchor pearls**:
|
||||
- `pearl_audit_unboundedness_for_implicit_asymmetry`
|
||||
- `pearl_event_driven_reward_density_alignment`
|
||||
- (NEW) `pearl_redefine_success_for_predictive_skill` — to be written post-validation
|
||||
- (NEW) `pearl_skill_bonus_must_be_alpha_bounded` — to be written post-validation
|
||||
- (NEW) `pearl_reward_quadrant_audit_required` — meta-pearl from v1 review
|
||||
|
||||
## v1 → v2 changes (reasons here for traceability)
|
||||
|
||||
| Issue | v1 | v2 |
|
||||
|---|---|---|
|
||||
| Alpha-vs-benchmark math broken — Long trades have alpha = −costs always (tautology against always-long benchmark) | reward = realized_pnl − benchmark_pnl_same_period | **Dropped.** Predictive-skill incentive carried by direction-skill bonus + correct-call asymmetry. Per-fold buy-and-hold delta tracked as HEALTH_DIAG metric only. |
|
||||
| Phase 0 confounded Hold removal AND aux_w boost — couldn't tell which helped | Single P0 commit | **Split:** P0a Hold-only (cheap test of user hypothesis), P0b aux probe (only if 0a partial) |
|
||||
| Skill bonus + luck-discount could produce positive reward for correct-direction expected-loser → reward gaming | α=β=1.0 raw constants | **Bounded:** `skill_bonus = sign × min(α_param, ratio × |alpha|)` with ratio=0.3 default. Worked examples below verify no negative-EV gaming. |
|
||||
| 5-epoch gate too short for aux to converge → false-negative risk | 5-epoch absolute threshold | **10-epoch trajectory criterion:** dir_acc trending up AND final > 0.52 |
|
||||
| Aux_w controller at hard cap forever in data-limited case → destabilises Q | `clamp(0.05, 1.5)` only | **Stagnation detector:** dual-EMA (fast/slow); if slow ≈ fast for ≥ N epochs, decay aux_w toward base |
|
||||
| No worked examples of full reward composition | Inline formulas | **Worked-example matrix** for 4 trade-outcome quadrants (correct/wrong × winner/loser) |
|
||||
|
||||
## TL;DR
|
||||
|
||||
The reward function defines what "success" means to the model. SP1-SP12 lineage rewarded
|
||||
**any positive bounded P&L** — including P&L from structural extraction (trail-stop
|
||||
arithmetic on directionally-random bets at 46% win rate). SP13 rewards **predictive
|
||||
skill specifically** via two surgical changes:
|
||||
|
||||
1. **Eliminate Hold action** — model must commit. Tests user's hypothesis that Hold's
|
||||
availability has been hiding directional signal across the entire SP1-SP12 lineage.
|
||||
2. **Direction-skill bonus + correct-call asymmetry** — bounded explicit reward for
|
||||
predicting direction correctly, bounded penalty for being wrong, and discount on
|
||||
wrong-direction lucky wins. Mathematically calibrated so the model cannot game by
|
||||
entering expected-loser trades.
|
||||
|
||||
Plus three structural supports:
|
||||
|
||||
3. **Aux head regression → binary classification** on next-30-bar return sign, wired
|
||||
into direction Q-head input.
|
||||
4. **ISV-driven aux_w controller** (replaces SP11 inverted formula) with deficit-driven
|
||||
amplification AND stagnation detector to prevent destabilization in data-limited case.
|
||||
5. **Per-fold buy-and-hold delta** tracked as HEALTH_DIAG metric (NOT reward signal).
|
||||
|
||||
~600-750 LOC across 4 commits (P0a, P0b, B, C). Greenfield refactor — replay buffer
|
||||
compatibility intentionally broken.
