Depeg Tracker and DEWS Changelog: v6 archive

Impact notes

• Native-origin events keep their structured quote identity through peak updates and recovery. Native confirmer peaks for USD-origin events are converted with the admitted peg reference; mismatched assets, currencies, or incomparable peak domains are withheld. Canonical currency spelling is accepted, and a changed peg identity resets pending onset.
• Policy-rejected pools cannot supply promoted peak statistics or source attribution. A single pool above the existing $5M exception must satisfy the same opposing-group vote as multi-group confirmation; TVL alone cannot outvote contradictory evidence.
• Historical event-floor and large-cap confirmation checks use USD-valued supply, including timestamp-matched FX or commodity references. Observed zero remains measured zero; unavailable supply is absent. A degraded non-empty history with no admitted supply snapshots preserves existing backfill events.
• Current at-peg summaries use the detector's raw-precision trigger boundaries, not rounded display deviations. Missing or nominal primary prices do not erase independently trusted DEX divergence from DEWS market evidence.
• The foundational stablecoins generation must pass its cadence-plus-runtime freshness budget. PSI amplification requires a valid sample within two producer intervals; stale or invalid retained samples hold the accepted DEWS generation rather than refreshing its clock or silently amplifying a new one.
• Failed or malformed canonical incident projections withhold recomputed history and analytics instead of substituting raw events or invented counts. Retained accepted analytics keep their original observation clock.
• Event archives are no longer limited to 20 pages, and Worst Depeg uses complete eligible incident history. Dated history exports identify retrospectively captured ledger state, not historical point-in-time knowledge.
• Historical DEWS diagnostic detection rates exclude unevaluable events and disclose partial coverage and missing supply anchors. Signal history distinguishes unavailable measurements from explicitly inapplicable signals.
• Formula weights and numeric trigger thresholds are unchanged. Owner-approved production-history replay and fixed-clock admission and band comparison remain release prerequisites; no replay or band-equivalence result is claimed.
• Activation is provisionally 2026-10-11 00:00 UTC (1791676800); re-date effectiveAt at release

Impact notes

• Without an eligible weekly baseline, both historical score and TVL anchors remain null/unavailable rather than using weak or out-of-window observations. This changes input availability and can change DEWS scores and bands.
• The live DEX history lower bound is now minus 8.5 days. Historical reconstruction reads DEX history from the earliest event UTC day minus 15.5 days; supply history bounds are unchanged.
• Weak public DEX history remains visible, and the existing confidence >=0.75 durability and digest gates remain unchanged. No collector, Liquidity, PSI, DDR or Safety scoring admission is promoted; depeg detection rules are unchanged.
• Release requires owner-approved production-history replay under U-C20-16. No production replay, mover, band-equivalence or runtime-health result is claimed by this entry.

Impact notes

• Per-asset decode diagnostics retain source, context, identity and reason; scoped rejections do not enter global degradedSources
• Quarantined latest rows are removed without rewriting sparse history, daily evidence or retained accepted publication buffers; healed assets rejoin normally
• Whole-source failures and unscopable cohort-invalidating inputs still withhold publication; an empty admitted cohort preserves the previous pointer
• Optional unreadable pool detail retains component-only unavailability; formula weights, trust floors and freshness budgets are unchanged

Impact notes

• Additive migration 0255 preserves old-Worker queries and records trusted off-peg and at-par intervals, the latest trusted price clock, and three-state observation metadata without a historical backfill
• Blind/missed observations, state transitions, or gaps over 1200 seconds break interval continuity; consecutive trusted at-par observations retain known at-par time instead of being labeled blind
• PegScore and its recent 90-day companion exclude unknown event spans from the time denominator and duration severity; historical measured peaks remain evidence, but stale open lifecycle alone no longer activates a current-depeg penalty
• Uninstrumented open events have unknown duration until new trusted observations arrive; closed legacy/replay events retain recorded durations
• Sub-$1M assets with observed prices receive currentDeviationBps and pegReference; event creation still requires $1M supply and all existing trust/confirmation gates

Impact notes

• The five known verdicts share one schema and separate PegScore/DDR eligibility policies
• PegScore still excludes false_positive and disputed, includes no_data, and retains legacy null-verdict evidence
• Unknown verdicts are excluded from direct scoring and SQL recomputation; strict archived parsing rejects them without coercion to null
• DEWS excludes nominal par references from price inputs; absent observed primary and DEX prices leave divergence unavailable rather than measured calm
• No detection threshold, DEWS formula, known-verdict weight, or stored historical row changes

