signal-platform/docs/research/regime-monitor-v4.md

# AI/Tech Risk Monitor v4 methodology

Named "Regime Monitor" until 2026-08-07; the filename's `regime` stem, the
`regime_monitor` job id, the `/regime` route and the `METHODOLOGY`/snapshot
fields keep the old word, because those are persisted or externally linked.

The AI/Tech Risk Monitor is an observational risk thermometer. It does not
gate entries, exits, position size, ranking, or alerts about individual setups.

**v4 supersedes v3** (2026-08-08). Unlike v3, whose calibration was ad-hoc and
never landed, every number below is reproducible:

    .\.venv\Scripts\python.exe scripts
un_regime_monitor_calibration.py ^
        --methodology v2_reconstruction,v3,v4 --cache-dir .calib-cache

The harness replays the 408 sessions ending 2026-07-24 from the live inputs
(Alpaca for all 33 symbols, FRED for VIX and HY OAS) with no database, and
reproduces the published v2 and v3 figures before it will emit anything:

| figure | published | replayed |
|---|---|---|
| v2 State avg | 22.6 | 22.68 |
| v2 State p80 | 35.1 | **35.1** |
| v2 State max | 91.2 | **91.2** |
| v2 P3 pegged | 39 | **39** |
| v2 W1 live | 108 | **108** |
| v3 State max | 87.4 | **87.4** |
| v3 band shares | 73.3 / 15.0 / 8.3 / 3.4 | 73.0 / 15.4 / 8.1 / 3.4 |

It refuses to emit a band recommendation, and exits non-zero, unless every hard
gate passes — 33 symbols fetched with full warm-up, the whole basket on every
session, the calendar anchors, 100% coverage on every row, and a row-wise
`state_v4 <= state_v3` invariant. Reading a calibration result out of a run whose
pipeline did not validate is meant to be structurally impossible.

## What changed in v4

**V1 stopped saturating at VIX 30.** `(vix - 15) / 15` reached 100 at VIX 30 —
the same defect v3 had *just* removed from P3, left in place one sensor over. VIX
30 is a bad week, 50 is a crisis and 82 was March 2020, and all three scored
identically. In the calibration window this flattened five distinct April-2025
prints (52.33, 46.98, 45.31, 40.72, 38.57) into a single 100. It pegged on 14 of
408 sessions; under the anchors below, none.

**The trend break is graded by depth, not a yes/no.** `_under_200` returned a
bare 0/100, so P1 printed 100 the moment SMH and QQQ were both under their
average — and because the price pillar takes `max(P1, P2, P3)`, that pinned the
pillar and stopped P3's anchored ladder resolving anything for the whole of a
selloff. It pegged on 46 of 408 sessions; now none. A 2% break reads ~30 where it
used to read 100.

`max()` was **kept**. The defect was the step function feeding it, not the vote
itself, and v3's "one capped vote for correlated reads" rationale still holds.
Over the window P1 is the sole price argmax on 17 of 408 sessions (4.2%), so it
informs the pillar without owning it — the `P1_SCORE_CAP` fallback considered
during design was measured as unnecessary and not shipped.

**The top State band moved 80 → 65.** See Calibration; this is the one change
that is about the band rather than a sensor.

**Scope.** All three are State-side. `WARNING_BANDS`, `WARNING_WEIGHTS`,
`QUADRANT_WARNING_DIVIDER` and the event study's frozen threshold are untouched.
`QUADRANT_STATE_DIVIDER` stays 50 because only `breaking` moved.

## What changed in v3

**Fundamentals left the score.** F1 (capex) and F3 (good-news-stock-down)
carried 12 + 8 of 100 Warning points. Pegged at maximum stress they produced a
Warning of exactly 20.0 — below the event study's 25.3 alarm threshold, and
still inside the "stable" band. The sourced observation could not change any
published conclusion, so refreshing it looked like it did nothing. They are now
a qualitative overlay reported beside the scores. Capex also stopped scoring
`raising` and `holding` identically at 0: `holding` is the deceleration case and
now scores 50, so a boom no longer reads the same as a stall.

**The drawdown sensor stopped saturating.** v2 used `dd_pct * 5`, reaching 100 at
a 20% drawdown — the 90th percentile of the observed distribution. 39 of 408
sessions sat at exactly 100 with no resolution left, and the price pillar showed
the top band on 13.5% of sessions. v3 uses named anchors with headroom past the
observed 36% maximum, and blends leader/confirm 2:1 as P1 and P2 already did
instead of taking `max()`. P3's realized share of State falls from 65% to 40%,
matching its nominal weight.