|
||||
|
||||
**Hypothesis under test** (user, 2026-05-04): directional signal IS in the data; ~46% WR
|
||||
persists because Hold hides it. P0a smoke decides cleanly:
|
||||
- WR climbs > 0.52 with Hold removed → ✅ confirmed, hypothesis was right
|
||||
- WR stays flat at 0.46-0.50 → P0b adds aux amplification as second probe
|
||||
- Both P0a + P0b flat → ❌ data limit, halt SP13
|
||||
|
||||
## Background — why current system fails
|
||||
|
||||
50-epoch validation `train-multi-seed-pmbwn` and 22-epoch SP12 v3 validation
|
||||
`train-multi-seed-vd2v5`:
|
||||
|
||||
- val_dir_dist: hold=0.45 (lazy default), Long+Short=0.36 (active bets)
|
||||
- win_rate=0.4651 (slightly below random for direction)
|
||||
- profit_factor=1.24 (winners 24% bigger than losers)
|
||||
- aux_w stuck at 0.09 (suppressed by SP11-era `0.1 × health × (1 − sharpe)` controller)
|
||||
- aux next_bar_mse=3.83 (regression error suggests poor signal extraction OR aux never got real gradient)
|
||||
|
||||
**The mechanism**: model picks Hold ~half the time. Of the ~37% directional bets, only
|
||||
46% are correct. Profitability comes from trail-stop arithmetic, not predictive skill.
|
||||
|
||||
## The architectural shift
|
||||
|
||||
| Current "success" | Proposed "success" |
|
||||
|---|---|
|
||||
| Maximize positive bounded P&L | Maximize bounded P&L + bounded direction-skill credit |
|
||||
| Hold is safe default | Must commit (no Hold option) |
|
||||
| Wrong direction + lucky exit = full reward | Wrong direction + win = 0.3× reward (luck discount) |
|
||||
| Aux head suppressed when sharpe high | Aux amplified when WR below target, decayed when stagnant |
|
||||
| 1-bar reward horizon (next-bar MSE) | 30-bar prediction horizon (sign classification) |
|
||||
|
||||
## The five changes
|
||||
|
||||
### Change 1: Eliminate Hold action
|
||||
|
||||
**Current**: 4-way direction `{Short=0, Hold=1, Long=2, Flat=3}` × magnitude × ord × urg.
|
||||
**New**: 3-way `{Short=0, Long=1, Flat=2}` — Hold removed, Short stays at 0 (preserves
|
||||
canonical anchor used across 30+ kernel comments and tests), Long shifts down 2 → 1,
|
||||
Flat shifts down 3 → 2.
|
||||
|
||||
**Action semantics in 3-way world** (replaces `project_hold_action_design`):
|
||||
- `Long` at t when prior position = Flat → open Long
|
||||
- `Long` at t when prior position = Long → keep Long (same as old "Hold-Long")
|
||||
- `Long` at t when prior position = Short → flip to Long
|
||||
- `Flat` at t when any position → close all positions
|
||||
- `Short` symmetric
|
||||
|
||||
`Long` is no longer "open Long" specifically; it is "have Long position". `Long` is a
|
||||
state, not a verb.
|
||||
|
||||
**Affected** (atomic per `feedback_no_partial_refactor`, ~50–70 touch-sites mapped by
|
||||
Explore agent 2026-05-04):
|
||||
- Canonical Rust enum at `crates/ml-core/src/common/action.rs:124-132` and
|
||||
`NUM_DIRECTIONS: 4 → 3` at `crates/ml-core/src/state_layout.rs:48-49`
|
||||
- `state_layout.cuh` C/CUDA mirror: `#define NUM_DIRECTIONS 3` and `DIRECTION_SHORT=0,
|
||||
DIRECTION_LONG=1, DIRECTION_FLAT=2` (DIRECTION_HOLD removed)
|
||||
- Q-head output dim 4 → 3, weight/bias re-init, layout fingerprint bump
|
||||
- Action selection kernel argmax over 3 (`experience_action_select`)
|
||||
- C51/IQN per-direction-bucket loops (~6 `for d<4` sites become `< NUM_DIRECTIONS`)
|
||||
- **Action packing recompacts 12 → 9 slots**: `action_counts` buffer and `exp_idx =
|
||||
dir*3 + mag` packing in `financials.rs:197` shrink from 12 → 9 (3 dirs × 3 mags).