Impact notes

• `computePoolSignal` in `worker/src/lib/dews/signal-families.ts` divides the readable weighted sum by the readable weight (`componentCoverage` 0.75 or 1) and publishes `worstPool: null`, `worstPoolStatus` (`observed` / `empty` / `unavailable`), `worstPoolUnavailableReason` and `componentCoverage` on the pool signal; with every component readable the formula is numerically unchanged
• Top-pool parsing in `worker/src/lib/dews/scoring.ts` no longer defaults a missing `extra.balanceRatio` to perfect balance (1.0) or turns an unreadable entry into a zero-TVL balanced pool; pools without a balance measurement are excluded from the worst-pool max, and an unreadable list or entry records a per-asset `dex_liquidity.top_pools_json` diagnostic (`degradesRun: false`) that degrades only that asset's pool signal, never the whole generation
• Pool smoothing now also requires the previous reading's `componentCoverage` to match the current one; a previous reading with different or unrecorded coverage (every pre-6.29 row) is not averaged in, and the outcome is published as `smoothing`. The first post-deploy cycle therefore publishes unsmoothed pool values
• Reproduction: balance ratio 0.5 and average stress 50 with absent pool detail previously scored 37.5 with `worstPool: 0`; it now scores 50 with `worstPool: null` over 75% component coverage, matching the same data with an observed $1M pool at ratio 0.5
• Captured-data sizing: 18 of 378 DEX assets have a raw pool-signal increase of one third, including USDT 29.5 -> 39.3, USD1 29.3 -> 39.0, XAUT 30.3 -> 40.4 and CADC 54.2 -> 72.3. These are pool sub-signals, not composite DEWS scores; the original DEX capture did not retain the other DEWS signals or prior smoothing state
• A separate current-time estimate uses the 2026-09-27 22:56:35 UTC published cohort (237 assets), its 22:26:35 predecessor and concurrently captured pool inputs. Replaying the old composite reproduces all 237 published scores; applying the new unsmoothed pool signal to those same inputs changes SBC CALM 14 -> WATCH 17, XSGD CALM 15 -> WATCH 16 and KRWQ WATCH 35 -> ALERT 37. Other signals and PSI are held fixed, with the evidence cap and both contagion passes recomputed; this is not a forecast of changing market inputs
• For a frozen-input next-cycle comparison using the latest published row as smoothing history, only SBC and XSGD change band between methodologies: old KRWQ already reaches ALERT 36, versus new ALERT 37. Neither SBC nor XSGD is Telegram-alertable. KRWQ is alertable, but the captured Telegram last-alertable baseline is already ALERT, so none of these transitions produces a new DEWS dispatch candidate under the existing rules; no suppression machinery is added

Impact notes

• `deriveDecisionContext` in `worker/src/cron/depeg-detection/decision-engine.ts` reads supply through `getCirculatingRawOrNull()` instead of the zero-defaulting `getCirculatingRaw()`; a `null` reading marks the open event seen, emits a per-asset `Kept live event ... current supply is unavailable` warning, and issues no persistence command, so recovery timers, peaks and `ended_at` are untouched
• Previously an empty supply record read as $0 and closed the open row with `close_reason = 'coverage-lost-supply'` and `recovery_price = NULL`, removing it from PSI's open-event set and the resolver's live board although no supply observation had fallen below the floor
• Observed sub-floor supply, the $1M threshold, recovery windows, trust gates and every DEWS formula are unchanged; previously stored coverage closures are not rewritten by this change

