The reactive-coverage cap (3e523e1) has run unchanged since it went live at
10:10 ET on 2026-10-02, so it is part of the strategy under test rather than an
open deviation; only TNON that morning traded without it. Dated closed with a
resolution note so the daily report stops listing it.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LXxioW7HA18AY1gqnzA2HN
On 2026-10-02 the live bot bought 7 TNON at 4.62 and the Tradier sell-stop filled at
4.30 forty-nine seconds later, before a position record existed, so no position_closed
event was written and the 80-trade scorecard silently dropped a -1R loser. The daily
equity-vs-P&L gap ($2.31 that day, cents everywhere else) is what exposed it.
load_trades now merges the journal with a cross-check of the broker's filled orders
(account_orders_updated snapshots, one row per broker order id). For any session day and
symbol with a filled buy and sell but no journaled close, reconstruct_missing synthesises
the trade from the fills: fill VWAPs for entry/exit, the sell-stop's stop price for R.
Such trades are labelled "[broker fill; journal missed]" and the daily report alerts on a
same-day miss. The paper side of the exit A/B skips the check: that store's order
snapshots run to 2.2 GB over the window and its Alpaca journal is reconciled by
alpaca_paper/sync.
Corrected test state: 16/80, mean R -0.08, realized +$0.91 (was 15/80, -0.02, +$3.15).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LXxioW7HA18AY1gqnzA2HN
The live (Tradier) stack has no backup service. Only the paper stack backs up its
own DB, so the 80-trade test's whole record lived only on the TrueNAS pool.
- deploy/backup_live_db.py, run by Task Scheduler "Momentum\LiveDbBackup" at 20:40 ET
nightly: pg_dump -Fc (PostgreSQL 18.6 client) of the live DB to
C:\Users\Admin\momentum-backups. It is verified with pg_restore --list (the events
and config_overrides data must be present), keeps the newest 7, and writes
status.json. A broken dump is deleted rather than kept.
- research/daily_report.py alerts when the last backup failed or is more than 36 h
old. When there are alerts it sends them to Telegram, using the stack's own sender
and the credentials in .env. --no-telegram suppresses that for dry runs.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01B3PXXn2tHBzdv6rYtGPz86
qwen3:14b rated "halted on circuit breaker to the upside, stock now up 172%" at
conviction 0.8 / sentiment +1.0, a bigger rank boost than most real news. So
names that had already run were ranked up.
- Prompt: halt notices, movers roundups and articles that only report a price move
are not a new catalyst (conviction <= 0.1). If one names a new announcement,
classify it with conviction <= 0.3.
- A deterministic cap that does not depend on the model: halt type or halt wording
caps conviction at 0.1; after-the-fact wording ("why is X stock up", "X shares
surge", "stocks moving") caps it at 0.3. Dilution flags are never touched.
On the 4,046-headline news history the pattern catches 100% of halt notices and 82%
of roundups, 0% of analyst items and 0.8% of earnings items. Against qwen3:14b on a
fixed 20-headline set: reactive mean boost 0.318 -> 0.055 (max 0.80 -> 0.21); real
catalysts 0.353 -> 0.362; offerings 3/3 still flagged dilutive.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01B3PXXn2tHBzdv6rYtGPz86
qwen3 thinks by default. With format=json, Ollama then returns "{}", which the parser
silently read as a confident "other" (sentiment 0, conviction 0), so every headline
scored zero with no error. Two changes:
- OLLAMA_THINK (true/false). It is sent as Ollama's "think" flag only when set, so other
models are untouched. qwen3:14b needs OLLAMA_THINK=false.
- A JSON object with none of the expected keys is now a failure (None, retried
next pass), not "other".
Checked through the live code path against qwen3:14b on mason (think=false): 1.4 s per
headline. 5/5 offerings (including an ATM sales agreement and a PIPE) flagged
dilutive. 'Stocks moving' roundups scored no catalyst. Earnings, guidance, analyst,
FDA and contract items were labelled correctly.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01B3PXXn2tHBzdv6rYtGPz86
research/daily_report.py, run by Task Scheduler "Momentum\DailyReport" at 16:25 ET
on weekdays, writes data/reports/daily_<date>.md with ALERTS on top:
- the live opening-range bot's frozen 80-trade test: verdict, mean R with CI,
P&L, equity headroom over the floor, today's trades, open deviations, settings
drift;
- the judge v2 paper trial: judge.review for every counted day, cumulative
paper-vs-sim execution parity, today's flagged trades, trader errors and
restarts;
- the old-vs-new exit A/B (alerts at 50 pairs);
- stack health: down at the open, silent gaps of 10+ min, the press-release
recorder.
