hjnormey/sentiment-engine
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
cd57abb089b48e5d517415da4b08ff14a3287386
sentiment-engine/prod/docs/V7_RISK_DOMINANT_PREEMPTION_REFERENCE.md

16 KiB
Raw Blame History

V7_RISK_DOMINANT — Preemption Reference & Economic-Policy Note

Status: verified against the live r7 tape (f13_v21_terminal_max_hold_r7_overnight_20260828). Scope: which subsystem preempts a correctly‑resting TP, why, and what the realized economics were — plus the one policy lever that is actually in play. Constraint: the kernel worktree /root/uv-wt/f13-hl is READ‑ONLY here. This doc cites LoCs for review; it does not tune live r7 (see ·Tuning caveat).

TL;DR (corrected): V7_RISK_DOMINANT is not a bug and is not ADVSL/TP_FLOOR. It is the RETRACT reason emitted by BLUE's AlphaExitEngineV7 when its exit_pressure metric exceeds retract_pressure_threshold (default 1.7). In r7 it fired on two ETH SHORTS and cancelled their correctly‑rested ×1.40 TPs (never hit), closing the legs at −0.0060% then −0.0201% (residual market close) — net‑zero, slot recycled. The TP was placed correctly (tp_exit_modulation ×1.40); the move simply never reached it. The right question is an economic‑policy review of the retract_pressure_threshold (1.7), not a change to TP/ADVSL/MAX_HOLD.

1. What V7_RISK_DOMINANT is

A reason string produced by BLUE's vol‑normalised exit engine, not a kernel LoC. It is returned as the reason field of the engine's decision dict and surfaced by r7 as v7_reason=V7_RISK_DOMINANT in the RETRACT‑edge log line.

Artifact File (READ‑ONLY) LoC Role
AlphaExitEngineV7 (the engine) /mnt/dolphinng5_predict/nautilus_dolphin/nautilus_dolphin/nautilus/alpha_exit_v7_engine.py :364 class; :447 evaluate; :753‑754 the V7_RISK_DOMINANT RETRACT branch emits {"action":"RETRACT","reason":"V7_RISK_DOMINANT",...}
AlphaExitV7Config thresholds …/alpha_exit_v7_engine.py :163‑170 exit_pressure_threshold=3.0, retract_pressure_threshold=1.7, extend_pressure_threshold=-0.5
V7 calibration notes …/alpha_exit_v7_engine.py :721‑741 13,000+ RETRACT events → 1.0→1.7; ≥1.7 == "correct retractions" (avg PnL negative)
VioletExitV7Adapter (VIOLET boundary) /root/uv-wt/f13-hl/prod/clean_arch/violet/exit_v7_wrapper.py :68 class; :71‑84 init (default config); :111 evaluate; :132‑138 priority=2, reason passthrough wraps the engine; returns ExitDecision(reason=…, priority=2)
Wiring into r7 /root/uv-wt/f13-hl/prod/clean_arch/violet/uv/tpsl_ticker.py :732‑766 _evaluate_v7; :496‑523 V7 EXIT path; :596‑632 V7 RETRACT edge where the RETRACT decision turns into a cancel + v7_retract close

1.1 The engine's decision tiering (verbatim, alpha_exit_v7_engine.py:742‑759)

if exit_pressure > cfg.exit_pressure_threshold:        # 742  (3.0 default)
    if mae_risk > mfe_risk and mae_risk > 0:
        _reason = "V7_MAE_SL_VOL_NORM"
        return {"action": "EXIT",   "reason": _reason,              **base_out}   # 745 full
    # V7_COMPOSITE_PRESSURE — only downgrade to RETRACT if MAE
    # exceeds the floor.  Negligible adverse movement (< 2 bps)
    # is likely microstructure noise, not a real exit signal.
    if mae >= cfg.composite_pressure_mae_floor:        # 749  (0.0002)
        return {"action": "RETRACT", "reason": "V7_RISK_DOMINANT",   **base_out}   # 750 <-- primary RETRACT
if exit_pressure > cfg.retract_pressure_threshold:     # 753  (1.7 default)
    return {"action": "RETRACT", "reason": "V7_RISK_DOMINANT",      **base_out}    # 754 <-- primary RETRACT

if exit_pressure < cfg.extend_pressure_threshold and pnl_pct > 0:                 # 756 (-0.5)
    return {"action": "EXTEND",  "reason": "V7_DIRECTIONAL_EDGE",   **base_out}    # 757
return {"action": "HOLD",        "reason": None,                  **base_out}    # 759