Impact notes

• `decideRecovery` in `worker/src/cron/depeg-detection/decision-engine.ts` gains a recovery lane for assets with no fresh trusted aggregate DEX row, alongside the authoritative/fresh-multi-source primary and the corroborated aggregate-DEX row: `derivePoolChallengerEvidence` reports `recoverySupported` when the fresh published challenger snapshot holds at least `POOL_CHALLENGE_CONFIRM_MIN` independent protocol/source-family groups inside the recovery band (50 bps USD, 75 bps non-USD) that strictly outvote the groups still crossing the trigger in the event direction (`corroboratingProtocolGroupsOutvote` in `worker/src/lib/constants.ts`); the `>= $5M` single-pool carve-out still vetoes, and while a fresh trusted aggregate row is available it remains the only DEX recovery lane while the snapshot only vetoes
• The lane persists the TVL-weighted median of the corroborating pools as `recovery_price` with `close_reason = 'recovered-dex'`, and logs a `Pool-challenger majority recovery` diagnostic naming the group counts; the 15-minute continuous recovery window, the 1200-second continuity tolerance, and the deadband rule are unchanged
• Root cause: `computeDexPrices` only publishes an aggregate `dex_prices` row for an asset whose primary price already clears the primary trust gates (`loadTrackedStablecoinMaps` feeds the publisher a trust-filtered price map and the row is withheld as `primary-missing` otherwise), so the aggregate-DEX recovery lane was structurally unreachable for exactly the `confirm_required` primaries that need independent corroboration
• Reproduction from live D1: event 90781 (`usdb-blast`, CoinGecko single-source, open since 2026-08-20 at +129 bps peak) sat at -1 bps with five published challenger pools (thruster-v3, monoswap-v3-blast, blasterswap; three independent groups) inside the band and no aggregate row; event 90760 (`vnxau-vnx`, gold peg, open since 2026-08-05) sat at +3 bps with raydium and aerodrome inside the band
• Guard evidence: event 90786 (`hollar-hydrated`) stays open on the single $11.9M hydration-dex pool still printing -526 bps, and event 90777 (`audf-forte`) stays open with one corroborating curve group below the two-group bar; a 2-vs-2 tie resolves to no recovery and clears any partial timer
• Onset behavior is unchanged: opening or refreshing a depeg still requires the primary to cross the trigger threshold, and the challenger snapshot is not consulted for onsets

Impact notes

• The primary-recovery pool veto (`derivePoolRecoveryVeto` in `worker/src/cron/depeg-detection/decision-engine.ts`) fires only when the diverging challenger groups reach `POOL_CHALLENGE_CONFIRM_MIN` and are at least as numerous as the challenger groups inside the trigger threshold that corroborate the recovered price; the single-pool `>= $5M` TVL carve-out is unchanged
• Pool-only pending-depeg confirmation applies the same majority test against the groups contradicting the pending direction (opposite-direction or back inside the bar), so two diverging groups no longer confirm while four independent protocols sit at the reference rate
• The rule is one exported authority, `divergingProtocolGroupsOutvote` plus its complement `corroboratingProtocolGroupsOutvote` in `worker/src/lib/constants.ts`, shared with the pricing pool challenge (`selectReplacementProtocolGroups` and its confidence downgrade, pricing v6.32) so the pricing and depeg lanes cannot drift; it keeps the shared `POOL_CHALLENGE_CONFIRM_MIN` = 2 independent-group bar
• Reproduction: live event 90792 (`vchf-vnx`) opened 2026-09-18 from a +640 bps print that two dormant venues carried — the Celo Uniswap v3 VCHF/USD₮ pool (last trade 2026-03-15, provider-reported $4.7M reserve against ~$141 on-chain, 24h volume 0) and the ICP kongswap VCHF/ICP pool (24h volume 0) — against four live protocols at the ECB CHF rate; the veto had blocked the primary recovery on every subsequent detection run
• Trigger thresholds, recovery window timing, pool TVL minimums, the aggregate-DEX corroboration rule, and every DEWS formula are unchanged; only protocol-group precedence in the two pool-challenger decisions moves

Impact notes

• Replay skips a recomputed `source='backfill'` episode when its inclusive `[startedAt, endedAt]` interval overlaps a reviewed entry for the same coin and direction in `shared/data/depegs/backfill-replay-suppressions.ts`; that registry is validated at module load and guarded by the registry-integrity suite in `shared/data/depegs/__tests__/`
• Replay also skips an episode when an existing `source='live'` row for the same coin and direction overlaps it (`episodeStart <= liveEnd && liveStart <= episodeEnd`, inclusive; open rows collapse to their start second) — the two detectors read the same price series at minutes and hourly granularity, so same-side overlapping windows are one episode counted twice, while one second apart stays disjoint
• Both skips run in `executeBackfillForCoin` before the dry-run/apply split, so the preview, the run fingerprint (`expectedEventCount`/`expectedFingerprint`), and the inserted rows stay mutually consistent; each skip is ledgered as `backfill-depegs-episode-skipped` with `skipReason: reviewed-suppression` or `live-overlap`
• The four seeded suppression windows cover the Noon USN upside artifacts for 2025-10-07, 2025-12-27/28, 2026-02-14, and 2026-03-02, whose on-chain forensics found no economic-size USN trade above 1.0011; the ledgered 14 Feb live/backfill pair is caught by both rules
• Live detection, thresholds, confirmation policy, price extraction, and the PegScore/DEWS formulas are unchanged; a one-off read-only production dry-run found 85 overlapping live/backfill pairs across 52 distinct backfill rows on 21 coins, and each affected coin drops at most its distinct backfill-row count at its next window replay

