Live: https://bensumm.github.io/The-Ledger/

A self-contained web app for finding and managing OSRS Grand Exchange flips. Vanilla JS, no build step, no framework, no bundler — plain static files served by GitHub Pages. Installable as a PWA (works on desktop and iOS home screen), but primary development is now on desktop.

Every price/margin in the app is after the 2% GE sell tax (floored per item, capped at 5m, none under 50gp). Price convention follows flipper usage: Buy = the instasell price (where you place buy offers), Sell = the instabuy price.

• Finder — ranks flippable items by a budget-aware rating (profit/hr × a quality dampener blending ROI, liquidity, stability, and turnaround; hover the Risk grade for the per-factor breakdown). Stability here is a cheap live-price-vs-guide proxy — the full regime-drift check lives on the Trends page. Sortable by rating, profit/hr, margin, ROI, or volume. Typing a search query reveals every mapped match — including cheap items (soul rune ~300gp) the browse-mode price floor normally hides; those search-only rows show — for the rating columns and lean on the quote button + star.
• Trends — deep per-item analysis. A live "Suggested plan" (instant buy/sell + patient pricing that sizes a wider-margin offer off the recent 2h range), a regime-shift guard that warns when a recent price-level jump makes the hourly-timing stats unreliable, a plain-language guide-divergence readout ("Why this trend?"), 3-month price history, and a collapsible timing & seasonality section gated on a walk-forward backtest (hourly charts only appear when the timing edge is actually proven out-of-sample).
• Watchlist — star items to track their live margins each refresh; the same starred set is the repo-shared watchlist the pipeline scans. (The old Signals tab — live cheap-window buy signals — was removed in 0.57.0; Ben didn't use it.)
• Ledger — per-item grouped open/closed positions with after-tax realized/unrealized P/L, summarized in the "Coffer" header tiles. Item names link to Trends; multi-lot groups expand via a chevron; a "P&L by" Day/Week/Month control on the Closed-flips header drives a period strip whose buckets click to filter the table by sell date; the manual-entry form is a collapsible section and the closed columns are sortable.
• Fill-data pipeline — see pipeline/ (below): captures real GE trades from RuneLite to fills.json so the tool can eventually calibrate its predictions against actual fills.

• index.html — the app shell (markup only)
• styles.css — all styles
• js/ — app logic as ES modules: state.js (shared mutable state as one STATE object + constants + persistence + diagnostics), money-math.js (the tax/margin/bond MATH — canonical tax()/netMargin/netMarginQty + the ONE BOND tax exception BOND_ID/isBond/bondFee: a bond is tax-exempt but pays a 10%-of-guide retrade fee, so netMargin(low,high,{bond,guide}) = sell − (buy + fee), tax-free; plus the generic clamp/now), money-format.js (gp/number DISPLAY formatting — fmt/fmtSig/fmtP/fmtTurn/fmtHour/fmtHourRange + ukHourOffset/localTzAbbrev/pad2/parseGp/sgn/grade/gradeCls; split out of the old format.js in the R2 rename. fmtHourRange(startH,endH) labels a diurnal hour-of-day window in BOTH the runner's local zone AND the UK — e.g. 01:00–03:00 PDT / 09:00–11:00 UK — so peak/dip window narration can't slip between the Pacific-local hours the tools emit and the UK-driven demand basis; the offset is Intl-derived per instant, DST-correct for both zones), charts-static.js (static inline SVG — svgLine/svgBars, fixed-size, no interaction; still used by the Trends hourly seasonality charts + the quote sparkline), charts-interactive.js (CL — the reusable interactive SVG chart: createChart(container,{series,overlay,fillBetween,refs,bands,markers, kind,yFmt,xFmt,spans,span}) → {setSpan,destroy} handle, with pointer-drag PAN, wheel/pinch ZOOM about the cursor, a span selector, y-axis auto-rescale to the visible window, a hover tooltip + crosshair, and a DEFAULT_SPANS export. Max zoom-in is density-floored (~4 sample points, medGap*4) so you can't zoom into empty space; explicit span buttons bypass that floor to hit their exact duration. Degrades to a no-op on a missing container or empty series. An optional overlay second line + fillBetween shading draws the Forward-forecast low/high CONE (0.61.0; absent config ⇒ byte-identical single-series). Consumed by trends.js — the Recent movement (2h), Price history (1/7/30/90d windows), Diurnal timing (7d/28d toggle), and Forward forecast (cone) charts; more surfaces adopt it over time. ADDITIVE — charts-static.js stays intact), marketfetch.js (shared browser fetch layer — one timeout-guarded jget