Read-only everywhere (PG_SKIP_SCHEMA=1, SELECTs only). Replayed on 2026-09-30 it
flags that day's real problems: the judge error loop, the paper stack down until
13:51, and a 25-min gap.
judge/PAPER_TRIAL.md, written before any counted trade: trades count from
2026-10-02. At 60 trades the paper trial must show execution parity with the
simulator (paper minus sim >= -0.15R) before a 150-trade performance read. Real
money is only considered after that, and only with the owner's go-ahead.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01B3PXXn2tHBzdv6rYtGPz86
The host ran out of memory (843 MiB free; Ollama held 7.8 GB at a 131k context)
and the owner turned ollama2 off on 2026-09-30 evening. Strategy settings are
unchanged, but the LLM catalyst score is inert while it is off, and the
dilution veto falls back to its regex floor.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01B3PXXn2tHBzdv6rYtGPz86
TRADING_GAP_SQUEEZE_ENABLED=0 from the next session. The >=150% gap names lost
in both halves of the Jun-Sep history (-0.21R / -0.30R) and in our real fills,
while the 10-35% band made money. 2026-09-28's four trades ran with the
exception on, so they are recorded as a previous period and not counted.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01E3S5EFKThFVrjmN4HHsHMN
At 09:34 ET on 2026-09-28 Alpaca's movers screener returned 'TSX:STC', a
Toronto listing. The discovery filter only dropped symbols containing a dot,
so it joined the universe. The live ingest fetches bars for every symbol in
one batched request, and Alpaca rejects the WHOLE batch with 400 "invalid
symbol". From 09:35 every ingest pass stored 0 rows and the loop ran on
stale bars ("INGEST STALE: freshest bar 458s old").
- Discovery keeps only plain 1-5 letter US tickers from both screeners.
- The ingest never sends a non-US or invalid symbol. If Alpaca still rejects
one ("invalid symbol: X"), it drops X and retries the rest (up to 3 times),
so one bad name costs one symbol, not the session.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01E3S5EFKThFVrjmN4HHsHMN
The 2026-09-24 period ran on settings the 5372585 deploy changed: live volume
was inflated 80-160x, squeezes were verified on a mean volume baseline, and
the first ingest began at ~09:15. It also had no premarket scan and no
fail-fast exit (now 60s/0.3R). Trades from before the restart would mix two
systems, so the 80-trade count starts again on the first session under the
new build. The four earlier trades are kept in the freeze file for the record.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01E3S5EFKThFVrjmN4HHsHMN
Discovery finds premarket squeezes far too late. Across 53 trades by a
professional small-cap momentum trader, our scanner had found only 4 of his
names before his entry, and 27 of his 28 entries were premarket (07:00-09:30
ET). KITT on 2026-09-25 ran $3.50 -> $8.39 at 09:00-09:01 ET (prior close
$0.49); the loop first "Discovered KITT via manual" at 09:35. The existing
sources (Alpaca movers / most-actives screeners) only reflect the regular
session, so a premarket squeeze is invisible until the open.
research/ingestion/snapshot_scan.py sweeps every active, tradable, listed US
equity (~11.9k names; OTC, warrants/units/rights and DISCOVER_EXCLUDE dropped;
asset list cached per ET day) through GET /v2/stocks/snapshots in 2,000-symbol
batches. The endpoint is bound by URL length rather than symbol count (8,000
worked; ~12.7k returned 414). A name qualifies when:
- the LATEST TRADE price is inside WATCHER_PRICE_MIN/MAX (never the prior
close, which is routinely < $1 on these names),
- % change vs the prior close >= DISCOVER_SNAPSHOT_MIN_PCT (20),
- it is trading now: last trade <= 180s old, latest minute bar <= 300s old
with >= 5,000 shares, and >= 100,000 shares since 04:00 ET (summed from
minute bars for the short list, one extra request).
Qualifiers are ranked by % change, capped per sweep, and announced once per
day. dailyBar OHLC ignores extended-hours prints (KITT's dailyBar high was
$3.24 against a premarket high of $8.39), so in premarket it can still be
yesterday's bar. The prior close is prevDailyBar.c only when dailyBar is dated
today, otherwise dailyBar.c.
run_live_paper.py runs it on its own daemon thread (04:00-16:00 ET weekdays,
every 45s), using the plan-019 pattern: its own DB connection, the thread-safe
Watchlist and the generic interval driver. Each new name emits
symbol_discovered with source="snapshot_scan" and price/prev_close/pct/minute
and cumulative volume in symbol_data, so first-sighting latency can be
measured. It gets a daily-history backfill and joins the watchlist, so the
normal new-symbol ingest path gives it its first full-session pull. This is
discovery only; no trading gate changes.