Impact notes

• When any non-bootstrap source fails or a core persisted input row is malformed, the run computes and stores rows but does not advance `dews:published-generation`; `/api/stress-signals`, the `/depeg` panel, PSI and Telegram keep serving the previous proven generation and its scores
• Renormalisation over the surviving signals is unchanged; it simply no longer reaches the published surface, so an outage in the 0.35-weight liquidity block cannot publish a lower score as a calmer market
• The withheld cycle is ledgered on `surface_publication_generations` as `rejected` with its degraded sources, and the freshness sentinel keeps its existing healthy-run-only gate, so publication recency and input quality now agree
• The score formula, weights, amplifiers and band ladder are unchanged

Impact notes

• Pool and divergence smoothing now requires the previous generation's signal to have been available; after a cycle without pool or DEX-price evidence the next fresh reading is published un-dampened instead of halved
• An absent, blank or unrecognised `priceConfidence` scores as the worst known tier (80) rather than the best (0), and the signal carries a `price-confidence-unmapped` marker
• Yield warnings that sum to no scored stress — an unrecognised key, or the recognised non-stress `zero-yield` flag — mark the yield signal unavailable instead of publishing a clean zero
• A tracked coin whose cache row carries no circulating buckets at all is skipped for the cycle; only an explicitly present zero supply retires its current and 7-day stress rows
• Signal weights and the CALM/WATCH/WARNING/DANGER band ladder are unchanged; every effect moves scores in the conservative direction or removes a signal from the weighted average

Impact notes

• Suppressed EURC mirror-zero rows no longer contribute to DEWS 24-hour or 7-day event counts or create digest blacklist candidates
• Unsuppressed rows retain their existing time windows, event-type eligibility, and canonical stablecoin attribution
• Both recent counts and the historical spike baseline change; lower event counts do not guarantee a lower stress score

Impact notes

• Native-origin candidates still use their direct native quote and the universal 15-minute persistence window, preserving detection of moves that USD peer references can mask
• At promotion, the fresh canonical USD price is normalized through the authoritative FX reference and must cross the full threshold in the same direction
• An independent normalized canonical price at peg or in the opposite direction rejects the candidate; stale, cached, fallback, or CoinGecko-only canonical evidence cannot promote it
• Repeated observations from one CoinGecko native feed now prove temporal persistence only, not independent source confirmation

Impact notes

• Depeg suppression, pending confirmation, challenger-pool checks, and promoted DEX pricing now share the 75-minute maximum trusted age
• The $1M aggregate DEX depth gate, protocol corroboration rules, $250K public Price Check gate, direction logic, and temporal confirmation windows are unchanged
• Rows older than 75 minutes remain excluded rather than being carried forward as current confirmation evidence

Impact notes

• Primary deviation, DEX deviation, and cross-source spread are read from the canonical depeg signal, so DEWS and the detector can no longer drift apart on the same price pair
• The three sub-signal readings are now canonical whole basis points rather than unrounded values truncated to two decimals; sub-bps differences no longer move the divergence curve
• A non-positive price or DEX price now yields no deviation instead of a nonsense magnitude from the inline arithmetic
• The divergence curve, the non-USD dampening, the previous-reading smoothing, and every other sub-signal family are unchanged

Impact notes

• Small-cap, multi-source, extreme, direction-change, and native-currency onsets all enter `depeg_pending`; no source combination bypasses the 15-minute temporal window
• Secondary confirmation uses the full trigger threshold, and DefiLlama's `coingecko:{id}` mirror no longer counts as independent evidence for a CoinGecko-family primary
• Live recovery begins only within 50% of the trigger threshold and persists `recovery_first_seen_at`; deadband or renewed-depeg observations reset the timer before closure
• Trigger comparisons use unrounded basis-point deviation while persisted and displayed values remain rounded
• Telegram depeg dispatch deduplicates at stablecoin incident level and formats native-domain EUR, GBP, and JPY prices in their stored currency