|
||||
Real shape change, not slot-3 dead-space. Consumers in `monitoring.rs:341` reindex.
|
||||
- **Direction-bias signal at `experience_kernels.cu:5159`**: 4-tuple per-direction
|
||||
bias weights `[Short=1.0, Hold=0.5, Long=1.0, Flat=0.0]` re-derives to 3-tuple
|
||||
`[Short=1.0, Long=1.0, Flat=0.0]`. Hold's 0.5 was a softening gate; with Hold gone
|
||||
the gate is no longer needed (Long/Short are full-weight, Flat is the soft branch).
|
||||
- SP5 Pearl 8 per-direction trail-distance ISV range `TRAIL_DIST_PER_DIR_BASE=270..274`:
|
||||
producer rewrites at new indices `{0=Short, 1=Long, 2=Flat}`; slot 273 dead-space
|
||||
(per-direction-indexed ISV ranges keep length 4 to avoid cascade-renumber of all
|
||||
downstream slots — consumers read by direction index ∈ 0..2 so dead slot is never
|
||||
touched). Recompaction deferred to SP13 close-out.
|
||||
- HEALTH_DIAG val_dir_dist 3-bucket emit: `metrics.rs:949`, `magnitude_distribution.rs:88,217`,
|
||||
`distributional_q_tests.rs:790`
|
||||
- KL divergence calc (if present)
|
||||
- 30+ stale doc/comment lines across kernel headers — implementer audits in same commit
|
||||
|
||||
### Change 2: Direction-skill bonus (bounded)
|
||||
|
||||
```
|
||||
if action ∈ {Long, Short}:
|
||||
correct = (sign(action) == sign(realized_30_bar_return))
|
||||
skill_cap = SKILL_BONUS_CAP_RATIO_INDEX × |alpha_capped| # default 0.3 × |alpha|
|
||||
if correct:
|
||||
skill_bonus = +min(DIR_SKILL_BONUS_ALPHA_INDEX, skill_cap)
|
||||
else:
|
||||
skill_bonus = -min(DIR_SKILL_BONUS_BETA_INDEX, skill_cap)
|
||||
else: # action == Flat
|
||||
skill_bonus = 0
|
||||
```
|
||||
|
||||
**Calibration check** (worked examples — α=β=1.0, ratio=0.3, SP12 cap [-10,+5]):
|
||||
|
||||
| Trade outcome | alpha (capped) | skill_bonus | luck-disc | reward | sane? |
|
||||
|---|---|---|---|---|---|
|
||||
| correct-Long-winner (big win) | +5 | +min(1.0, 1.5) = +1.0 | n/a | +6.0 | ✓ |
|
||||
| correct-Long-winner (small win) | +0.5 | +min(1.0, 0.15) = +0.15 | n/a | +0.65 | ✓ |
|
||||
| correct-Long-loser (small loss) | −0.5 | +min(1.0, 0.15) = +0.15 | n/a | **−0.35** | ✓ still negative |
|
||||
| correct-Long-loser (big loss) | −10 | +min(1.0, 3.0) = +1.0 | n/a | **−9.0** | ✓ still very negative |
|
||||
| wrong-Long-winner LUCKY (big) | +5 | −min(1.0, 1.5) = −1.0 | × 0.3 | (5−1)×0.3 = **+1.2** | ✓ reduced from naive +6 |
|
||||
| wrong-Long-winner LUCKY (small) | +0.5 | −min(1.0, 0.15) = −0.15 | × 0.3 | (0.5−0.15)×0.3 = +0.105 | ✓ |
|
||||
| wrong-Long-loser | −10 | −min(1.0, 3.0) = −1.0 | n/a | **−11.0** | ✓ full penalty + skill penalty |
|
||||
| Flat in any direction | 0 | 0 | n/a | 0 | ✓ |
|
||||
|
||||
**Invariant verified**: every correct-direction loser has reward < 0; every
|
||||
wrong-direction expected-loser has reward < 0. The model cannot game by entering
|
||||
trades it expects to lose. ✓
|
||||
|
||||
### Change 3: Conditional correct-call asymmetry (luck-win discount)
|
||||
|
||||
```
|
||||
if trade_pnl > 0 AND direction_was_wrong:
|
||||
reward *= LUCK_WIN_DISCOUNT_INDEX /* default 0.3 */
|
||||
# else: reward unchanged
|
||||
```
|
||||
|
||||
Penalises specifically lucky wins (positive outcome despite wrong direction prediction).