  • one cached fetchTs/fetch24h store, A2), market.js (price/guide fetch + scoring; keeps the bond in the catalog — searchable — with a bond-aware Finder margin via bondMarginOpts; AP4 desirabilityOf computes the Finder's shared rank + Desirability grade off js/estimators.mjs+js/rating.mjs. vol24Of(id) (0.74.2) is THE ONE HOME for turning STATE.VOL24 into a daily volume — every reader goes through it so the absent-vs-unavailable rule cannot be re-derived per call site: an item ABSENT from a PRESENT /24h map traded ZERO (measured — 0.0% of absent items traded during the day the map covers), while a null return means the MAP is unusable and the consumer must keep its wide prior. Consumed by desirabilityOf and by trends.js's item header), trends.js (archive + seasonal analysis + regime/patient/backtest — renders the Trends view; pure analytics live in trendcore.js; TV — also renders the Diurnal timing section via the shared windowread.mjs hourProfile/deriveDiurnalRange + a charts-interactive.js bar chart + an inform-only validate.mjs reachValidator note — the same computation the console prints), trendcore.js (TC1 — pure DOM-free Trends analytics: hourly/seasonal decomposition, the walk-forward backtestPlan gate, patientTargets offer sizing, bestWindow/median; moved out of trends.js for pipeline/test/trendcore.test.mjs), quotecore.js (DOM-free quote model + canonical market-table cells — computeQuote/regimeDrift/quoteCells; shared byte-for-byte with the node analysis scripts; also home to recentDirection (DP1), flushSignal (DL2 — the reactive liquid-flush firing read, consumed ONLY by pipeline/commands/watch-positions.mjs --dip, no app import), and nominateDip/reconcileDipPool/pruneDipPool (DL4 — the scan's flush-SUITABILITY nomination + the quality-ranked, self-pruning pool write; selectNominations is the legacy dedup/cap, retained + tested but no longer on the write path — all consumed ONLY by pipeline/commands/screen-flip-niches.mjs, no app import); also the ONE type-7 quantile/median home (SF-1): quantileSorted (pre-sorted input) + quantileOf/median (sort a copy) — termstructure.mjs re-exports it as quantile, retrojoin.mjs aliases quantileOf; and home to grossFromNet (PLAN-SALE-LOG-TAX — the EXACT sell-side inverse of tax(): the smallest integer g with g − tax(g) === round(net), found by a ±2 scan around round(net/0.98) then net + TAXCAP (the cap region falls out of tax saturation). NOT breakEven (that answers ≥ and lands 1gp high where two consecutive g share one net — the per-item floor makes ~every-50th pair collide, so an ask at an exact-2% point recovers 1gp low: display-only, realised never routes through it; fractional net inverts the rounded value). The FALLBACK gross-recovery for .json-era net worth in the reconstruction + the raw-row display sites — since SLT C5 those prefer the plugin's own recorded tax (worth + tax, exact) when the row carries it; bond-blind like tax() — the quarantined-bond caveat applies), forward-reach.mjs (the shared FORWARD-SCORING primitives over the 1h archive — touchedAt (bid side, avgLowPrice), reachedWithin / maxHighWithin (ask side, avgHighPrice), covers, firstIndexAfter, and the END-OF-WINDOW bail pair endLowWithin / endHighWithin off one lastPrintWithin(series, from, windowH, field) (chunk 4 homed them here; js/reach-surface.mjs had carried a private endLowWithin twin). The two bails are DIFFERENT POLICIES for a seller whose ask was never reached — cross the spread into a standing bid (avgLow, chunk 1's choice) or rest at the ask level (avgHigh) — and join-exit-ev.mjs measures that they rank contenders differently, so a consumer must say which it charges rather than inheriting one. Lives in js/ as of PLAN-REACH-SURFACE chunk 0 (it was pipeline/lib/market/forward-reach.mjs; that path is now a one-line re-export shim so every pipeline importer resolves byte-identically, the pipeline/lib/signal/estimators.mjs pattern). The move is what lets js/reach-surface.mjs build the surface from the same walk the joiners score with — js/ never imports pipeline/, so a primitive both sides need has to live here. covers is the load-bearing one: an unresolved window is DROPPED, never counted as a miss, because counting it biases every rate DOWN by exactly the truncation at the end of the archive. Side convention is pinned by js/quotecore.js — a BUY fills against avgLow, a SELL against avgHigh. Pure: no fs, no fetch, no clock; the caller supplies the ts-ascending series. ⚠ It is shared by join-asym-outcomes.mjs + join-reach-outcomes.mjs + js/reach-surface.mjs ONLY: join-reach-basis.mjs still carries its own independent scoreForward walk, and pipeline/lib/market/printed-at.mjs a third — so a fix here does NOT reach every reach joiner. Consolidating them is unfinished work, not a claim to repeat; printedAt additionally has the better tristate (null when no bucket exists, which reachedWithin collapses to false)), reach-surface.mjs (PLAN-REACH-SURFACE chunk 1 — the empirical reach surface p(ask, H) for ONE item, replayed from its own 1h archive. buildReachSurface(series, opts) walks origins every strideH hours and scores a z-grid of ask levels against maxHighWithin, returning the p(z,H) grid plus refHigh/disp, bailZOnMiss (the miss payoff, in z), the per-horizon refusal, and a reported-never-applied 1h-vs-5m grainBias. surfaceProb(surface, ask, H) reads a live ask off it; surfaceShape reduces the curve to z50/z20/spread; referenceAsOf exposes the point-in-time reference. It re-derives nothing: refHigh is windowread.mjs's recentQuant(days,'ask',0.5,3) and disp its iqr (exported from that module for this — it was module-local), both fed by windowStats, whose pt.timestamp shape is bridged from the archive's .ts here rather than in windowStats, where every other caller already passes timestamp. Everything is @provisional-api until chunk 3's read-exit-surface.mjs consumes it — chunk 4 scores the surface before anything prices off it. Four properties are load-bearing and each has a killed mutant in pipeline/test/reach-surface.test.mjs: (1) no look-ahead — refHigh/disp are re-derived at every origin from complete days strictly before it, so a level scored 90 days ago never sees day 91; (2) unresolved windows are DROPPED, never scored as misses, and an origin whose window held no printing bucket is dropped too and COUNTED (noPrintDropped) because the archive cannot tell a quiet hour from an unfetched one — that drop biases p UP; (3) refusal is a WIDTH bound — a cell is thin by its Wilson half-width (Wald reads 0 at p=0 and would price an empty cell as certain), computed on nIndep (origins thinned to non-overlapping windows) rather than the raw count; a HORIZON refuses on its decision cell, the one nearest p=0.5, because "every cell thin" can never fire; (4) the grain bias is reported, never applied — passing fiveMin must not move a grid cell, and fiveMinCoverage rides beside every delta since a ~0 delta on a thin item means NOT MEASURABLE rather than unbiased. The z axis is (ask - refHigh)/disp, not a % grid: a raw %-above-median grid inverts the trend split. z-monotonicity needs no cleanup — it holds by construction (one origin set, one threshold per origin; measured 0 violations in 22,500 adjacent pairs, so the specified isotonic pass was deleted rather than shipped inert); the H axis DOES invert because its origin set shrinks with H (155 violations over 250 items, max 12.1pp) and keeps a running max. REACH IS NOT FILL — queue position is invisible in bucketed aggregates, so p bounds P(fill) from ABOVE as a working bound, not a theorem: the folded plan's §6.1 RETRACTED the theorem reading — the 1h-average instrument pushes p the other way on liquid items, so the net error's direction is item-dependent there (read it off grainBias, never assume it) — and every consumer must say so. What chunk 1 MEASURED — including that PLAN-REACH-SURFACE §1.5's taxonomy premise did not survive and its ordering inverted — is §1b of the folded plan (git show bdea911:plans/PLAN-REACH-SURFACE.md, folded into PLAN.md 2026-08-30); don't restate it here), exit-ev.mjs (PLAN-REACH-SURFACE chunk 2 — the inversions that turn a reach surface into a PRICE. evCurve(surface,H,{bailNet,delayCost}) scores every level of one horizon at EV = p·net(ask) + (1−p)·(net(bail) − delayCost) (tax from money-math.js, the ONE definition); askStar is its argmax, askForHorizon the highest level clearing pTarget, and horizonForAsk the SMALLEST horizon clearing it — returned with the full p-by-H row, so the threshold never travels alone; p is read through reach-surface.mjs's surfaceProb, the ONE interpolation, so every row carries its ciHalf. Six properties are load-bearing and each has a killed mutant in pipeline/test/exit-ev.test.mjs: (1) EV has an interior maximum on real curves — the property the co-log scorer lacked, and without it nothing can rank; (2) the miss payoff is PER-CELL (bailZOnMiss) because that is the conditional expectation the decomposition asks for — and its MEASURED DIRECTION is the opposite of the intuition: at a low ask only a catastrophic window misses, so E[bail|miss] is WORST there and rises toward the unconditional value as the ask climbs, which means per-cell prices at or ABOVE the unconditional form, never below (an earlier version of this entry, the module header and the test all asserted the reverse); (3) delayCost is charged to the MISS branch only — on both branches it is a constant at fixed H and cannot move the argmax at all, so the asymmetry is what makes waiting cost anything; (4) a maximum on the last SCORED z is a refusal, not a price (the edge is the scored one, not the declared grid — a dropped cell would otherwise pass as interior); (5) net() is the ONE tax definition