Gated by DISCOVER_SNAPSHOT_SCAN_ENABLED (default 1); knobs in .env.example.
Live read-only sweep at 13:10 ET: 11,871 names in 8 requests / 6.3s cold
(7 requests / 3.5s warm); top hits KITT +376%, MSGY +262%, INLF +60%.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
KITT on 2026-09-25 (+557%, 9.3M shares, the day's biggest mover) was benched
as "gap 557% squeeze unverified (rvol 1.54)". Two separate faults:
1. The rvol baseline was the 20-day MEAN, and two 11-14M-share days on
09-01/09-02 put it at 1.66M against a 333k median. Squeeze verification
(gap >= 150% needs rvol >= 5) now takes the stronger of the mean-baseline
reading and the median-baseline one. Both are cumulative volume over a
full day's baseline, not prorated, so they share the units the rvol >= 5
boundary was calibrated in. A reverse-split fake still reads ~0.01 on
both. This applies in the watcher pre-filter, the trigger book's squeeze
exception and the fast-path population gate. The ranking and the tod-rvol
arm signal are unchanged.
2. A name's first ingest reached back only to ~09:15 (09:30 + 15 min, or
the 20-minute refresh floor). KITT was found at 09:35, so its stored day
began at 09:13 and missed the 09:00 squeeze and ~6M premarket shares.
The first pass now covers the session from 04:00 ET.
With both fixed, KITT's volume at 09:35 is ~20x its median day.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With BROKER=tradier the loop's `client` is a TradierClient that delegates
market data to the Alpaca client in `data_client`. Two places read the feed
as getattr(client.settings, "feed", "iex"). TradierSettings has no feed, so
both read "iex" even though the container runs ALPACA_DATA_FEED=sip (verified
in the running loop and in config_overrides):
- step_vratio kept computing IEX->SIP ratios as if on the free feed
- ingest_live_minute_bars multiplied the real SIP volume by those ratios
Checked against the paper stack's SIP bars for the same minutes: prices are
identical and unscaled volume matches 1.00x, but scaled rows run x83 (GLND),
x117-126 (GRML) and x163 (SKYQ). The 2026-09-25 GRML signal logged a
cum_volume of 1.39 billion shares and an rvol of 64.6. For any name with a
ratio, every volume gate — rvol, the 300k day-volume floor, min dollar
volume, the 2.5%-of-volume liquidity cap, and the gap x rvol ranking — has
been reading inflated numbers since the stack went live on 2026-07-31.
data_feed() reads through data_client when present. The regression test
wraps a SIP client in a TradierClient and fails without the fix.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Pre-registers the real-money test that starts 2026-09-24, the first session
without commission drag. PASS needs at least 80 round trips with a positive
mean R and positive P&L. FAIL is any session the equity floor halts. Changing
any strategy setting in live_test_freeze.json restarts the count, and the
scorecard reports that drift.
The settings were frozen, not tuned, because a study of 239 real fills
re-simulated under the current exit found nothing worth changing:
- Capping how far above the trigger an entry may fill lost R at every cap
from 0.25% to 10%, with either skip or resting-limit handling. Fills more
than 5% above the trigger were the best bucket (+0.66R).
- No break-time factor (spread, price, VWAP distance, stop width, L1
bid/ask size, premarket range) kept its sign across both date halves and
both accounts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Tests whether a PROSPECTIVE forecaster (owner or local LLM) carries signal
the machine levers never found. Picks land premarket via dashboard panel or
research/llm_picker.py (Ollama); research/pick_scoring.py judges them under
a pre-registered bar (n>=40, >=30 sessions, mean >=+0.5%/pick net, median>0,
both halves positive). Picks after 9:30 ET stored but flagged invalid.
Shadow-only: human_picks is never read by the live loop.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Year-replay stop-width buckets: <8% +0.072R, 8-12% +0.120R, 12-16%
-0.054R, >16% -0.092R — the wide cohort holds every recent dollar
disaster (BIYA 16.7%, FCUV day-2 17.5%), where the exit ladder's blind
spot (BE/trail only arm on the way up) rides the full catastrophe
distance. New stopwidth_gate intervention (fail-open without levels) +
registered spec, same forward bar as the pm-high gate; the two cohorts
overlap so the adjudicator attributes which does the work. In-sample
counts for nothing.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
H3 (enter the pm-high break at 9:30 instead of the ORB) FAILED its
pre-registered bar on the year replay: paired delta -0.069R (n=689) —
entering earlier on the same names does not pay; the 5-minute wait stays.