|
||||
Forces the model to optimize for correct prediction, not just positive outcome.
|
||||
|
||||
ISV-driven discount factor; default 0.3 starts strong, can be relaxed by ISV controller
|
||||
if luck-win frequency drops below threshold.
|
||||
|
||||
### Change 4: Aux head — regression → binary classification
|
||||
|
||||
**Current**: `next_bar_return_mse` (regression). MSE 3.83 — worse than predicting zero.
|
||||
|
||||
**New**: `next_30_bar_return_sign` cross-entropy. 30 bars matches SP12 min_hold_target.
|
||||
|
||||
**Output**: 2-class softmax `[P(positive), P(negative)]`.
|
||||
|
||||
**Wired into direction Q-head**: aux signal `logit_diff = softmax[1] − softmax[0]` ∈
|
||||
[−1, +1] is concatenated into direction Q-head input layer via ISV[`AUX_DIR_PREDICTION_INDEX`].
|
||||
This explicitly couples predictive signal to action selection.
|
||||
|
||||
**No information leakage**: aux head trained with hindsight labels (sign of
|
||||
price[t+30] − price[t], known at training time only). At inference, aux predicts using
|
||||
state s_t alone — no future info accessible. Same as existing next_bar_mse pattern.
|
||||
|
||||
### Change 5: ISV-driven aux_w controller (with stagnation detector)
|
||||
|
||||
**Current** (SP11-era, broken):
|
||||
```rust
|
||||
let aux_w = (0.1 * learning_health * (1 - sharpe_tanh)).clamp(0.05, 0.3);
|
||||
```
|
||||
Lowers aux when sharpe is high — exactly wrong for the failure mode (high sharpe + low WR).
|
||||
|
||||
**New** (deficit-driven + stagnation):
|
||||
```
|
||||
target = ISV[TARGET_DIR_ACC_INDEX] # default 0.55
|
||||
short = ISV[AUX_DIR_ACC_SHORT_EMA_INDEX] # fast EMA (α=0.3, ~3-epoch lookback)
|
||||
long = ISV[AUX_DIR_ACC_LONG_EMA_INDEX] # slow EMA (α=0.05, ~20-epoch lookback)
|
||||
|
||||
deficit = max(0, target - short)
|
||||
improvement = max(0, short - long) # positive when short pulling ahead of long
|
||||
stagnation = max(0, 1 - improvement / max(0.005, deficit))
|
||||
# 1 = fully stagnant (no improvement); 0 = improving fast
|
||||
|
||||
aux_w = base * (1 + 5 * deficit) * (1 - 0.7 * stagnation)
|
||||
aux_w = clamp(aux_w, base * 0.3, base * 3.0) # never zero, never beyond 3× base
|
||||
```
|
||||
|
||||
Rationale:
|
||||
- When `short < target` and improving (`short > long`): aux_w rises strongly to push harder
|
||||
- When `short < target` and stagnant (`short ≈ long`): aux_w decays back toward base — "we tried, it didn't help, stop destabilising Q"
|
||||
- When `short ≥ target`: deficit = 0, aux_w = base. Steady-state.
|
||||
- Hard floor at 0.3× base prevents complete suppression (different from SP11 inversion bug).
|
||||
|
||||
base = 0.5 in Phase 0b and beyond (SP11 base 0.05 was the suppression bug; lifting is
|
||||
the fix, not a confounder, because the SP11 formula was structurally wrong).