on BOTH legs — a tax asymmetry is 2% of price, larger than the whole EV spread being optimized over; (6) horizonForAsk reads horizons ascending regardless of grid order. Consumed by chunk 3's read-exit-surface.mjs, which is why the @provisional-api markers are gone from every export but askForHorizon — that one is now @test-only and has no production consumer BY DESIGN, since §1c forbids its ask as a price. askStar is an argmax over a PLATEAU, not a point — adjacent cells sit within a few basis points of refHigh of each other, decided off a p known to a few pp, so a consumer must present a band and never a false point; read-exit-surface.mjs owns that. The chunk-2 stop-or-go gate and what it measured live in the folded plan's §1c — PLAN-REACH-SURFACE folded into PLAN.md 2026-08-30; full text git show bdea911:plans/PLAN-REACH-SURFACE.md (don't restate its numbers here — they were, once). The one rule to carry out of it: pTarget must never pick a price. It answers "how long", never "how much". ⚠ And chunk 4 has now measured what askStar is worth: scored against realized net gp over the archive it LOSES to a deployed incumbent, decisively enough to fire the plan's pre-registered null branch. Everything above is still how the arithmetic works; none of it is a reason to price an offer off askStar. See join-exit-ev.mjs), windowread.mjs (P2 — pure window-range/reach math: windowReliability (DT4, 2026-08-10 — the split-half hours gate: parity-split the last WINDOW_RELIABLE_NIGHTS (14) days, hourProfile each half, Pearson-correlate the de-trended devLow/devHi 24h vectors, gate on min(rLow,rHi) ≥ WINDOW_RELIABLE_R (0.6, PLACEHOLDER). Returns a TRI-state reliable true/false/null — null = not measurable, deliberately distinct from a measured fail. Pins its OWN 14-day window off the RAW series, never the caller's nights. Consumed by diurnalTimedLap (lap.reliable), formatTimedLap, softBuyRead/softBuyHoursClause, the /scan digest soft-buy cell, read-schedule.mjs and read-window-range --profile; shadow-logged by suggestlog.mjs. DISPLAY-ONLY — it gates no grade, rank, verdict or screen.json field. It DOES move displayed prices on the ~0.8% of rows that PASS: DT4b refits those over the gate's window, and hours and levels come from one fit, so the soft-buy floor/diurnal levels shift there — see displayFitNights) + softBuyHoursClause (the ONE wording for the three hours states, full + compact styles) + windowStats/quantLow/quantHigh/touchedDays/reachedDays (EC2 2026-09-08: windowStats also returns forming — today's partial-day {key, low, hi, volLo, volHi}, still excluded from days as always but no longer silently discarded; floorCeilingTrack's monotone-certainty read consumes it) + the RC1 recencySplit/recentQuant reach-contamination guard + askExitRead (PLAN-POSITIONS-WINDOW-READ 2026-07-18 — the ONE ask-side "typical exit" assembly: daily-HIGH q50/q75/every-day levels + the scored list-price reach/placement + the ≥FIVE_MIN_MIN_DAYS 5m-grain reach; pure over already-computed windowStats results, so read-window-range.mjs's --ask block and quote-items.mjs --positions' auto-surfaced big-ticket ↗ windowExit note render from ONE definition instead of re-sequencing the primitives) + reachMargin (the fade check, 2026-07-20 — folded INTO askExitRead, so it rides both surfaces: the cushion TREND fading|stable|extending over the recent MARGIN_NIGHTS days at MARGIN_FADE_FRAC, the current-day cushion, and today's pace — live-now vs the reaching-day median at this hour-of-day off the in-hand hourProfile; symmetric ask/bid, inform-only, placeholders pending F1; the lean summary rides suggestions.jsonl via windowExitShadow) + avgBoundRead/formatAvgBound (2026-08-05 — the DEEP-BOOK reach-misread guard: touchedDays/reachedDays count days the per-day extremum of 1h-bucket AVERAGES crossed a level, a bias that is strict IN PROPORTION TO LIQUIDITY, so a low N/M on a deep book means "below every hourly average", not "never fills". Fires only when the limiting-side volDay clears REACH_RELIEF_MIN_VOL (reused from js/estimators/reach.mjs, never forked — passed in as an opt so windowread.mjs stays a leaf) AND the hit fraction is below AVG_BOUND_LOW_FRAC; names the averaged basis, the signed gap to the most extreme daily average (read against AC2's ~0.36–0.56% measured smoothing bias), the in-window competing pool, and the bid-side "low placement = deep patient entry" gloss. Rendered by read-window-range.mjs on every scored --bid/--ask/--exit AND mirrored into the --json/--out dump as avgBound (the machine path is the point — agents read verify.json). THIN books return null and print byte-identically, an asymmetry pinned by test in pipeline/test/windowread.test.mjs. INFORM-ONLY, n≈0 — gates nothing) + the hour-of-day diurnal profile hourProfile/deriveDiurnalRange (2026-07-09 — de-trended per-hour dip/peak detection, side-specific clustering, and the stale-to-live guard; the peak-timing engine screen-flip-niches.mjs auto-runs and windowrange --profile prints) + multi-peak windows (PLAN-MULTI-PEAK-WINDOWS 2026-07-23 — hourProfile gains ADDITIVE prominence-ranked peaks/dips arrays (length 1–2); peaks[0]/dips[0] deep-equal the unchanged peak/dip, peaks[1]/dips[1] is a SECOND window that cleared the SECOND_PROMINENCE_FRAC topographic-prominence gate; diurnalTimedLap mirrors it with index-aligned askReaches/bidReaches arrays — inform-only, n≈0, rendered as a trailing clause by emit.mjs formatTimedLap) + computeReality/realityClause (PLAN-DIURNAL-RECENCY-GUARD 2026-07-24 — each emitted peak/dip carries an additive reality level-check: spikeTop (a recent 1–2 day spike over-generalised into the quoted level) / staleOptimistic (an old high the current regime no longer reaches) + a recency-honest typicalLevel to quote instead, off the cluster's per-day HIGHS/LOWS with ZERO new fetch; realityClause is the ONE renderer across nine console call sites as of Chunk 2c (2026-08-13) — --profile window headers + its → BID/ASK line, the non---profile diurnal: summary, formatTimedLap's primary BID/ASK and its also ASK/also BID secondaries, quote-items' windowExit peak-level bit, the held-lot thesis-frame exit, and /schedule's Level column. (This cell read "the three console surfaces" through Chunks 2b AND 2c; every hand-enumeration of these sites has been short — treat the number as a floor, and see the plan's §10 for what is still bare.) Constants SPIKE_REACH_FRAC/SPIKE_PLACEMENT_PCTILE/ SPIKE_MIN_GAP_FRAC/REALITY_TYPICAL_QUANT/REALITY_TYPICAL_RECENTN are PLACEHOLDERS; inform-only, n≈0, never gates — the emitted level is unchanged. No longer console-only: Chunk 2c added the WRITE side, so reality now reaches suggestions.jsonl (timedLapShadow) and verify.json (result.profile, which already carried it) — that is what finally makes "does this flag predict anything" an answerable question) + hourConcentration/diurnalTimedLap (PLAN-DIURNAL-TIMING DT1 2026-07-23 — hourConcentration is a per-day argmin/argmax-hour CIRCULAR-concentration classifier (mean resultant length R∈[0,1] of each day's own trough/peak hour, HOURCONC_MIN_DAYS/HOURCONC_MIN_R placeholders), distinct from hourProfile's aggregate cluster width; diurnalTimedLap is deriveDiurnalRange's output EXTENDED with net/roi (the TIMED trough→peak lap, netMargin from js/money-math.js) + instantNet/instantRoi (the SAME-HOUR/churn margin — both surfaced, since a big-ticket item can show a NEGATIVE same-hour margin beside a POSITIVE timed one) + bidReach/ askReach (recencySplit scored against the chosen dip/peak levels' own windowStats slice) + lowTrend/hiTrend (projectTrajectory) + dipPool/peakPool + trancheComfort/ trancheCeiling (DT_TRANCHE_COMFORT_VOL_PCT=0.5%/DT_TRANCHE_CEILING_VOL_PCT=1% of volDay, borrowed from js/estimators/reach.mjs's n≈6 reach-relief knee, not validated for diurnal specifically — rendered tranche ~X clean · ~Y price-knee, and read the two disclaimers on emit.mjs's caveat before quoting either number: it is a price-degradation knee rather than a clearing cap, and a ROUND-TRIP bound rather than a per-leg one, both of which have been misread off it in a live session) + hourConcentration's clean verdict; degrades to {degraded:true, reason}, never a throw. DT2 (2026-07-23) wires this into screen-flip-niches.mjs for EVERY flip-niche survivor (was top-picks-only via raw hourProfile+deriveDiurnalRange), rendered through the ONE shared pipeline/lib/render/emit.mjs formatTimedLap — see that file's README entry. DT3 (2026-07-23) wires the SAME diurnalTimedLap+formatTimedLap pair into quote-items.mjs's bare-quote kind:'diurnal' note (prof/dr themselves stay — they still feed extraEst.diurnal, the window-clear peak window, and the forward E4 inputs), and swaps watch-positions.mjs's two direct hourProfile+ deriveDiurnalRange call sites (the shadow-log bid/ask co-log, the diurnalAsk cycle-fallback exit) for diurnalTimedLap — those two are VALUE consumers, not note-render sites, so only the underlying computation moved) + softBuyRead/formatSoftBuy/SOFT_BUY_CUE_TEXT (2026-07-22 — the ADD-while-holding soft-buy timing read off the SAME hourProfile: the dip-cluster FLOOR level + a live-vs-dip-floor @floor/+X% marker at SOFT_BUY_AT_FLOOR_PCT, ending in a cue, with the diurnal DIP window in a trailing parenthetical labelled attended. DT2 (2026-08-09) made the render LEVEL-FIRST and re-scoped the window: it does NOT time a resting offer (71.2% in-window touch vs 70.5% for a random window of the same width; waiting forfeits ~29% of bid fill-days at an identical price), so a resting bid is placed at the level now and the hours are for attended market-TAKING. The @floor