**Warning gained a sensor with range.** The HY OAS *level* is pinned at zero
below the 3.5 mild anchor (2.77 at the cutover), so credit contributed nothing
in a calm tape. Its 20-session rate of change still does, and spread widening is
a classic lead.

**The credit percentile leg was removed.** Its reference window silently shrank
from 10 years to 3 when ICE restricted the upstream series in April 2026, after
which it scored 20 points of stress at a spread the same sensor's anchors call
"mild". See Calibration below.

**Breadth loss counts during declines.** v2's divergence gate was
`price_ret >= 0`, so the sensor zeroed during every selloff. On 2026-07-24 the
basket shed 10 points of participation in 20 sessions while SMH fell 11.9% and
Warning printed exactly 0. v3 tapers to a floor instead: deterioration counts
fully when price masks it (true divergence, the dangerous pre-top case) and at
35% when price confirms it. Breadth *level* lives in State, but breadth
*velocity* appears nowhere else, so this is not double counting.

**Bands are per axis.** v2 Warning never exceeded 64.9 in 408 sessions while
State reached 91.2, yet both used 30/60/80 with quadrant dividers at 60. The
upper half of the Warning axis was unreachable.

## Outputs

**State** — current structural stress:

- Price structure, 40%: `max(P1, P2, P3)`, one capped vote for correlated reads.
- Fixed-basket breadth level, 25%.
- HY option-adjusted credit spread level, 20%.
- VIX level, 15%.

**Warning** — deterioration and divergence:

- Fixed-basket breadth divergence, 45%.
- 60-session SMH/SPY relative-strength deterioration, 30%.
- HY OAS 20-session widening, 25%.

Combined, RSP/SPY (former F4), and the NVDA canary (former P6) do not enter v3 or v4.

## Calibration

### Interpolated sensor tables

All three are `(x, stress score)` pairs read by `_interpolate`, flat outside the
first and last anchor.

| sensor | anchors |
|---|---|
| P3 drawdown (% below the 52w high) | 0→0, 4→10, 8→25, 16→50, 28→78, 40→100 |
| **P1 trend break** (% below the 200-DMA) | 0→**20**, 3→35, 8→55, 15→75, 25→100 |
| **V1 volatility** (VIX level) | 15→0, 20→20, 25→38, 30→55, 40→80, 55→100 |

P1's floor of 20 at the crossing is deliberate: the break itself is a genuine
binary event and deserves a floor; only the depth past it is graded. P1 is
calibrated to sit alongside P3 rather than swamp it — the 200-DMA lags, so a 20%
drawdown typically coincides with ~10% below the average, where P1 reads ~61
against P3's ~59.

V1 reaches full scale at 55 rather than at 2020's ~82: anchoring the top at a
once-in-a-generation print would make VIX 50 — a genuine crisis — read only ~70.
The anchors encode the long-run distribution as constants, the same argument the
credit level uses. Unlike P3 and V1, whose slopes ease off monotonically, P3's do
not (2.5, 3.75, 3.125, 2.33, 1.83) — its gentle onset is intentional and the
monotone-slope test excludes it.

Credit impulse is relative (+35% over 20 sessions = 100) rather than absolute,
because +0.5pp means something very different at an OAS of 2.7 than at 8.0.

### Bands

Round, meaning-anchored numbers, **not** percentile fits — those would drift on
every rebuild and silently rewrite what past snapshots meant.

**Why `breaking` moved 80 → 65.** With credit calm, `f2_credit_spreads` returns
`0.0` (not `None`), so it keeps its full 20 points pinned at zero. Price, breadth
and volatility at *literal maximum* therefore sum to:

    (100×40 + 100×25 + 0×20 + 100×15) / 100 = 80.0 exactly

`band_for` uses `>=`, so v3's top band was reachable only by touching its floor
to the decimal, with nothing above it. The band was fit on v2, when credit's
since-removed percentile leg still contributed regularly; the sensor is not
wrong — a calm-credit selloff genuinely *is* less stressed than one with credit
contagion — the threshold was stale.

