Relationship analytics

all-time records within the loaded span · colored = which side holds it · ● = set in the last 30 days · click a tile → that day

the most recent occurrence of each signal · colored = who it came from · ● = still lit (within the last 30 days) · click a tile → that moment

cuts = sustained changepoints on RAW weeklies - volume (grey), warmth (orange), msg-share balance (purple), conversation friction (red, from the model's summaries) · pace/warmth cut dates are refined to the best DAY near the pivot week; balance/friction resolve to the week · ≥4wk eras, ≥20-msg weeks · the smoothing dropdowns move the line, never the cuts

columns = the changepoint eras chapters detects (RAW weeklies - the smoothing dropdowns don't move them) · color = how the era ranks against the couple's other eras (green better) · balance rows (kai initiates / labor / expressed emotion) red the further from 50/50, whoever leans · warm closes / vulnerability / friction / labor read the model's conversation summaries and blank when an era has <5 of them · era grade = the column's average standing on this couple's own curve · reply/positivity definitions as on their own charts

composite = weighted mean of the component rows below, rescaled 0–100 · red <40 · amber 40–60 · green >60 · rows red→green = 0→1, heaviest weight on top (hover for weights) · drag/lift = the component that moved the composite most over the last 28d, in composite points · ▼ = a rupture conversation, deep red = unrepaired (hover for the model's summary, click → Discord)

composite = weighted mean of the 13 balance signals below, −1…+1 · 0 = even · beyond ±0.25 = entrenched imbalance · above 0 = kai more invested · rows lean sam ← even → kai in their colors (who CARRIES the signal, not good/bad), heaviest weight on top (hover for weights) · shift → = the signal that moved the composite most toward that side over the last 28d · ▼ = a rupture conversation (hover for the model's summary, click → Discord)

momentum = fast pace − slow baseline · fast window = ¼ of the smoothing window (≥3d) · a pace flip only counts once momentum escapes the grey noise band (±10% of baseline) - in-band wiggles keep the regime · warmth lane = the same fast−slow read on per-message net sentiment, so "louder" and "warmer" are separate claims · ▼ = a rupture conversation (hover for the model's summary, click → Discord)

positivity ratio = positive ÷ negative emotion mass (mass = mean score × messages), capped 10× · green ≥3.5 healthy (Gottman), yellow 1–3.5, red <1 · peak skips sparse stretches where the ratio only pegs the cap · tracked = top emotions by count-weighted score · ▼ = a rupture conversation (hover for the model's summary, click → Discord)

streak = consecutive calendar days with (talk lanes) / without (silence lanes) a message · color depth ∝ length within its lane · outlined bar + → = still running · click a bar → its boundary message (talk: the streak's first; silence: the one that broke it) · ★ = a new longest-daily-contact record (hover for what it beat)

negative = below this chat's own 10th-pct valence (-0.72) · flashpoint = ≥9 negative within 30 scored messages, both sides, ≥12 total · ✕ = chained into a 20h+ silence

stacked share of conversations on the model's conflict-severity ladder, 28-day average - filled = friction, the gap to 100% = calm · tension = friction that held · rupture = a real break, repaired (it mended in-conversation) or not · ▼ = one rupture conversation (hover for the model's summary, click → Discord) · faded = few conversations in window · dotted verticals = chapter cuts

conflict = the model read tension/criticism in the conversation (not just negative words) · green = repaired (acknowledged or smoothed over before the end) · red = left hanging · line = share of the last 5 conflicts that got repaired

reopening = the first message of the conversation after one the model read as tension/rupture · left: share of reopenings made, count-weighted rolling · right: one dot per reopening (log wait), green = the next conversation read calm, gold = friction carried over, bar = median · hover for the friction's summary, click → the reopening message · faded = few events in window · dotted verticals = chapter cuts

rupture = no messages for ≥20h · marker = whoever spoke first after · hollow = no reply within 8h · recovery = 7-day pace back at 70% of the pre-rupture month

each summarized conversation bucketed by the silence since the previous one's end · warm = warm_close ending or a bonding connection move · friction = a tension or rupture read (the two can co-occur) · faded = thin buckets (<10 reunions) · outside labels = the rate, n under the axis

cell = how much more (red) or less (blue) likely than its base rate that tier is for the NEXT conversation, given this one's read (lift, log-colored) · outlined diagonal = the weather persisting · · = never observed where chance predicted under 3 anyway · rows need ≥5 occurrences · consecutive summarized conversations only

each line = share of conversations the model read as carrying that mode, 28-day average (modes overlap - not a stacked share) · title arrows = lifetime → this era share · this era = since Jun 04 2026 (58d, “intimacy, substance, and self”) · rupture = a real break in the bond · tension = friction that held · support = comfort sought/given · closeness = pulled them closer · faded = few conversations in window · dotted verticals = chapter cuts

each line = share of summarized conversations the model read that way, 28-day average (independent rates, not a stacked share) · resolved = thread settled · warm = warm last stretch · deepened = went deeper than it opened · left hanging (dashed, the counter-signal) = unfinished business (open loops) · faded = few conversations in window · dotted verticals = chapter cuts · this era = since Jun 04 2026 (58d, “intimacy, substance, and self”)