|
||||
|
||||
**ISV slots used**: 372 (target), 373 (short EMA), 374 (long EMA). Both EMAs reset per fold.
|
||||
|
||||
## ISV slot allocation
|
||||
|
||||
8 new slots in `[372..380)`:
|
||||
|
||||
| Slot | Name | Purpose | Reset |
|
||||
|---|---|---|---|
|
||||
| 372 | `TARGET_DIR_ACC_INDEX` | Directional accuracy target (default 0.55) | static |
|
||||
| 373 | `AUX_DIR_ACC_SHORT_EMA_INDEX` | Fast EMA of dir_acc (α=0.3) | per-fold |
|
||||
| 374 | `AUX_DIR_ACC_LONG_EMA_INDEX` | Slow EMA of dir_acc (α=0.05), stagnation detector | per-fold |
|
||||
| 375 | `AUX_DIR_PREDICTION_INDEX` | Aux head logit_diff per bar (concat into Q-head input) | per-bar |
|
||||
| 376 | `DIR_SKILL_BONUS_ALPHA_INDEX` | Skill bonus magnitude (correct direction), default 1.0 | per-fold |
|
||||
| 377 | `DIR_SKILL_BONUS_BETA_INDEX` | Skill bonus penalty (wrong direction), default 1.0 | per-fold |
|
||||
| 378 | `LUCK_WIN_DISCOUNT_INDEX` | Lucky-win discount factor, default 0.3 | per-fold |
|
||||
| 379 | `SKILL_BONUS_CAP_RATIO_INDEX` | Cap ratio: bonus ≤ ratio × \|alpha\|, default 0.3 | per-fold |
|
||||
|
||||
Bump `SP5_SLOT_END = 380`, `ISV_TOTAL_DIM = 380`.
|
||||
|
||||
**Slot 371 (was BENCHMARK_PNL_CUMULATIVE_INDEX in v1) — DROPPED.** Per-fold buy-and-hold
|
||||
delta is computed at fold-end in HEALTH_DIAG emit (trivial: `price[fold_end] −
|
||||
price[fold_start]`). No kernel, no slot, no reward dependency.
|
||||
|
||||
## Implementation phases
|
||||
|
||||
### Phase 0a — Hold elimination only (1 atomic commit)
|
||||
|
||||
**Purpose**: clean test of user's hypothesis. Hold removed; aux_w controller, dir_acc
|
||||
instrumentation, skill bonus, aux classification all UNCHANGED from current main.
|
||||
|
||||
- ISV slot 372 (target, static), 373 (short EMA), 374 (long EMA), 375 (aux prediction —
|
||||
but produced by the existing regression aux head, just `tanh(scalar_pred)` for now to
|
||||
fit the [−1, +1] interface)
|
||||
- `aux_dir_acc_reduce_kernel` — single-block tree-reduce of (correct_count, pos_pred,
|
||||
pos_label, valid_count) → 3 output floats per epoch
|
||||
- HEALTH_DIAG `aux_dir_acc accuracy=… pos_pred=… pos_label=…` per epoch
|
||||
- Hold elimination atomic refactor (state_layout, Q-head, action select, C51/IQN bucket
|
||||
state, replay encoding, val_dir_dist, KL, Rust enum)
|
||||
- ~300 LOC
|
||||
|
||||
**Phase 0a smoke** (10-epoch L40S):
|
||||
- ✅ **Pass**: `val_dir_acc[ep9] > val_dir_acc[ep1]` AND `val_dir_acc[ep9] > 0.52` AND
|
||||
`val_win_rate[ep9] > 0.51` → user's hypothesis confirmed; proceed Layer B.
|
||||
- ⚠️ **Partial** (one trending up, the other flat): proceed P0b.
|
||||
- ❌ **Both flat** (`dir_acc < 0.51` AND `win_rate < 0.50` after 10 epochs): proceed P0b
|
||||
to add aux amplification. If P0b also fails, halt SP13.
|
||||
|
||||
10 epochs (not 5) because aux head regression has been suppressed for entire SP1-SP12
|
||||
lineage — needs time even at unchanged aux_w to show whether signal is recoverable.
|
||||
Cost: ~€0.60.