Chosen by scenario arithmetic on unchanged weights (`_scenarios` in the harness
computes these, so they are machine-checked, not prose):

| scenario | price | breadth | C1 | V1 | State |
|---|---|---|---|---|---|
| Ordinary tape (3% dd, breadth 65%, VIX 16, OAS 2.8) | 7.5 | 0 | 0 | 4.0 | **3.6** |
| 10% correction, calm credit (2% below, breadth 35%, VIX 24) | 31.2 | 62.5 | 0 | 34.4 | **33.3** |
| **2022-style drawdown, calm credit, no death cross** | 90.8 | 100 | 0 | 60.0 | **70.3** |
| **same, with death cross** (P2 pegged) | 100 | 100 | 0 | 60.0 | **74.0** |
| Credit event on top (OAS 6.0, VIX 45) | 100 | 100 | 75.0 | 86.7 | **93.0** |
| March 2020 (everything pegged) | 100 | 100 | 100 | 100 | **100** |

Rows 3 and 4 are the case this monitor exists to measure, and they must print
`breaking`. At 80 they do not. **65** clears them under either P2 assumption,
which matters because P2 is set by the 50/200-DMA gap and no drawdown figure
implies it; 70 would have left 0.33 points of headroom in row 3, reproducing the
defect being fixed.

Realized shares, **reported not fitted**, over the 408 sessions to 2026-07-24:

| Axis | stable | watch | elevated | breaking | thresholds |
|------|--------|-------|----------|----------|------------|
| State (v4) | 78.9% | 13.0% | 4.7% | **3.4%** | 20 / 50 / **65** |
| Warning | 69.4% | 19.6% | 7.6% | 3.4% | 20 / 40 / 60 |

The v4 `breaking` share lands on 3.4% — the same as v3's — having been chosen by
scenario reasoning rather than aimed at that number. Sensitivity: 60 gives 5.1%,
70 gives 1.2%.

Quadrant dividers sit at each axis's watch/elevated boundary: State 50,
Warning 40. Only `breaking` moved in v4, so the dividers and every alert
threshold are unchanged. `test_quadrant_dividers_match_the_band_boundaries` now
enforces that relationship, which nothing did before.

Scores renormalize over available fixed weights, but a band is published only at
75% or greater coverage. Trend deltas are suppressed when the participating
pillar set changes. Zero means ordinary/healthy; only stress contributes.

Credit level is the named HY OAS anchors alone: 3.5 mild, 5.0 elevated, 7.0
stressed, linear between, and nothing else. v2 blended those anchors at 70% with
a 30% upper-tail percentile over a nominally 10-year window.

That leg was removed rather than repaired. ICE restricted FRED to a rolling
3-year window for `BAMLH0A0HYM2` in April 2026 — the series metadata states it
outright ("Starting in April 2026, this series will only include 3 years of
observations"), and an unbounded request returns the same 795 observations as a
30-year one. The v2 percentile therefore ranked the current spread against three
uniformly tight years (range 2.59–4.61 over the calibration window), which made
it fire early and saturate absurdly: at an OAS of 3.50 — the level the anchors
call *mild*, scoring zero stress — the blended sensor read 20.1, and the
percentile leg pegged at 100 by an OAS of 4.5. Across the 408 sessions it
roughly tripled the credit sensor's average (2.70 vs 1.00) and more than doubled
its nonzero days (60 vs 27).

The anchors already encode the long-run distribution as constants, so the
percentile was a second, noisier estimate of the same thing. What it was
genuinely reaching for — "unusual versus recent history" — is now W3 on the
Warning axis, computed as a rate of change, which is where deterioration
belongs. Removing it moved State's average by −0.4 and its maximum by −3.8, left
Warning bit-identical, and did not shift any band threshold.

A long-history alternative (`BAA10Y`, Fed-published, 7,712 observations back to
1997) was considered and rejected: ranking an HY spread against investment-grade
history is not a coherent statistic, and it would rescue a leg that is redundant
anyway.

Every snapshot now records `data_quality.credit_history_days` and
`vix_history_days`. This defect was invisible for roughly three months because
nothing asserted the window the code claimed; the spans make a future upstream
truncation show up in the record instead of quietly reshaping a sensor.

**Survivorship caveat.** The basket was frozen 2026-07-15 but the calibration
window reaches back to 2024, so names were partly selected for having done well.
Every distribution above inherits that bias. It is the same bias v2 carried, so
the v2/v3 comparison is like-for-like, but the absolute band shares are
optimistic.

## Point-in-time record

The first run under a new `METHODOLOGY` rebuilds the latest 400 trading sessions
with sufficient sensor warm-up; routine runs thereafter insert/update only the
latest trading date. The history API and main chart show only snapshots matching
the current methodology, so a bump reseeds the series rather than splicing two
formulas into one line.

The fundamental overlay keeps its effective date (normally the next session after
collection) and is never replayed backward, so a rebuild cannot stamp today's
observation onto historical snapshots. Because the observation is stored in a
single slot, a refresh replaces the previously effective record: the snapshot
therefore reports the overlay as `pending` until the new effective date.

Two functions, deliberately: `fundamental_overlay` is the **record** and keeps
the gate — it runs for every replayed date during a rebuild, so it must never
grow a bypass flag. `current_observation` is the **live reading** behind
`fundamental_context`, and *reports* the effective date instead of blanking the
content.