cue is FLOOR-AWARE (the fang under-read fix) — softBuyRead takes an optional fc (a floorCeilingTrack result the caller already computed; NO re-derived slope) and resolves buy now (flat/ranging), ▲ favorable (rising floor — price-trend-only, never a green-light), or ▽ caution — floor breaking ↓ (broke/crash-risk — a dump artifact, not a discount); SOFT_BUY_CUE_TEXT maps the cue to its wording so both surfaces phrase it identically. quote-items.mjs renders it as the ⏳ softBuy note on held lots + bare quotes (threading the fc pushTrajectory returns), and screen-flip-niches.mjs's digest soft-buy column delegates to the SAME helper (ONE implementation); inform-only, n≈0, null profile / no fc ⇒ plain cue) + asymPair (PART II PLAN-GRADE-REACH 2026-07-12 — the day-level deep-bid/high-reach-ask realizable pair + P_ask/P_bid, consumed by js/estimators.mjs asymEstimate for the ◆ asym fill inform line + the asym suggestions-ledger shadow field) + clearableAsk (PLAN-DEPTH-EXIT DE1 2026-07-15 — the percentile-DEPTH exit: reconstructs a per-day price→volume distribution from the 1h bucket point masses and answers "what can I actually BOOK at?" for a given lot size; the reach count is its qty→0 limit, and a thin book collapses to a null-with-reason; feeds the --depth "BOOK AT ≤X" line + the LIVE DE3 depthExit shadow on watch/quote held lots; the DEPTH_* constants module-internal placeholders. depthDays/clearableBid were REMOVED 2026-07-22, PLAN-REMOVE-DEPTH-PRESSURE-READS — git-revivable) + demandPressure/reachableBand (PLAN-DEPTH-EXIT Extension A PB1 2026-07-15 — the pressure-driven reachable band: s=ln(medVolHi/medVolLo) sets each side's headroom base ± band·φ(±s)·reliability off the recent central daily level (RC1 reused)
  • the daily-high/low IQR; a thin-VOLUME book collapses to the smoothed center via the sample-reliability guard (no peak-cap); the PRESSURE_* constants are exported n≈0 placeholders and the Soul-rune/sell-heavy reasonableness pins live in the test; its surviving consumers are the co-log shadow (bid/band), the read-window-range.mjs --pressure inspector and reverse-flip's reverseListBand — the pressure SELL model it once priced was RETIRED 2026-08-30 (join-exit-ev.mjs's criterion; CHANGELOG 0.76.0). The Extension-B hourlyPressure/demandRegime per-hour demand-cycle classifier was REMOVED 2026-07-22, PLAN-REMOVE-DEPTH-PRESSURE-READS — git-revivable) + trajectoryRead (2026-07-21, the fang under-read fix — the shared multi-day SHAPE read over a windowStats().days series: classifies rising/falling/oscillating/based/elevated + the window floor/ceiling (with the day each printed) + where a liveRef sits between them; HEURISTIC/inform-only, never gates. Its rendered ⌁ read: line was RETIRED at both emitters by R6 — floorCeilingTrack below supersedes it; what survives is the floor/ceiling/livePos fields that note now carries) + floorCeilingTrack/formatFloorCeiling (PLAN-DRIFT-VS-CRASH, 2026-07-22 — the phase-aligned floor+ceiling slope-asymmetry classifier that trajectoryRead's single min-low/max-high collapse washes out: reads the daily-LOW track and daily-HIGH track SEPARATELY, each a recent-window least-squares slope (windowed, so a short trailing wiggle cannot flip it; a volatile END day can — OLS endpoint leverage) classified rising|flat|falling + a raw-sign trailing micro-run for duration, plus a discrete floor-break flag (latest completed low vs the prior-lookback floor); combines the two slopes + the break into crash-risk (break dominates) / healthy-trend / compressing-up / mild-cooldown / cooling / ranging. REQUIREMENT #1 phase-alignment: the forming/incomplete day is DROPPED from the slope/break and surfaced separately as provisional. EC2 (PLAN-ENTRY-CONFIDENCE, 2026-09-08 — the Avernic-hilt entry post-mortem): the live forming read arrives via the forming opt (= windowStats().forming, the partial-day aggregates windowStats always excluded from days and now RETURNS instead of discarding — before this, every production days was pre-stripped, so fc.forming was structurally null and the "provisional" clause never rendered anywhere); todayKey survives as the hand-built-days fallback. On it rides formingContradiction, the MONOTONE-CERTAINTY read: a forming day's low only falls, so forming.low < a completed-days statistic is decided at any hour — under-trough (forming low < min(prior floor, latest): the discrete break is CERTAIN at day end) or under-last-low (the floor label is STALE — measured over 21,013 archive item-days at 21:00, the next 3 days print below the current floor 88.5% when it fires vs 53.1% when it doesn't; a "cheaper entry likely" read, NOT a trend-reversal predictor — forward slope is unpredicted either way, and 44.1% of rising labels carried the contradiction). formatFloorCeiling stale-marks a rising-flavored