V7_RISK_DOMINANT fires on two paths — both RETRACT:

  1. :750 — COMPOSITE_PRESSURE downgrade: exit_pressure > 3.0 AND mae ≥ composite_pressure_mae_floor (0.0002).
  2. :754 (PRIMARY for r7) — exit_pressure > retract_pressure_threshold (1.7).

RETRACT is a half / de‑risk action (see §3), not a full EXIT.

1.2 Calibration evidence (verbatim, :721‑741)

# EXIT pressure threshold: 3.0 (raised from 2.69, 2026-05-12)
#   79 COMPOSITE_PRESSURE EXIT events: total -$963 across 78 trades.
#   9 trades had MAE < 0.0002 (2 bps) at exit — avg PnL +17% (V7 units),
#   meaning they were IN PROFIT when cut … false positives from micronoise.
#   MAE guard (composite_pressure_mae_floor = 0.0002) prevents these 9.
#   remaining 70 exits at MAE >= 2bps avg PnL -$12.35 — genuine adverse
#   movements where the cut saves against MAX_HOLD loss.
#
#   RETRACT threshold raised from 1.0 to 1.7 based on 13,000+ RETRACT events:
#   ep=1.0-1.6 avg PnL was POSITIVE (+0.07 to +0.12 V7 units) →
#   "false retractions on profitable trades".  ep=1.7+ avg PnL turns
#   negative, indicating correct retractions.

So the engine asserts ≥1.7 pressure == a "correct" cut (it is saving against a bigger MAX_HOLD loss on average). The r7 question is whether that aggregate holds for r7's regime (see §5).

2. Where it fires in r7 — the dispatch ordering (READ‑ONLY)

alpha_exit_v7_engine.py is wrapped by VioletExitV7Adapter (exit_v7_wrapper.py:71‑84 — constructed by r7 with no config arg, so the engine uses the 1.7 default; the 2.0 seen at alpha_exit_v7_engine.py:416 is only a docstring example for a per‑asset override that has no callers — get_effective_config :410 / set_asset_config :421 are stubs).

r7 evaluates it once per bar in tpsl_ticker.py:_evaluate_v7 (:732‑766):

tpsl_ticker.py:482   # V7 (BLUE parity, both roles) — one evaluation per bar
tpsl_ticker.py:492   if has_market_price and self._v7_authority_enabled:
tpsl_ticker.py:493       v7_raw = self._evaluate_v7(pos, bars, price, ingress)
tpsl_ticker.py:494       v7_action = str(v7_raw.get("action","")).upper()

The exit_authority dispatch this bar (inside the if not _expired / bars<max_hold branch) is strictly ordered (comments are in tpsl_ticker.py):