The decomposition found the real signal: ORB fires where RTH price also
reclaimed the premarket high ran +0.316R (n=689) vs -0.298R (n=477) where
it never did — 100th percentile of a 2000-draw random-gate null at the
same keep-rate (null p99 +0.127), sign-consistent 14/14 months. First
lever of 15 to survive both checks in-sample.
Because the split was found in this dataset, it counts for nothing until
fresh fires confirm: new 'pmhigh_gate' intervention kind (fail-open on a
missing premarket_structure row, RTH-only reclaim check at/before the
entry bar) + registered spec with expR_delta >= 0.05 at n >= 200 forward
bar. Cohort self-registers on the next nightly experiments run.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The pre-registered premarket hypothesis H3 (the source trader's actual
entry) made executable: both legs run per symbol-day over the year of
labeled gapper sessions, sharing the deployed exit engine (config parity)
and +1-tick adverse fills so the paired delta isolates ENTRY choice.
Comparison metric locked in the module docstring before the first run:
paired mean R delta (H3-ORB) where both fire, pass at >= +0.05R with
n >= 300 pairs, plus a stop-unit guard (median stop-distance % and
captured-move % printed for both legs — R parity from a tighter stop is
a fail, the vwap-reclaim lesson).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Four of ten armed triggers had NO criteria events ever — the trigger book
arms from the live gapper scan while the watcher scores only the
discovery watchlist, so scanner-only names (TCX/ZTG class) showed an
empty reasoning panel with nothing behind it.
- arm pass injects every complete armed candidate into the watchlist
provider's extra_symbols (its standing manual-add hook)
- provider returns injected symbols FIRST: the watcher truncates at
max_symbols, so extras appended after a full research list were
silently dropped — 'always include manually injected symbols' was
untrue exactly when the list was full
60 watcher/watchlist/trigger tests pass.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
New 'pyramid' intervention kind in the replay engine: at +add_at_r on
initial risk, add add_frac x size (+adverse fill), whole-position stop to
blended breakeven; scratch there if a later bar trades through, else both
legs exit at the baseline exit. variant_r is per unit DEPLOYED
(P&L / ((1+add_frac) x risk)) so leverage alone cannot fake a pass.
Spec pyramid_1r_v1: add 0.5x at +1R, bar expR_delta >= +0.05 at n >= 300,
futility kill at n >= 500 below -0.02. Hypothesis notes the known failure
mode (median winners scratched at blended BE on first pullback). Shadow
only — no live-path changes; cohort self-registers on the next nightly
experiments run.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Distills each gapper's premarket tape (first-hit price/time, pm high/low,
run %, open-vs-pm-high, VWAP distance, jackknife count) into a
premarket_structure table for entry-conversion analysis. Motivated by
FCUV 2026-07-31: our DB captured the full $1.98->$10.49 premarket ramp
while the ORB entered at $11.89 into the reversal.
Pre-registered hypotheses H1-H3 locked in the module docstring before any
data was examined; report command prints descriptive buckets only, with
n>=30/bucket floor and Phase-2 shadow bar required for promotion. Gates
nothing, trades nothing.
Wired into the nightly scheduler after premarket_eval. One session scan
per build (per-symbol queries were ~100x slower); partial session index
created post-close only (mid-market CONCURRENTLY build deadlocked against
live ingest).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
eod_reconcile, backfill_journal and the admin health block were hardwired
to AlpacaPaperClient — on a BROKER=tradier stack they would silently
reconcile/report against the wrong broker. make_trading_client() honors
BROKER env with no silent cross-broker fallback (missing Tradier keys =>
skip with reason, never fall back to Alpaca). Admin UI shows the active
broker label; JSON key stays 'alpaca' for compatibility.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Read-only nightly-scheduler job (Friday-gated): buckets realized P&L by entry
time (open/late-morning/midday) and, for each market_exit close, computes the
post-exit run-up capturable under a breakeven stop (0 if price undercuts entry
first; intrabar ties count as stopped). Emits a weekly_tally event + Telegram
summary. Evidence accumulator for the dead no_new_entries_after cutoff and the
weakness-exit winner-clipping question — no live trading path touched.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The DB-config wiring was unnecessary there anyway — market_context reads
zero trading knobs; the guard test (plan 009) rightly rejected the import.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Version: CI now tags images YYYY.MM.DD-<shortsha> alongside :latest/:sha
and passes GIT_SHA/BUILD_TAG build-args; the running app reports them via
dashboard /api/health and admin /api/status (TrueNAS custom apps only show
'custom/1.0.0', so the app self-reports what it actually runs).