|
||||
|
||||
### Phase 0b — Add corrected aux_w controller (only if 0a partial/red, atomic commit)
|
||||
|
||||
- Replace SP11 inverted formula at `training_loop.rs:4103-4104` with deficit-driven +
|
||||
stagnation formula from Change 5
|
||||
- Bump aux_w base 0.05 → 0.5
|
||||
- ~80 LOC
|
||||
|
||||
**Phase 0b smoke** (10-epoch):
|
||||
- Same gate criterion as P0a
|
||||
- If still red: halt SP13. Update `project_hold_hides_directional_signal_hypothesis.md`
|
||||
status to DISCONFIRMED. Pivot to data-augmentation work (separate spec, SP14).
|
||||
|
||||
### Layer B — Aux head regression → classification (atomic, only on green Phase 0)
|
||||
|
||||
- Aux head output dim 1 → 2 + softmax + cross-entropy loss
|
||||
- Label generation: `sign(price[t+30] − price[t])` as 0/1 class
|
||||
- `aux_dir_acc_reduce_kernel` adapted: read 2-class softmax instead of regression scalar
|
||||
- Wire `logit_diff = softmax[1] − softmax[0]` → ISV[375] → direction Q-head input concat
|
||||
(in_dim grows by 1)
|
||||
- ~120 LOC
|
||||
|
||||
### Layer C — Reward composition (atomic)
|
||||
|
||||
- ISV slots 376, 377, 378, 379 + state-reset registry
|
||||
- Direction-skill bonus device-inline function (`compute_direction_skill_bonus_bounded`)
|
||||
- Luck-win discount device-inline function (`compute_luck_win_discount`)
|
||||
- Replace SP12 reward composition at exit-event sites: keep SP12 asymmetric cap, ADD
|
||||
skill bonus, ADD luck discount
|
||||
- HEALTH_DIAG: add `per_fold_alpha_vs_buy_and_hold = total_pnl − (price[fold_end] −
|
||||
price[fold_start])` — METRIC ONLY, not reward signal
|
||||
- 12+ GPU oracle tests for new reward math (extends SP12 sp12_reward_math_tests scaffold)
|
||||
- ~150 LOC
|
||||
|
||||
### Layer D — 30-epoch validation + close-out (no commit until green)
|
||||
|
||||
- 30-epoch L40S full validation
|
||||
- 3-action SP12 control run for fair baseline (~€1, optional but recommended for clean
|
||||
before/after comparison)
|
||||
- If WR ≥ 55% AND profit_factor improvement ≥ 15% over 3-action SP12 control →
|
||||
close-out commit: audit doc + 2 new pearls (`pearl_redefine_success_for_predictive_skill`,
|
||||
`pearl_skill_bonus_must_be_alpha_bounded`) + meta-pearl
|
||||
`pearl_reward_quadrant_audit_required`
|
||||
- If WR plateaus 50-54%: tune horizon (next_60 instead of next_30) or skill bonus α/β,
|
||||
re-validate
|
||||
|
||||
## Validation targets
|
||||
|
||||
**Phase 0a smoke (10-epoch)** — primary gate:
|
||||
- val_dir_acc trajectory: monotone non-decreasing OR final > start
|
||||
- val_dir_acc[final] > 0.52
|
||||
- val_win_rate[final] > 0.51
|
||||
- Hold action eliminated: val_dir_dist sums to 1.0 over {Long, Short, Flat}
|
||||
- No NaN/inf
|
||||
- mean(weights) controller still ≈ 1.0
|
||||
- aux next_bar_mse trajectory: not blowing up
|
||||
|
||||
**Phase 0b smoke (10-epoch)** — secondary gate (only if 0a partial):
|
||||
- aux_w controller responding: aux_w[ep5] > base when deficit > 0.05
|
||||
- Stagnation detector engaging: aux_w decays toward base when short EMA ≈ long EMA
|
||||
- Same dir_acc + win_rate criteria as 0a
|
||||
|
||||
**Layer B 5-epoch smoke**:
|
||||
- Aux CrossEntropy loss decreasing