  1. V7 EXIT (full) — tpsl_ticker.py:496‑523 (if v7_action == "EXIT"). "BLUE accepts ONLY the provider's EXIT (HOLD/RETRACT fall through); it outranks ADVSL and MAX_HOLD on this bar." → exit_fraction=1.0, trigger_type="v7_exit", reason=_normalize_v7_exit_reason(raw.reason).
  2. ADVSL — tpsl_ticker.py:524‑595. Shadow‑only by default; live actuation behind UV_ADVSL_LIVE=1 opt‑in (tpsl_ticker.py:543‑544). For r7 this was OFF (the boot arm EXIT PRE-EMPTION armed (ADVSL/TP_FLOOR…) is a generic notice and did not actuate on these trades — see §6).
  3. V7 RETRACT (half) — tpsl_ticker.py:596‑632: 
    # 596  # DE-RISK HALF on the HOLD->RETRACT transition only —
# 597  # fraction 0.50, never repeated while RETRACT is sustained
# 603  if v7_action == "RETRACT" and v7_prev_action != "RETRACT":
# 604      self._v7_retract_fires += 1
# 612      exit_fraction=0.5, exit_leg_ratios=(0.5,)
# 616      LOGGER.warning("V7 RETRACT (half): %s %s state=%s v7_reason=%s",
# 620      # ANTI-GHOST (2026-08-13): a RETRACT whose exit is REJECTED while
# 622      # venue truth is FLAT is a ghost → confirm_closed + forget.
# 625      if state == "rejected" and self._venue_flat(pos.asset):
    This is the V7_RISK_DOMINANT path: it submits an EXIT PRE-EMPT CANCEL that cancels the resting TP (so the TP never fills), then routes the de‑risk half through trigger_type="v7_retract".
  4. MAX_HOLD terminal — tpsl_ticker.py:633‑660 (Add.15 TERMINAL MAX_HOLD), gated by _max_hold_expired (tpsl_ticker.py:567). Not reached for these legs (closed via V7 RETRACT first).

Config knobs / telemetry (tpsl_ticker.py:299‑313, :1092): '_v7_authority_enabled = os.environ.get("UV_V7_EXIT_AUTHORITY","1")' (307‑08); counters '_v7_evaluations', '_v7_exit_fires', '_v7_retract_fires', '_v7_last_action' (311‑13) emitted at :1092.

3. The two r7 preempted legs — full tape (READ‑ONLY spool)

Spool: /root/dolphin_logs/uv_journal_runs/f13_v21_terminal_max_hold_r7_overnight_20260828/runner.log

force‑fd9731b2 (ETH SHORT) — entry 2491.50, TP rested at 2484.5238 (0.28% OTM, ×1.40)

t (2026‑08‑28) line log
13:58:27 513 TP REST-ON-ENTRY SUBMITTED: force-fd9731b2f9 ETHUSDT qty=0.008400 @ … (TP + entry placed)
13:59:07.458 558 V7 RETRACT (half): ETHUSDT force-fd9731b2f9 state=preempt_requested v7_reason=V7_RISK_DOMINANT
13:59:07.458 — EXIT PRE-EMPT CANCEL submitted: force-fd9731b2f9 ETHUSDT — v7_retract out…
13:59:08.931 560 EXIT PRE-EMPT CANCEL OUTCOME: trade_id=force-fd9731b2f9 verdict=confirmed (TP cancelled, slot released)
13:59:08.933 563 TPSL EXIT: v7_retract ETHUSDT force-fd9731b2f9 at 2491.65 (-0.0060%) ← half close
14:02:14.783 669 V7 RETRACT (half): … state=dispatched v7_reason=V7_RISK_DOMINANT
14:02:14.789 673 TPSL EXIT: v7_retract force-fd9731b2f9 at 2492.00 (-0.0201%) ← residual close
14:02:26.449 679 POSITION CLOSED (net-zero) + exit disarmed: ETHUSDT (trade force-fd9731b2f9)
  • TP 2484.5238 never filled — live‑tape low ≈ 2491.35 (only ~0.0006% favourable; TP sat 0.28% away).
  • Realised: −0.0060% then −0.0201% (residual) → net ≈ −0.020%, slot recycled. ✔ matches the Dev‑agent live review ("preempted at ~−0.006%, closed remaining partial at about −0.020%").