Config parity: every scheduler-run script (nightly_tune, eod_replay,
premarket_eval, regime, market_context, experiments cli) now loads
config_overrides at startup like the live services do — a bare env would
replay under DEFAULT exits (the documented 4.5x-baseline landmine).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Prices each closed round-trip against the nearest quote_snapshots NBBO:
per-trade spread cost, book-walk flag (order size > displayed ask), and a
cost-adjusted P&L total. Measured this week: 0.3-0.5% median spreads on
gate-passing names, ~12% profit drag — far below the static 2-4% microcap
proxy. Coverage gaps (quote recorder only follows tracked names) reported
per-trade, not silently dropped.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Dashboard backtest sources sessions from setup_events v1, which only the
labeler writes — and the labeler was never in the nightly scheduler, so
setups (and therefore backtests) froze at the last manual run (07-17).
Root blocker was _insert crashing on duplicate setup_ids on any re-run;
now ON CONFLICT DO NOTHING, so incremental nightly builds are safe.
Scheduler gains 'labeler build' + 'build-tod1' after eod_replay in both
compose files.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The 07-21 misses (CPHI +417%, DFNS +94%, GREE/WGRX +40%) were all static-gap
<10% names the validated gapper population excludes by design, but they share
a signature no refuted detector (liq1/run1) measured: rvol_tod (cum volume vs
median-20d prorated by elapsed session) at 50-1000x on a flat open. New
labeler setup tod1 fires on the rising edge of rvol_tod>=50 AND price>=
prev_close*1.20, restricted to static gap <10%, stop = 5-bar swing low,
scored via the shared _forward_labels. Shadow-only; pre-registered promotion
bar locked before forward data (median fwdmax60 >=+10%, median dd60 >-6%,
+1R>50%, top stratum rvol_tod>=200 clears alone).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Day-22 forensic (BIYA session): 6 entry submissions, 0 fills — IREX missed
+1.7R, IRE +1.65R by one cent. Two root causes fixed:
1. Slow-path limit rested AT the trigger while price ran 2% past it, and
retries reused the same stale price. Marketable model now prices from
max(trigger, live last) via price_provider; the anti-chase extension
gate gets the live ref too (was an entry_ref==trigger no-op on this
path). TRADING_ENTRY_FILL_MODEL flipped to marketable.
2. Fast trigger path armed nothing all day: legacy rvol (cum-so-far /
full-day MEAN) is ~0.1-0.3 minutes after the open for a normal heavy
name and its mean is poisoned by reverse-split/prior-squeeze days
(BIYA legacy 1.27 vs real activity 481x median). New rvol_tod =
cum / (MEDIAN(20d) * elapsed-session fraction) drives arm eligibility
and the scanner keep-filter; legacy rvol untouched for validated
boundaries (gap-squeeze rvol>=5).
Also: stagnation time-stop (TRADING_EXIT_STAGNATION_MIN/_R, default OFF)
in simulate_exit + manage_live with sweep_exits.py variants for
validation before any live enable.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Plan 017 Phase 0->3. Databento depth backfilled for 602 fires across 21
sessions -> depth_features table. Reweighted to true 23% population win rate,
gating on ask-overhang-above-break (keep>=0.034) lifts win rate 23->28% and
clears the random-gate null (+0.057 vs +0.05 on the conservative +1/-1
bracket). iv_depth_overhang_gate = keep-or-zero, skip-on-no-coverage (the
dilution-v1 lesson). This commit registers the REAL-exit harness run
(simulate_exit asymmetric trail) for the true live delta. Counter-intuitive
mechanism: more resting supply above the break = MORE likely to win
(absorption/fuel), invisible in bar data — the owner's 'size stacked above'
instinct, proven.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
research/attribution.py's _connect() read os.environ["DATABASE_URL"]
directly, bypassing the storage facade fallback chain (DATABASE_URL,
PG* vars, LAN host retries) every other module goes through — and
KeyError-ing (uncaught by name, though still 500'd by the endpoint's
broad except) whenever DATABASE_URL was unset.