label on under-last-low and WITHHOLDS the label on under-trough; softBuyFloorCue maps under-trough → caution and rising+under-last-low → the new stale-uptrend cue. The regime GATE (quotecore.js regimeDrift), driftExitFrom's slope read and join-exit-ev.mjs's replay deliberately pass no forming — gates and replays stay completed-days-only. HEURISTIC/n≈0/inform-only, never gates; formatFloorCeiling (fmt injected — windowread stays dependency-free) is the ONE line-render both read-window-range.mjs's trajectory block and quote-items.mjs's ⇅ note use); MOVED here from pipeline/lib/ so it is node- AND app-importable like quotecore.js; consumed by pipeline/commands/read-window-range.mjs, pipeline/commands/watch-positions.mjs, pipeline/commands/screen-flip-niches.mjs, js/validate.mjs and js/forecast.mjs (both now app-imported via js/trends.js, TV). PF1 (2026-07-10) added additive per-hour dispersion fields devMid/devLowSpread/devHiSpread (IQR of the deviation samples) so the forecast band isn't re-derived; every pre-existing field is byte-identical), forecast.mjs (PF1 2026-07-10 — the pure forward 12h/24h price projection: CONSUMES an hourProfile object and produces a diurnal+trend forecast — diurnalForecast(profile, ctx) → nextTrough/nextPeak {level, band, etaH, window, confidence, mode} + the per-hour projected series, plus whenBuyable/ whenSellable/fmtEta. The interpretable ADDITIVE model projLevel(h) ≈ baselineNow + trendPerHour·Δt + deTrendedHourShape(h); anchor from the live quote, shape/dispersion from up to 14d, trend from the recent slope. DEGRADES LOUDLY to {forecast:null, reason} on a spike/decay phase, a live band violation, a thin/short series, an unreliable quote, or a trend-erased dip (trend-only mode); the band widens with horizon. Claims ONLY "recurring diurnal shape + dumb trend extension" — never an exogenous shock. Imports only windowread.mjs (no quotecore — phase/mom/reliable arrive as plain ctx). INFORM-ONLY / console-only / provisional (n≈0, every constant a NAMED PLACEHOLDER pending the PF8 backtest); no consumer wired yet (PF2–PF8) and no app import → no APP_VERSION. Pinned by pipeline/test/forecast.test.mjs. Also homes PLAN-OSCILLATION-CYCLE Chunk 1 (the multi-week oscillator lane): driftAdjustedExit(fc, {ceilingSlope,floorSlope,holdHorizonDays}) composes diurnalForecast's next trough/peak with a multi-week drift NUMBER (never a direction label — NO phase/direction field), shifting ONLY by the RESIDUAL horizon past the diurnal eta (diurnalForecast already trend-extrapolates to the eta), and oscillationVsKnife(days) — a detector (REDESIGNED at F-A, 2026-07-22 — the original first-difference flip-fraction metric measured day-to-day NOISINESS, not harvestable oscillation, and mislabeled fang/blowpipe's smooth multi-day runs a false knife; see the header comment above the function for the full finding) that detrends the daily mids (same shared projectTrajectory slope, one-home) and splits the residuals into maximal same-direction LEGS, counting a leg as real only past OSC_MIN_LEG_DAYS + OSC_AMP_NOISE_MULT× the series' own day-to-day noise floor; oscillating fires at OSC_MIN_LEGS (≥2 direction reversals) — fewer legs is a monotone linear-fit hump (even a CURVED collapse), never a real cycle. Tells an oscillating-while-drifting shape (fang/blowpipe) from a monotone knife where floorCeilingTrack.oscillating structurally can't. F-H (2026-07-22) added OSC_DETECTOR_NIGHTS=21 (> the amplitude gate's AMP_NIGHTS=14): renderAmplitudeMode feeds the detector its OWN longer trailing windowStats(...).days window (off the same in-hand series, NO fetch) so it sees the ≥1.5 cycles / ≥3 legs it needs WITHOUT widening the gate's AMP_NIGHTS daily-range/ reach/recency read — a sample-size fix BOUNDED by the ~16-day /timeseries?timestep=1h endpoint, not a calibration. ⚠ That endpoint bound is now load-bearing, not incidental (2026-08-11): OSC_MIN_LEGS is an ABSOLUTE leg count with no length normalisation, so the label is a function of WINDOW LENGTH — measured 59.5% OSC at 14d → 99.9% at 60d on the real archive, and ~66% at 14d → ~100% at 30d on a synthetic DRIFTLESS RANDOM WALK containing no cycle at all. Amplitude cannot matter — every threshold is homogeneous of degree 1 in price, so the criterion is scale-free BY CONSTRUCTION and length is the only free variable (an earlier note offered amplitude-invariance as corroborating measurement; it is an algebraic identity, not evidence). The ~15d cap is the only reason the knife temper still discriminates; feeding the detector a deeper archive.mjs series (F-H calls this a noted-not-built follow-up, and renderAmplitudeMode already has that archive open) would take it to ~100% and silently delete the guard. Normalise the criterion BEFORE widening the window — full note in the function