Until 2026-08-07 the live reading called the gated function, so a just-collected
observation stayed hidden until the next weekday — three days over a weekend —
and refreshing appeared to do nothing. That was the opposite of what this section
already claimed. Showing it early cannot leak into a published number, because
nothing in the overlay is scored (see "Fundamentals left the score").

`current_observation` gates on `observed` (a non-null `fetched_at`, the one field
every path writing real content stamps). Without it, the default override —
`unknown` for every hyperscaler and `mixed` for the reaction — was reported as a
live observation with `available: true`, so the card presented placeholders as a
collected reading. Those are the absence of an observation, not an observation of
absence. `fundamental_overlay` never had this problem: no observation means no
effective date, which means `pending`, which already blanks the content.

Each snapshot stores the fixed basket symbols, hash, and freeze date.
Reconstructed history before that freeze date is retrospective/exploratory.

## Presentation

The page is deliberately thin: two gauges, one chart card, one pillar table, the
overlay, and a provenance strip. Time and Path are two projections of the same
snapshot series and share one card and one query key — they were previously two
panels, which read as two datasets. Methodology rationale lives in this document,
not on the page; page text is limited to what changes how the reader interprets
today's number. The quadrant dividers rendered in Path view come from
`quadrant_config` and are the same constants the alert path consumes
(`alert_service`), so the chart cannot drift from what actually fires.

## Warning study

The study calls the outcome a **10% correction**, not a regime break. The first
70% of sessions freezes the 80th-percentile warning threshold; alarm episodes are
measured on the final 30%. Because v3 dropped fundamentals from the score, the
study now measures exactly the live Warning score rather than a technical-only
approximation of it, and both are computed from one shared sensor definition
(`warning_sensor_scores`) so they cannot drift apart.

A cached report is discarded when its methodology no longer matches, so the panel
reverts to "not run yet" after a bump rather than showing stale numbers. **Re-run
the Event Study job after cutting over to v4.**

### Reading the result

The report carries a `reliability` block and the UI renders its warnings, because
the headline numbers invite over-reading in two specific ways.

**The holdout is thin.** The study detects 11 corrections across 5 years but the
70/30 split leaves only 4 in the test period. Recall is therefore one event away
from a materially different headline, and in practice the event that flips is
decided by where the frozen threshold happens to land rather than by whether the
score saw anything. The v3 cutover run illustrates it: v3 scored 2/4 against v2's
3/4, but "v3 without the credit sensor" scores 3/4 at a *higher* threshold
(35.5) than shipped v3 misses it at (32.3) — because the alarm rule needs a
rising edge, and a lower threshold can mean the alarm already fired outside the
20-session horizon and never reset below. Below `MIN_EVENTS_FOR_CONFIDENCE`
holdout events the report says so explicitly.

Some events carry no information at all for comparison: in that run every
variant caught 2026-03-06, every variant missed 2026-06-05, and every variant
"caught" 2025-11-20 with a 1-session lead, which is coincident rather than a
warning.

**Sensor coverage can straddle the split.** The score renormalises over available
sensors, so a training window predating a sensor's history freezes the threshold
on a different construct than the holdout is measured against. At the v3 cutover
only 39% of training sessions had all three Warning sensors versus 100% of the
test period, because credit history begins 2023-07-25.

Restricting the threshold to sensor-matched training sessions was tried and is
*not* the fix: those sessions are a calm recent stretch, so the threshold drops
from 32.3 to 22.5 and false alarms rise from 3.3 to 8.6 per year. It trades a
coverage bias for a regime-selection bias. The honest position is that the
threshold is hypersensitive to window choice at this sample size; the report
states its limits rather than pretending to a precision it does not have.

## Resolved in v4 (raised 2026-08-07, shipped 2026-08-08)

The three questions this section used to hold are now answered. Kept here
because the reasoning that resolved them is not obvious from the code.

**1. `breaking` had zero headroom — resolved by moving the band, not the sensor.**
`f2_credit_spreads` returns `0.0`, not `None`, below the 3.5 mild anchor, so
credit stays *available* at weight 20 and is pinned at zero on roughly 93% of
sessions rather than being renormalized out. Price + breadth + volatility at
literal maximum therefore summed to exactly 80.0 — v3's threshold, to the
decimal.