Delegate to storage.db.get_connection().raw instead, so DB resolution
matches EventStore/research.multi_schema, and a genuinely unreachable
DB fails with a catchable psycopg2.OperationalError rather than a
KeyError.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Null (the H2 lesson): zeroing share f of a negative baseline mechanically
yields expR_delta = f*|mean baseline|; a PASS on expR_delta_min now also
requires expR_delta_vs_null >= the same margin whenever zeroed_share >= 5%.
Exposure-shrinking can no longer masquerade as selection; non-gate
experiments unaffected. Honest disclosure: regime_meta_v1's FIRE-weighted
pass would not clear this bar either — its standing rests on the session-
level walk-forward OOS evidence, which no random gate reproduces.
H6: CKS order-flow-imbalance from quote_snapshots bid/ask SIZES —
size-weighted book imbalance over the 10-min pre-trigger window, zero
ask-tilted books; skip-on-no-coverage; judged under the null from day one.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
H4 (3-book): keep-or-zero on trigger-bar volume concentration vs prior 5-bar
sum (>1.0 = completed one-shot burst). H3 (5-book, direction disputed):
fragility-camp primary — zero fires whose fill-time relative spread exceeds
2x the name's own 10-min median from quote_snapshots (which carry bid/ask
SIZES since 07-02 — the note's data-upgrade #1 was already satisfied).
Coverage design learns the dilution_veto_v1 lesson: no-quote fires are
SKIPPED (variant None, excluded from n), so thin coverage cannot false-FAIL.
H5 deliberately deferred until H1's verdict (its incremental value is defined
net of H1's catches). 6 tests.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
From docs/research/orderflow_notes.md — the order-flow shelf distillation.
H1 (5-book convergence): exit on first post-fill 1-min close back below the
trigger within 5 bars — the transitory-impact/recross signature — else the
standing exit stack. H2 (4-book): keep-or-zero on pre-trigger volume-weighted
bar-sign persistence (<0.30 over 15 bars = choppy one-shot burst, zeroed);
strictly-before-entry window, fail-open on thin history. Both pre-registered
at the house bar (+0.05R, n>=300, standard fidelity); nightly adjudication
decides. 7 tests.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
replay_window reconstructed fires from ALL setup versions, so liq1/vr1/run1/pm1
(refuted shadow tracks, never traded live, 8:1 volume vs v1, tight-stop
artifacts) dominated /backtest: 17k trades / 45 sessions at -1.8R/trade —
garbage scale vs the harness-validated -0.08R/fire. reconstruct_fires_from_setups
gains an optional setup_version filter (default None: experiments unchanged);
replay_window pins it to v1.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
premarket_eval's shadow enrichment call site queried session_date = current_date
(wrong on UTC-server Postgres at night, so it never matched any symbols) and
shared one try-block for float_cache + dilution filings (a float_cache failure
silently skipped the dilution refresh too). Fix: compute the ET session date
in Python, split into two independent try-blocks, lift into a testable
_refresh_shadow_enrichment(con, session_date). float_cache now applies its own
DDL unconditionally on use instead of only when an API key is present.
Adds iv_dilution_gate (additive, mirrors iv_regime_gate's keep-or-zero shape
keyed on symbol instead of session_date, fail-open on unfetched symbols) plus
its once-per-run preload, and registers the pre-registered dilution_veto_v1
experiment via the existing experiments/*.json self-registering cohort. Live
veto stays OFF; this only wires the shadow measurement.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adds a read-only /backtest page showing how the bot WOULD have traded past
sessions under the CURRENTLY configured live settings: theoretical ORB fires
from setup_events, replayed through the shared exit engine (simulate_exit +
ExitConfig.from_env(), same adverse-fill accounting the exit-experiment
harness uses), screened through the shared population gate and (optionally)
the live regime gate. No new simulation math — replay.replay_window() is pure
assembly of existing pieces (reconstruct_fires_from_setups, _fire_in_population,
_load_regime_favorable, simulate_exit); api/main.py adds the /backtest and
/api/backtest routes plus a dedicated-lock 10-min cache so a slow replay never
blocks the live dashboard's SSE stream. Also extends /api/chart with an
optional date= anchor so the trade drill-in chart modal can reach sessions
older than the existing 72h-from-now window, and adds a one-line BACKTEST nav
link to the live dashboard header.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The 07-15 live gate never made a real decision: live_session_favorable read
today's breadth_rvol from today's OWN setup_events rows, which only the
NIGHTLY labeler writes — always None intraday, fail-open every session (the
07-15 "validation block" was computed pre-guard on a zero-filled feature).