header; measured in RANGE-PERSISTENCE-FINDINGS.md. INFORM-ONLY, wired into NO gate in Chunk 1 (gating is Chunk 3); pinned by pipeline/test/oscillation-cycle.test.mjs. Chunk 2 adds driftExitFrom(profile, days, ctx, opts) — the ONE slope-sourcing + drift-adjusted-exit COMPOSITION (imports floorCeilingTrack from windowread to pull the ceiling/floor slope off an in-hand windowStats().days, NO fetch; builds the diurnalForecast wrapper; calls driftAdjustedExit) — the reusable caller pattern the amplitude lane established and Chunk 6 REUSES; PURE/tax-free (the after-tax margin stays the caller's concern). Pinned by pipeline/test/oscillation-shadow.test.mjs. Chunk 3 (THE ONLY GATE) turns that margin into amplitudeGate's margin-below-floor reject (amplitudeDriftMargin().margin <= 0, direction-agnostic, the floor already inside the margin) sequenced after trend/knife, computed ONCE at the gate stage in renderAmplitudeMode and reused for the shadow-log, and TEMPERS the knife guard with oscillationVsKnife (a drift-riding oscillator is not a false knife → falls through to the margin gate). Pinned by pipeline/test/oscillation-gate.test.mjs. Chunk 6 REUSES driftExitFrom per-thesis (band/churn/scalp/value drift-adjusted-exit INFORM notes — see js/flip-niches.mjs; console-only, no gate). Chunk 5 folds the drift-adjusted exit LEVEL into the SHARED formatFloorCeiling note path (an optional drift opt — the caller passes a pre-computed driftAdjustedExit() result off its in-hand prof+days, so windowread keeps its one-way arrow) so it rides beside EVERY price suggestion (quote-items.mjs trajectory note, read-window-range.mjs, and — APP-VISIBLE — js/trends.js renderForecast); a projected LEVEL never a direction verdict, display-only, degrade-clean; APP_VERSION-bumped (reaches js/trends.js). Pinned by pipeline/test/oscillation-render.test.mjs). F-F adds a trough-vs-decay DISPLAY annotation to the amplitude reach cell (reachPhaseNote in screen-flip-niches.mjs — recent+full both-leg reach plus a 3-signal phase note off oscillating/floorSlope/margin, direction-agnostic in the knife bucket, console-only; pinned by pipeline/test/oscillation-reachphase.test.mjs), validate.mjs (P2 — the pure VALIDATOR REGISTRY (ctx)→{status:pass|caution|reject,reason,evidence} run on EVERY surface: reachValidator wraps windowread reach + RC1 into caution/reject WITH the reach evidence; floorValidator (P3, BUY-side) rejects/cautions a buy parked above the durable floor — and durableFloorRead(vres) (2026-08-06) is the ONE canonical extraction of that verdict ({status, ranges, lookback}) for consumers that need the LEVEL read without re-deriving it: the soft-buy unproven-base cue, the ⚠N×floor row probe, and the bucketed caution footer all compose it rather than fork it (windowread cannot import validate — the arrow is one-way — so the caller hands it down); trajectoryValidator (TV1 2026-07-09, BUY-side) is the SHAPE policy over termstructure's classifyTrajectory — knife/oscillating/based/elevated; valueAmplitudeValidator (TV1) the recent-week amplitude+proximity for value; limitValidator (LM1) the rolling-4h buy-limit; dipPostureValidator (DP1, BUY-side, INFORM-only/NEVER-REJECT) the dip DIRECTION read via quotecore.js's recentDirection (a reverting dip → caution "cross or pass"; wired inform on band/churn). All degrade to pass on missing data and never throw. runValidators(ctx,{specs}) drives a PER-THESIS plan ({key,mode,window} from js/flip-niches.mjs) — gate (verdict stands) vs inform (informFlags: annotate-only, clamped to pass, would-have verdict logged via leanValidators). worstStatus/flags. Screens DROP reject + FLAG caution + SHOW inform notes; explicit asks/held/watchlist never hidden. App-imported via js/trends.js (TV) — a behavioral change here needs the smoke test + the APP_VERSION rule), termstructure.mjs (P3 — pure DOM-free multi-day term structure over a daily-mid [{ts,mid}] series: the 1/3/7/14/28d termStructure (median/low/high/pctInRange per lookback), a durable floor (low quantile of the longest multi-week lookback), a robust ceiling (P5 — the symmetric high quantile q85, so a lone spike can't inflate a range), a typical fluctuation (IQR), and a trajectory SHAPE (TV1 — classifyTrajectory: knife/oscillating/based/rising/elevated/flat, attached as ts.trajectory); degrades to hasData:false/unknown on a short series. Plus (DT6, PLAN-DIURNAL-TIMING §6, 2026-07-23) basePosition(ts) — a pure, LIGHT read of an already-computed ts: reuses ts.lookbacks[14].pctInRange (the SAME field classifyTrajectory already scores) + a 3-way coarsening of ts.trajectory.shape (+ts.recentTrend.dir to split a falling-drift oscillation as "decaying") onto range-bound/ trending↑/trending↓/decaying, for the screen-flip-niches.mjs Base position note on band/ churn/amplitude survivors (rendered by <