The sensor is **deliberately unchanged**. A calm-credit selloff genuinely is less
stressed than one with credit contagion, so scoring it lower is correct; what was
stale was `STATE_BANDS`, fit on v2 while credit's since-removed percentile leg
still contributed. Making credit `None` when calm was considered and rejected: it
would leave State on 80% coverage, which still publishes, but consumes the whole
buffer — any *second* missing pillar would then suppress the band, and the 7d/30d
trend deltas would null out every time OAS crossed 3.5, because `_delta`
suppresses on a change of participating pillars. See Calibration for the
scenario arithmetic behind 65.

**2. V1 saturated at VIX 30 — resolved with an anchor table.** See "What changed
in v4".

**3. `max(P1, P2, P3)` defeated P3's anchoring — resolved by grading `_under_200`,
keeping `max()`.** The `max` was deliberate ("one capped vote for correlated
reads") and survives; the binary step feeding it was the defect.

**Its limit, stated precisely.** `_death_cross` is `clamp(-gap_pct * 20)`, so P2
pegs at a −5% 50/200-DMA gap — routine in a real downtrend. In a *deep* selloff
the price pillar therefore still reaches 100 via P2 even with P1 graded. What v4
repairs is the shallow-to-moderate break, which is where resolution was most
obviously missing: a 10% correction 2% below the average now scores 31 where v3
scored 100. It would be wrong to claim "the price pillar no longer pegs".

P2 did not peg once in the 408-session calibration window, so this is a property
of the sensor rather than an observed problem. Grading P2 the same way is the
natural next item if it starts binding; the replay reports a P2-pegged census
alongside P3 and V1 so the evidence accumulates.

## Fixed 2026-08-07: the OAS fetch window did not cover a rebuild

`HY_OAS_WINDOW_DAYS` was 400 **calendar** days, but a rebuild replays
`leader_series[-REBUILD_SESSIONS:]` — 400 **trading** sessions, about 579
calendar days. The oldest ~180 calendar days of any rebuild therefore got no OAS
data at all, so `f2_credit_spreads` and `w3_credit_impulse` both returned `None`.
Verified: State then lands at 80% coverage and Warning at exactly 75.0% —
`MIN_COVERAGE` — so **both still publish bands**. The rebuilt series would look
homogeneous while its oldest rows had been scored without credit, the tell being
a null `data_quality.credit_history_days` on exactly those rows.

The window is now 700 days: it must cover the oldest replayed date (~579) plus
W3's lookback and slack, while staying under ICE's ~3-year cap so FRED still
honours the request. This required **no methodology bump** — C1 reads
`oas_values[-1]` and W3 reads `oas_values[-21]`, both indexed from the end, so
widening only prepends older observations and every live score is bit-identical.
Confirmed by evaluating both windows against a varying synthetic series: today's
C1/W3 match exactly, while the oldest rebuild row goes from `None`/`None` to real
values.

Expect `credit_history_days` on new snapshots to rise from ~400 to ~700. That is
the widened request, not new upstream history — and it makes the chip a better
truncation canary, since a 700-day request returning ~1095 days' worth is now
the visible ceiling.

**Widening the window alone does not repair stored history.** Routine runs
recompute only the latest trading date, and `rebuilding` was keyed on "no v3
snapshot exists at all" — which is false once the cutover has run — so every row
already written would have kept its credit gap indefinitely. `SENSOR_REVISION`
fixes that: it is stamped into each snapshot, snapshots predating it read as 1,
and a stored revision below the current one triggers exactly one reseed.

It is deliberately not `METHODOLOGY`. That constant partitions the history API
and discards the cached event study; neither is warranted here, because the study
recomputes its Warning series from source (`_warning_series` calls
`warning_sensor_scores` against freshly fetched prices and OAS) rather than
reading snapshots, so a reseed cannot stale it.

The reseed is bounded by `REBUILD_LOOKBACK_DAYS` in calendar days rather than a
session count, because the binding constraint is the OAS fetch: each replayed row
needs W3's 20-business-day lookback inside `HY_OAS_WINDOW_DAYS`. At 672 days the
replay reaches ~464 sessions, W3's oldest requirement lands exactly on the first
fetched OAS day, and the ~400-session series the v3 cutover wrote is fully
covered. A test asserts that relationship so the two constants cannot drift into
recreating the gap.

The fix was sequenced deliberately: acting on items 1–3 above bumped
`METHODOLOGY`, which fires `rebuilding`, which would have baked the credit-less
rows into the fresh series. Fixing the window first meant the v4 reseed replayed
a clean window; doing it the other way round would have meant reseeding twice.

## Operator rule

Quadrant alerts default off for new/reset configurations. When enabled they
require fresh inputs, at least 75% coverage on both axes, two consecutive daily
confirmations, hysteresis, and cooldown. Every alert states: **Risk thermometer —
not a trade signal.**