Worse, breadth_rvol was unit-corrupted: _session_context pooled all setup
versions and pm1 stores RAW cumulative premarket volume in relative_volume
(median ~261k) vs the ratio tracks (~0.8) — ~240/267 historical sessions'
breadth was actually median premarket volume, and plan 009's v1 RVOL fallback
was about to flip forward sessions to ratio units.
Fixes (TDD, 21 regime tests):
- breadth_rvol is now pm1-ONLY: median premarket volume — the dominant
historical source, unit-consistent, and strictly pre-open knowable.
- _live_breadth_rvol recomputes today's pm1 fires IN MEMORY from premarket
minute_bars (pure compute_premarket_setups, no insert); previous_close
falls back to the latest prior daily close when today's daily_bars row is
absent intraday (else the gap basis silently flips to the premarket open).
- _regime_reject: verdicts before 9:50 ET are PROVISIONAL (re-scored ~5 min;
ingest lag can truncate the premarket frame) and the first verdict at/after
9:50 is FINAL even when None — caps the scorer at a handful of runs/day
instead of ~78 on quiet days (hot-path stall guard). Throttle arithmetic
moved inside the fail-open try (naive/aware mixing must not escape).
- Live fit excludes breadth_rvol-NULL training rows (zero-fill taught a
"zero premarket volume" class the serve path can never produce).
Prod re-validation on the cleaned feature (regime_features rebuilt, scores
rewritten): regime_meta_v1 fire-weighted expR_delta +0.0570 (pre-registered
bar +0.05, n=2058, PASS — the contaminated +0.0965 was inflated), keeps 37%;
recent-80-session delta +0.149R/session (always −0.141 → gated +0.008).
First genuine live verdict 2026-07-16: UNFAVORABLE (pm breadth 612,819).
Known accepted skew: live scaled-IEX volumes vs nightly SIP tape — the
meta_live-vs-EOD comparison is the standing forward measure.
Also backfilled the missed 07-15 nightly job supervised (SIP refetch, 50 v1
+ 6 pm1 setups, SPY/QQQ ingest, regime row) — plan 009's accrual clock is
unbroken.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Wires the regime signal into the money path — fresh entries are BLOCKED on
sessions the walk-forward model scores unfavorable. Session-level: evaluated
once per day, cached. Enabled in both compose templates (paper).
Correctness (blocker 1 resolved): the LIVE feature set drops tape_regime (the
only non-strictly-open feature — AND measured 0.000 importance) + breadth_gap
(also 0.000). The 5 remaining features are all non-look-ahead at entry time:
breadth_rvol = median(relative_volume=vol_upto/avg_vol_20d) over the session's
setups so far (the SAME up-to-trigger calc the live bot runs, NOT day-end),
spy_* lagged to prior close, trailing from PRIOR sessions. The live single-
session fit reproduces the batch meta_v1 verdict exactly on the last 6 sessions.
Safety: fail-open everywhere — disabled, unfit model (<60 prior sessions),
missing breadth_rvol (pre-setup), or ANY exception returns None and never
blocks. The gate can only SUBTRACT trades on bad-regime days.
- research/regime.py: live_session_favorable() (refit on priors <today, predict
today; writes regime_scores['meta_live'] for forward validation) + LIVE_FEATURE_COLS
+ the breadth_rvol-missing fail-open guard.
- trading_execution.py: TRADING_REGIME_GATE_ENABLED / _P_MIN settings + _regime_reject()
(cached per session) + a REGIME gate after SESSION in _evaluate_entry_gates.
- docker-compose.yml + truenas-compose.yml: TRADING_REGIME_GATE_ENABLED=1.
- tests/unit/test_regime_gate_live.py: block / pass / fail-open(x3) / scored-once-per-day.
KNOWN train/serve skew (documented, validate forward via meta_live vs EOD label):
model trains on full-day breadth_rvol, live uses setups-so-far. Small edge on a
~266-session sample; paper only. Suite: 504 passed, 15 pre-existing loopback fails.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The session-level metalabel P(our ORB profits today) — the one lever that
survives out-of-sample. Backtested on 266 prod sessions (walk-forward):
trade-always (trained subset): -0.111R/session, 36% hit
meta_v1 GATED (trades best ~32% of days): -0.004R/session, 48% hit
delta +0.107R in-sample, +0.128R OOS (recent 80 sessions), improves across folds
It does NOT make the bot profitable — it takes a bleeding strategy to roughly
flat by sitting out the worst ~2/3 of sessions. threshold_v1 rule gates too
little to matter; meta_v1 (walk-forward GBM) is the signal. Leaky full-fit
(+0.61R) >> walk-forward confirms the discipline was load-bearing; the OOS
holding is the evidence the +0.078R is real, not look-ahead.