Impact notes

• Live fiat FX references take precedence over peer medians for detector, peg-summary, detail, and feedback-verification paths; peer medians remain a fail-closed fallback requiring at least three contributors
• Feedback verification uses the same 100 bps USD and 150 bps non-USD threshold policy as live depeg detection
• Reviewed replay coverage can explicitly set a coin's PegScore denominator; BRLA coverage begins at the audited 2026-06-28 replay window instead of implying stable history before Pharos observation
• Peg summary publishes current reference evidence, coverage provenance, and a coverage-aware recent 90-day view
• Stablecoin history and API responses distinguish public incidents from their constituent threshold-crossing rows

Impact notes

• Native-quote events update peaks only from subsequent native quotes against the stored `1.0` peg reference
• A recovered native quote closes the event with `recovered-native` and the same-domain native recovery price
• When only a USD primary or DEX recovery is available, the event can still close but persists `recovery_price = NULL` rather than mixing quote units
• The known BRLA event 90509 mixed-unit recovery value is cleared through a guarded data migration

Impact notes

• Direct native CoinGecko quotes remain able to veto, sustain, and resolve non-USD depeg rows; they can now also initiate the row when they are the clearest fresh signal
• Small and mid-cap native-quote events open immediately with the native quote stored against a `1.0` peg reference, matching the historical native-fiat replay model
• Large-cap or extreme native-quote triggers still route through `depeg_pending` before public event creation
• This prevents BRL peer medians from masking true BRLA/BRL discounts while preserving the existing 150 bps non-USD threshold

Impact notes

• A mature 30-day baseline no longer produces available zero flow stress when the mint/burn pipeline has no fresh 24-hour row
• Fresh zero-volume 24-hour rows still contribute zero flow stress when the source is current
• Stale mint/burn hourly input is recorded as a `mint-burn-hourly-freshness` source failure and surfaced through `sourceAges.mintBurn` and `staleFlags.mintBurn`
• `flowBaselineDays` remains the baseline coverage diagnostic, while `flowDataAgeDays` now reports true mint/burn source freshness age in days

Impact notes

• Single-source or same-family extreme prints still route through `depeg_pending` before public event creation
• Fresh severe moves with at least two independent primary depeg source families, such as CoinGecko plus DefiLlama, can open a live event immediately below the $1B large-cap floor
• Large-cap assets still require the existing pending-confirmation path even when the severe move is multi-source
• The change prevents confirmed small- and mid-cap crashes from being hidden until DEX confirmation catches up

Impact notes

• A single high-TVL pool (>= $5M) or at least two independent protocol/source-family pool groups can veto primary-price recovery closure
• This covers cases where aggregate off-chain prices briefly print near peg while a large venue remains materially depegged
• Small single-pool disagreements remain advisory and do not block authoritative primary recovery by themselves
• The change hardens long-running incidents such as the APXUSD June 2026 underpeg against transient near-$1 blips

Impact notes

• Authoritative primary recovery still closes an event when there is no qualifying DEX contradiction
• When trusted aggregate DEX price and at least two protocol groups still cross the threshold in the existing event direction, the event remains open and the contradictory recovery is logged
• This prevents brief off-chain or consensus near-peg prints from splitting unresolved long-running incidents such as the APXUSD June 2026 underpeg
• DEX-confirmed recoveries remain supported: ambiguous primary recoveries can still close when aggregate DEX and enough protocol groups are back inside threshold with no challenger contradiction

Impact notes

• The `structured-high-risk-venue` (+25) and `structured-medium-risk-venue` (+10) branches are unchanged, but `sourceRisk.venueRiskTier` is now derived from a 5-category Yearn-style score across 61 venues instead of 12 hand-set tiers
• Newly-scored high-risk venues (e.g. Clearpool, Goldfinch, 3Jane, Aries Markets, Curvance, Avantis) now contribute `structured-high-risk-venue` to the Yield Anomaly sub-signal for coins routing yield through them
• Newly-scored medium-risk venues (e.g. Fluid, Dolomite, Fraxlend v2, Felix, Centrifuge, Jupiter Lend, Sovryn) now contribute `structured-medium-risk-venue`
• Reviewer-set cross-venue dependency concentration raises the source-risk penalty, which can cross the existing +20 `structured-source-risk-penalty` threshold for affected rows
• No DEWS code, weight, threshold, or cap changed; unknown/low venues remain a no-op

