Depeg Duration Resolver Changelog: v4 archive

Impact notes

• Unique promoted pending outcomes contribute first-seen and last-seen deviation at their recorded times; matching assessments contribute current deviation at assessment time. Untimed final event peaks, pending peaks, and assessment aggregate peaks cannot establish earlier landmark depth.
• Retained samples are sparse evidence, not a reconstructed price tape. Historical incidents without timed support remain unsupported through grouping and duration estimation instead of manufacturing comparable observations.
• Historical training filters false-positive, disputed, no-data, and unknown non-null audit verdicts before the row cap. Confirmed, repaired, and legacy null-verdict events retain eligibility.
• Live-price admission rejects nominal references and unavailable or nonpositive supply instead of fabricating an at-peg deviation. Resolution weights, duration landmarks, and support thresholds are unchanged, but changed input admission can move assessments and future verdicts.
• The current-policy accuracy gate changes accuracy, duration error, and horizon calibration only. Whole-universe accountability, missed locks, publication debt, and scoreable-coverage reporting remain independent of that accuracy cohort.
• Calibration breakdowns use full-universe counts within the headline policy. Live forecast totals include valid predictions only; pending, deferred, no-call, and invalidated audit states do not contribute verdicts or durations.
• Per-coin track records use producer-computed counts and median duration errors from before the public row cap. Retained snapshots without those aggregates show unavailable until refreshed, never a statistic reconstructed from the browse sample.
• Release requires fixed-clock replay, mover attribution and owner review, plus fresh timestamp-evidenced duration calibration. Archived golden histories without timed provenance are not evidence of current calibration; no successful recalibration is claimed.
• Activation is provisionally 2026-10-11 00:00 UTC (1791676800); re-date effectiveAt at release

Impact notes

• DDR continues to exclude false_positive, disputed, and no_data while retaining confirmed, repaired, and legacy null verdicts
• Bound SQL eligibility is derived from the shared audit vocabulary and the DDR-specific exclusion policy
• Migration 0252 rejects new unknown verdict writes without rewriting archives or changing sealed prediction history
• Resolution weights, duration landmarks, and existing invalidation authorization requirements are unchanged

Impact notes

• A curated optimistic label no longer overrides an adverse published mint posture in duration strata or mint-risk signals
• V9's mint-scoped none-resolved posture maps to none-resolved-mint unless the curated review independently establishes whole-chain absence, preserving the distinction between weak and strong recovery anchors
• Published unknown posture remains authoritative; an absent derived posture retains the curated fallback
• Factor weights, duration landmarks, and sealed prediction history are unchanged; new structural classifications can move where derived and curated postures differ

Impact notes

• `stateAssignedPct` and `finalizedCoveragePct` group review rows by `incidentKey` and give each incident exactly one unit, however many review rows it carries, assigned from the rows' single `predictionState`; an operational cause is an overlay on that state, not a second assignment, and neither share is weighted by row count or depeg duration
• Both shares are at most one by construction rather than by clamping: a data-quality gap or orphan closure attributed to a cron gap, system deferral or missed lock previously counted in both its state bucket and the missed-lock bucket, publishing an assigned share above one
• An incident whose rows disagree about its state is reported as unassigned, so the shortfall stays visible instead of inflating the assigned share
• Published values fall wherever the old double count applied: `finalizedCoveragePct` drops for any window where an operational cause overlapped a data-quality or orphan-closed state, and `stateAssignedPct` no longer exceeds one — it still reads below one when policy-universe incidents have no review rows or disagree about their state
• Per-state counts, `missedNoPredictionCount`, `operationalMissRatePct`, prediction/no-call rates and all accuracy metrics are unchanged

Impact notes

• Same-direction pending episodes reset after an observation gap greater than 1200 seconds, so a blind interval cannot backdate a confirmed onset
• Recovery confirmation stores both first and last qualified observations, resets after a gap greater than 1200 seconds or contradictory evidence, and keeps events open through missing data
• Explicit recovered-primary, recovered-dex, and recovered-native reasons define recovered labels even when native quote-domain policy stores no recovery price
• Direction supersession, coverage loss, orphan cleanup, and unknown explicit closures cannot enter the recovered duration corpus or DDRR recovered outcomes through a stray recovery price
• Legacy rows with null close reasons retain recovery-price compatibility so historical recovered labels are not erased

Impact notes

• Posture set membership moved to `isFragileMintPosture` / `isUnboundedMintPosture` / `isNoPrivilegedMintPosture` / `isNoPrivilegedMintChainPosture` in shared/lib/safety-score-v9/mint-posture.ts; K1's risky-minter leg, K1's severe-surge rung and the structural-class legs all read the same source, so a future posture value is classified once instead of at each site
• 14 wrapper assets annotated `none-resolved-mint` (sUSDe, sDAI, sGHO, scrvUSD, stUSDS, sBOLD, sDOLA, srUSD, sdUSD, said, steakUSDC, steakUSDT, yvUSDC, eEARN) now publish the R1 recovery anchor at the weak rung during an event, labelled `No privileged mint authority on this token (wrapped supply can still expand)`
• The strong R1 rung remains whole-of-chain only: a wrapper's parent can still print, so `none-resolved-mint` cannot satisfy the strong-anchor leg of the tier rule. Only strong anchors enter that rule, so no verdict, sealed prediction or duration stratum moves
• Structural-class stratification is unchanged: `none-resolved-mint` still confers no robustness of its own, matching the whole-of-chain definition in shared/types/core.ts

Impact notes

• K1's risky-minter test reads the published V9 mint posture band; the concentrated and exposed bands stay the risky set, so the rule is unchanged and only its input moved engines
• A run with no installed V9 publication derives the band from the curated authority posture instead, so a degraded or held publication cannot silently drop K1's band leg; an installed publication is authoritative, including when it publishes no band for an asset
• K5 inputs are unchanged: it already read published Safety Score outputs and continues to do so
• No factor weights, severity bands, duration landmarks, or incident-lifecycle rules change

Impact notes

• Issuer wind-down evidence and measured exit or supply stress provide earlier terminality context, while backing concentration is gated by mechanism and observed impairment
• Duration labels follow canonical incident grouping, typical ranges use the 15th-85th percentiles, and comparable histories are deduplicated by coin
• Recovered pre-lock incidents can close and resurrect within the merge window, while regime-escalating tails begin a separate incident and prediction
• DDRR review rows expose repaired and split lineage and publish expected-versus-observed horizon calibration alongside realized outcomes