force‑475d4ed5 (ETH SHORT) — entry 2492.0, TP rested ≈ 0.28% OTM

t line log
14:03:15 709 FORCE_ENGAGE: SYNTHETIC ENTER ETHUSDT SHORT @ 2504.48 u=0.8
14:03:16 / :17 716/717 TP REST-ON-ENTRY SUBMITTED → ACKED (reduce‑only TP on book)
14:06:04.213 806 V7 RETRACT (half): force-475d4ed5ce state=preempt_requested v7_reason=V7_RISK_DOMINANT
14:06:05.464 808 EXIT PRE-EMPT CANCEL OUTCOME: verdict=confirmed
14:06:05.469 811 TPSL EXIT: v7_retract force-475d4ed5ce at 2491.15 (0.0060%)
14:07:04.661 848 V7 RETRACT (half): state=dispatched v7_reason=V7_RISK_DOMINANT
14:07:04.665 851 TPSL EXIT: v7_retract force-475d4ed5ce at 2493.15 (-0.0743%)
14:07:19.304 865 POSITION CLOSED (net-zero) + exit disarmed: ETHUSDT (trade force-475d4ed5ce)

Same shape: V7_RISK_DOMINANT cancels the resting TP, closes half then residual at near‑entry → net‑zero, slot recycled, FORCE re‑engages. A third leg (force-e3d6af52, DOGE) cycled the same way (:1733 POSITION CLOSED net-zero 14:31:28).

4. The TP was placed correctly (×1.40) — and was never hit

The SHORT entry‑TP level is the soft‑TP curve widened in the CASCADE branch:

# tp_exit_modulation.py:92-102
curve  = compute_soft_tp_pct(base, our_leverage)     # 92
dynamic = curve                                        # 93  branch = "NEUTRAL"
if ingress.cascade_count > 0:                          # 100
    dynamic *= 1.40                                    # 101  branch = "CASCADE"

For the r7 SHORTs base ≈ 0.20% × 1.40 (CASCADE) = 0.28% TP ⇒ 2484.5238 from a 2491.50 entry — exactly the "correctly rested TP" the live review cites. The TP order is reduce‑only on the book (:513/717 TP REST-ON-ENTRY SUBMITTED … reduce-once ACKed at the venue). It was not hit because tape low (2491.35) never approached 2484.52; it was cancelled by the V7 RETRACT edge (EXIT PRE-EMPT CANCEL OUTCOME verdict=confirmed).

5. Economic‑policy question (the lever)

V7_RISK_DOMINANT fires when exit_pressure > retract_pressure_threshold (1.7, default). "Pressure" is the engine's composite risk score (alpha_exit_v7_engine.py:694‑719 base_out keys: exit_pressure, rv_comp, mae_thresh1, bounce_score, bounce_risk, … — built from MAE tiers §5.2, velocity/acceleration, Bollinger squeeze, fractal/Hurst, OB microstruct, EXF macros). The calibration at :721‑741 treats ≥1.7 as a "correct" cut — but that is an aggregate claim (13,000+ retractions).

The r7 tape shows the cuts landing at −0.006% / −0.020% while a correctly rested ×1.40 TP 0.28% away sat untouched (tape moved <0.001% favourable). So:

Policy review (do NOT change while r7 is running): should retract_pressure_threshold (1.7 → 2.0?) let the correctly‑rested ×1.40 TP survive small adverse excursions, or is the ≥1.7 cut still right for this regime? Raising it = fewer preempts, more TP‑hits‑attempted (also more MAX_HOLD exposure if pressure is real). Lowering it = more near‑breakeven cuts.

The engine already records the signal for the review: every RETRACT bar emits exit_pressure via last_raw_decision → v7_raw (see tpsl_ticker.py:761‑762 raw["ob_imbalance"], and v7_raw.get("exit_pressure") logged at :513), plus the counters _v7_retract_fires / _v7_exit_fires (:313, :1092). A sample of (exit_pressure, pct_move, max_fav, realised) at each RETRACT edge is enough to decide without touching r7.