Changes:
- research/regime.py: batch minute-bar load per session (64k per-fire round-trips
-> ~266 queries; fixed a >5min build to ~3min) + incremental label pass (skip
already-labeled sessions; --rebuild forces full). Nightly cost now ~one new day.
- docker-compose.yml + truenas-compose.yml: scheduler runs build-features + score
(both models) before `experiments run`, nightly. ALL-SHADOW — does NOT gate live
trading (that stays owner-gated Step 5).
- admin_api.py + admin.html: POST /api/regime/run (detached) + GET /api/regime/status
+ a Run/Status button — the panel had no way to trigger this.
- plans/011 + docs/research/regime_notes.md (6-book distillation) + plans 008/009 docs.
Prod state (shadow, rollback = DROP regime_features/regime_scores + DELETE
experiments WHERE id LIKE 'regime%'): 266 feature rows, 532 scores, 2 experiments
registered.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
research/regime.py: Step 0 probe (labelable-session/SPY-coverage STOP gate),
Step 1 build-features (regime_features: pre-open breadth/tape + prior-close
SPY momentum/vol + trailing bot expR, label = mean _sim_baseline_r over the
session's setup fires) + walk-forward threshold_v1 rule, Step 2 walk-forward
meta_v1 (shallow sklearn GBM), Step 4 OOS/leakage-audit helpers
(oos_split/sequential_folds/session_expr_delta/full_sample_fit_threshold).
Step 3 (HMM) deferred per scope — statsmodels stays absent.
replay.py: new iv_regime_gate intervention (keep-or-zero a fire on a
pre-computed session regime score, fail-open on no score row) wired into
_compute_fire + run(), mirroring rank_one_bullet's shape.
experiments/regime_threshold_v1.json + regime_meta_v1.json register the two
model variants against the shared population/bars/fidelity contract.
All new code is unit-tested against isolated EventStore(":memory:") schemas
only — no production DB read or written. Live-trading paths untouched.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
iv_exit_config replays ANY target/breakeven/trail combo through the shared exit
engine vs the live-exit baseline — reusable way to adjudicate exit numbers.
Counterfactual over setup fires: v1 (slow, 1.37R median winners) gains most from
prior_low trailing (win 17->24%, +0.036R); pm1 (fast, 2.64R median, 10R tail)
gains most from NO target + step0.5 (+0.273R, win 24->34%) — the fixed target
throws away pm1's fat tail. Registered both as tracked experiments. +test.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
research/conversion/{__init__,dataset,signal_probe}.py — the leakage-safe probe
plan 009 Step 5 re-runs scoped to a live setup_version:
- dataset.load_dataset: setup_events JOIN outcome_labels; label =
reached_1r_before_minus_1r; features = trigger-time-knowable setup_events
columns ONLY, dropping >70%-null and <2-distinct columns. HARD leakage rule —
every outcome_labels column is blocklisted AND asserted absent from X, so a
schema change can't leak the answer in.
- signal_probe.run_probe: expanding-window CV over ~5 sequential session-date
blocks (whole days, train strictly precedes test), HistGradientBoostingClassifier
(max_iter=250, lr=0.05, max_depth=4, l2=1.0, categorical_features=indices),
per-fold + mean ROC-AUC, top-decile +1R lift vs base, permutation importance on
the last fold. Drops columns constant within a train slice (HistGBM binning).
CLI: python3 -m research.conversion.signal_probe [--setup-version v1].
Tests: the critical leakage guard (zero outcome columns in features), AUC>0.5 on a
synthetic signal + ~0.5 on pure noise, and a static live-path guard. All :memory:.
No live-path change.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Step 3 of plan 009 (config + code; no live-path change):
- research/ingestion/market_context.py: pulls SPY/QQQ full-session minute bars
(SIP, post-close embargo-safe) + daily bars via the existing upsert helpers, so
the labeler's market-context features have a FORWARD source (index ETFs never
enter the gapper universe the live loop ingests). Mirrors sip_backfill's shape.
- both compose scheduler blocks: run it after premarket_eval, guarded || echo.
- NEWS_ENRICH_ENABLED on: added to docker-compose x-app-env anchor (loop +
scheduler inherit it); truenas-compose + .env.example already had it. Catalyst
is a FEATURE source only, never a live gate; enabled-but-down never blocks
trading. Enrichment OUTPUT is verified by the orchestrator (needs live Ollama).
Compose statically validated (yaml.safe_load). Ingest self-checked via a
fake-client :memory: round-trip through _preload_spy.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