|
||||
- val_dir_acc continues at or above Phase 0 level
|
||||
- aux signal feeding Q-head (verify by perturbing aux output → direction action distribution shifts)
|
||||
|
||||
**Full validation (30-epoch)**:
|
||||
- val_dir_dist: roughly {Long: 0.4, Short: 0.4, Flat: 0.2} — model committing
|
||||
- **win_rate ≥ 0.55** PRIMARY SUCCESS CRITERION
|
||||
- profit_factor ≥ 1.5 (SP12 baseline 1.24 × 21% improvement floor for "skill-driven" claim)
|
||||
- per_fold_alpha_vs_buy_and_hold > 0 sustained (the dropped reward signal, now metric)
|
||||
- max_drawdown < 0.005
|
||||
|
||||
## Operating principles (NON-NEGOTIABLE)
|
||||
|
||||
- **`feedback_no_partial_refactor`**: P0a, P0b, B, C each single atomic commit
|
||||
- **`feedback_isv_for_adaptive_bounds`**: target_dir_acc, aux_w base/cap, bonus α/β,
|
||||
luck-discount, skill-bonus-cap-ratio — ALL ISV-driven
|
||||
- **`feedback_no_quickfixes`**: this IS the proper fix per pearls
|
||||
- **`feedback_no_feature_flags`**: behavior change unconditional
|
||||
- **End-to-end training**: NO pre-training of aux head
|
||||
- **`pearl_audit_unboundedness_for_implicit_asymmetry`**: SP12 asymmetric cap [-10,+5]
|
||||
preserved
|
||||
- **`pearl_event_driven_reward_density_alignment`**: skill bonus + luck discount fire
|
||||
only at trade-exit events, not per-bar
|
||||
- **(NEW) Reward-quadrant audit**: any reward composition change must include the 4-
|
||||
quadrant table (correct/wrong × winner/loser) showing all signs and inequalities. This
|
||||
rule will become `pearl_reward_quadrant_audit_required` post-validation.
|
||||
|
||||
## Estimated cost
|
||||
|
||||
**Implementation**: ~10-12h (~700 LOC across P0a + P0b + B + C + D close-out)
|
||||
**P0a smoke**: ~€0.60 (10-epoch L40S)
|
||||
**P0b smoke** (if needed): ~€0.60
|
||||
**Layer B smoke**: ~€0.30 (5-epoch)
|
||||
**30-epoch validation**: ~€1.00
|
||||
**3-action SP12 control** (recommended): ~€1.00
|
||||
|
||||
**Total best case** (P0a green, no P0b): ~€2.90 + 10-12h
|
||||
**Total worst case** (P0a red, P0b green): ~€3.50
|
||||
**Total disconfirm case** (both P0a + P0b red): ~€1.20 — cheapest learn
|
||||
|
||||
## Open questions
|
||||
|
||||
1. **Aux prediction horizon**: 30 bars matches SP12 min_hold_target. If 30-epoch
|
||||
validation shows partial WR climb (50-54%), Layer D iteration tries 60-bar horizon.
|
||||
2. **Skill-bonus cap ratio**: 0.3 chosen so worst-case bonus contribution to reward at
|
||||
|alpha|=10 is ±1.0 — comparable to one tick of P&L. ISV-driven, can be tightened to
|
||||
0.2 if Layer B/C reveals reward gaming.
|
||||
3. **Stagnation detection threshold**: 0.7 decay coefficient is heuristic. Worth probing
|
||||
in Layer C smoke (read aux_w trajectory).
|
||||
4. **Whether to keep SP12 min-hold penalty**: SP12's `min_hold_factor` curriculum still
|
||||
applies. Spec recommends KEEP — it's orthogonal to skill incentive and prevents
|
||||
compulsive churn. Re-evaluate at Layer D close-out.
|
||||
|
||||
## Decision needed
|
||||
|
||||
Approve SP13 v2 spec → update implementation plan to match → dispatch P0a implementer
|
||||
chain → 10-epoch smoke → branch on outcome.
|
||||
Reference in New Issue
Block a user