Impact notes

• A lone non-USD coin's peer-median reference equals its own price (deviation always ~0) and a 2-coin group mirrors half of any real move onto the healthy peer; detection has always failed closed on these, but peg-summary, the depeg tracker, and the coin-detail hero kept publishing the masked number
• When the gate fails, currentDeviationBps is withheld (null) and the new pegReferenceUnavailable flag drives an explicit 'reference unavailable' readout in the tracker table and detail hero
• USD, commodity, VAR/OTHER pegs and groups with a live FX fallback or at least 3 peer contributors are unaffected
• Depeg detection, PegScore, and event history are unchanged; this aligns the displayed numbers with the detection engine's existing trust policy

Impact notes

• Structured Yield Intelligence evidence with `sourceRisk.venueRiskTier = "medium"` now adds +10 to the Yield Anomaly sub-signal
• Medium venue-risk rows emit the `structured-medium-risk-venue` warning
• The existing high-risk venue branch remains +25 with `structured-high-risk-venue`
• The final Yield Anomaly sub-signal still caps at 100 after adding warning-string, source-risk, and rank-attribution evidence
• Missing, malformed, unknown, low, or otherwise neutral venue-risk evidence remains a no-op and does not create an available zero-stress yield signal

Impact notes

• Thin non-USD fiat peg references without FX fallback use the stored pending `peg_reference` when valid instead of recomputing confirmation against a 1-2 coin peer median
• Pending rows with neither an authoritative refreshed reference nor a valid stored reference wait for a safer reference instead of being deleted or promoted
• Duplicate open-event repair merges only same-direction rows and closes older opposite-direction rows at the newer direction boundary with `recovery_price = NULL`

Impact notes

• Coins with no prior supply-history anchors no longer receive an available zero-stress supply signal by default
• When only one prior anchor is present, DEWS still computes the available side and treats the missing side as zero velocity contribution
• Explicit finite zero anchors remain available and continue to produce zero velocity stress instead of divide-by-zero behavior

Impact notes

• A coin that depegged above the live-event floor no longer stays LIVE forever after supply shrinks below $1M
• Below-floor closures keep `recovery_price = NULL`, matching other coverage-lost closures where Pharos cannot assert a price recovery
• New below-floor coins still cannot open fresh live depeg rows; the change only resolves existing open rows whose coverage leaves the live-event universe

Impact notes

• Blacklist counts are hydrated from event provenance (`config_key` / `contract_address`) instead of being grouped by bare symbol
• Same-symbol siblings that do not have direct blacklist-tracker coverage now keep the blacklist sub-signal unavailable
• Legacy event rows without provenance fall back only to a single tracker-owned stablecoin id for that symbol, avoiding fan-out across same-symbol PSI assets

Impact notes

• Absent previous-day or previous-week supply history still defaults to the current supply so new or incomplete rows do not create false contraction
• Finite zero buckets in the stablecoins cache now remain zero when the supply-velocity input is assembled
• The supply sub-signal continues to treat zero baselines as zero velocity stress, avoiding divide-by-zero while keeping the input provenance explicit

Impact notes

• Curated launch dates remain the preferred tracking anchor, followed by the earliest supply_history snapshot
• For priced assets without either anchor, the stablecoins cache now contributes a durable first-observed valid-price timestamp through the same first-seen cache used by PegScore
• Newly observed priced assets still need at least 7 days of tracking before PegScore is rated; missing-price assets and pure NAV tokens remain NR
• Low-cap assets below the live depeg-event floor can receive a PegScore anchor, while the existing coverage-limited flag continues to warn that empty event history is not full depeg coverage

Impact notes

• Backfill runs persist replay status, counts, fingerprints, source providers, quote mode, peg-reference/supply source, confidence tier, and public provenance projections
• Audit verdicts are direction-aware and persist confirmed, disputed, false_positive, no_data, and repaired-style outcomes without deleting rows solely because audit data is weak or absent
• PegScore excludes false-positive/disputed events and downweights low-confidence rows while surfacing quality-adjustment counters to callers
• DEWS calibration metrics now use stored stress_signal_history as the primary source and report precision, recall, false-positive days, false-negative incidents, lead-time percentiles, churn, and cohorts