6. Definitely NOT this (corrections from the live tape)

  • ADVSL did not actuate. UV_ADVSL_LIVE defaults off (tpsl_ticker.py:544); ADVSL ran shadow‑only. The EXIT PRE-EMPTION armed (ADVSL/TP_FLOOR…) boot line is a generic pre‑emption notice; for these trades the specific cause was v7_reason=V7_RISK_DOMINANT, logged at the same millisecond as the CANCEL submit (:558). Do not chase tp_exit_modulation.py / ADVSL knobs for this.
  • TP_FLOOR never fired. (It is a tpsl_ticker.py MAX_HOLD floor, unrelated.)
  • My earlier 6.02% max_favorable_pct reading was stale/mis‑associated — the live tape shows tape low ≈ 2491.35 vs TP 2484.52; there was no 6% favourable move on fd9731b2. Retract that claim.

7. Observation / tuning cheat‑sheet (read‑only until policy decided)

  • See the V7 authority fire: grep -nE "V7 RETRACT \(half\)|TPSL EXIT: v7_retract|EXIT PRE-EMPT CANCEL OUTCOME" runner.log
  • Toggle V7 on/off (coarse; affects V7 EXIT, RETRACT, EXTEND): env UV_V7_EXIT_AUTHORITY (tpsl_ticker.py:307‑08, default "1").
  • Fine‑tune the threshold: AlphaExitV7Config.retract_pressure_threshold (alpha_exit_v7_engine.py:165, default 1.7; per‑asset set_asset_config is a stub :421 with no callers; r7 constructs the adapter with no config → exit_v7_wrapper.py:84 → engine default 1.7).
  • Counter telemetry: v7_evaluations / v7_exit_fires / v7_retract_fires (tpsl_ticker.py:311‑13, emitted to heartbeat at :1092).

LoC index (exact)

  • Reason producer (engine decision tiers): nautilus_dolphin/nautilus/nautilus/alpha_exit_v7_engine.py — :742‑759 (tiering), :750 & :754 (V7_RISK_DOMINANT RETRACT), :163‑170 (thresholds), :721‑741 (calibration), :447 (evaluate), :364 (class).
  • Boundary (priority=2, reason passthrough, default config): prod/clean_arch/violet/exit_v7_wrapper.py — :71‑84 (init), :110‑138 (evaluate), :132‑138 (priority=2 / RETRACT-not-promoted), :125‑128 (raw decision capture).
  • Dispatch into r7 (the preemptor path): prod/clean_arch/violet/uv/tpsl_ticker.py — :482‑495 (V7 eval/arm), :496‑523 (V7 EXIT full, outranks ADVSL/MAX_HOLD), :524‑595 (ADVSL shadow + UV_ADVSL_LIVE), :596‑632 (V7 RETRACT half + ANTI‑GHOST guard), :633‑660 (MAX_HOLD terminal), :299‑313 (V7 knobs/counters), :307‑308 (UV_V7_EXIT_AUTHORITY), :732‑766 (_evaluate_v7), :756 (engine .evaluate), :761 (last_raw_decision), :1092 (counter telemetry).
  • TP placement (correctly rested ×1.40, untouched): prod/clean_arch/violet/uv/tp_exit_modulation.py — :92 (compute_soft_tp_pct(base, leverage)), :100‑102 (CASCADE ×1.40). TP submit at entry: journal TP REST-ON-ENTRY SUBMITTED (e.g. runner.log:513, :717).
  • Tape evidence (READ‑ONLY spool): /root/dolphin_logs/uv_journal_runs/f13_v21_terminal_max_hold_r7_overnight_20260828/runner.log — :558/:560/:563 (fd9731b2 preempt), :669/:673/:679 (fd9731b2 close), :806/:808/:811 (475d4ed5 preempt), :848/:851/:865 (475d4ed5 close), :1733 (DOGE cycle close).

·Tuning caveat (hard)

r7 is live. No config/env edits, no code edits, no restart until policy is decided. Collect a RETRACT sample (exit_pressure, pct_move, max_favorable, realised) from the counters above; decide threshold offline; then relaunch r7 against the patched tree. The author of this doc is read‑only on the kernel worktree and will not mutate it.