warmth = smoothed conversation sentiment rescaled to 0-100 · resolved = count-weighted resolved rate · 28-day average · the shaded band is the gap: warm-but-unresolved when it widens · faded = few conversations in window

share of conversations the model read as each person carrying the emotional labor - repair, soothing, checking-in, 28-day average (the tooltip's balanced/unclear = the model named no side) · band = the leader's color · lines crossing = a handoff, ◆ = a confirmed one (the new leader opened a ≥10pp lead) · note: the call partly tracks who wrote more · faded = few conversations in window · dotted verticals = chapter cuts

share of conversations the model read as each person leaning on a victim stance - aggrieved, blaming, hard-done-by, 28-day average (the tooltip's balanced/unclear = the model named no side) · band = the leader's color · lines crossing = the victim role changing hands, ◆ = a confirmed one (the new leader opened a ≥10pp lead) · note: the call partly tracks who wrote more · faded = few conversations in window · dotted verticals = chapter cuts

pursuit = double-texts (re-messaging before a reply) · withdrawal = dead-air hours (leaving the other waiting ≥3h) · each normalized to its cross-user p95 · above 0 = chasing, below = pulling away · lines cross when the roles flip · smoothed 28-day average

each dot = a checkpoint sampled every 28d · size ∝ msgs/day · color faint→solid = early→recent · green corner = balanced & healthy · grey shading = where it lived (density of days) · ◇ = a chapter cut (hover: what shifted)

strip = one conversation, midnight wraps to the next day · led = ≥60% of its messages · labeled marathons by message count

one tile per conversation, reading like text (oldest top-left) · pale = calm · gold = tension · orange = rupture repaired · red = unrepaired · □ = a chapter opens · row label = that row's first conversation · hover for the model's summary, click → Discord

opener = first message after >=2h quiet · ignored = one ping, no reply · monologue = kept talking, still no reply · converted = both spoke · tone = last message's net sentiment (warm >=+0.3 / cold <=-0.3) · click a ribbon → most recent example

each converted conversation (both spoke) flows open -> time-of-day -> length -> end -> after · after = a >=20h silence followed · length short <10 / medium <30 / marathon 30+ · category counts in parentheses · ribbon colour = how it ended (green warm / grey neutral / red cold)

opener = first message after ≥2h of quiet · converted = the other side joined that run (≥2 msgs, both spoke) · unanswered pings and monologues count as fizzles · blank stretches = too few openers to rate

trail = chapter eras in order, dot size ∝ openers in the era (≥5); faint dots = single months · diamond = lifetime · converted = the other side joined (≥2 msgs, both spoke)

fizzle = a run-opener the other side never joined (>=2 msgs, both spoke = converted) · daily count smoothed · grey = multi-day silences

bars cover every month (explicit zeros) · the duration line only exists where conversations do

fan = monthly p25–p75 band + median · monthly p95 spikes toggle on via the legend · session = one gap-split conversation · dashed refs = lifetime medians

conversation = messages <2h apart, both sides spoke · burn uses chats with ≥8 msgs · dotted diagonal = perfectly even burn · t½ = half the messages sent

marathon = 100+ messages or 2h+ · top: the 20 biggest conversations - dot size = duration, color = who dominated · hover for the model's summary, click → that conversation in Discord

chain depth = links in one consecutive quoted-reply thread (4+ bucketed) · bars = monthly replies by depth, black = deepest chain that month · line = daily-smoothed mean depth

moment = two messages ≤10s apart · rate per 1,000 msgs discounts volume · ref = mean of the drawn daily-rate series

bid = a question, share, affection phrase, or conversation reach-out · toward = reply within 2h that isn't negative · ignored = no reply in 2h · against = reply whose dominant emotion is negative · line colored by the BIDDER

alternation = share of adjacent messages (within 1h) where the speaker changes · dotted 50% = even ping-pong dialogue · low = block monologues · run = consecutive messages before the other replies (or an hour passes) · days with <5 adjacent messages excluded

ender = last word before ≥2h of quiet · solid = ending, dashed = starting · green 40–60% balanced · red ≤25% / ≥75% pursuer–distancer · % lines faded/blank = few conversations in window

silence = one side quiet ≥3h, merged until the silent side speaks · colored by who went quiet · ▼ = a rupture conversation (hover for the model's summary, click → Discord) · pre-spans caches fall back to raw per-gap events

a long silence spreads across the days it spans (~24h each, no spike) · monthly bars = per-calendar-day mean (zeros included) - same scale as the daily line · green ≤3h · red ≥8h

gap = kai's − sam's daily median reply time · fill colored by the slower side · green |gap| ≤10m · red ≥30m entrenched

median time to answer a partner message INSIDE a conversation (within the 2h gap), by what that message was · ×1 = that side's own all-reply median · red ≥1.15× slower, green ≤0.87× faster, grey ≈ baseline · mood = net GoEmotions valence of the message answered - flat mood rows mean latency is processing cost, not feeling · levers need ≥50 replies

reply times capped at the 2h conversation gap · lifetime stats = median of the raw daily percentiles · shifts = this era vs lifetime · green ≤5m · red ≥60m · this era = since Jun 04 2026 (58d, “intimacy, substance, and self”)

each dot = a one-sided silence and how long until that person broke it · line + band = trailing-90d median and p25-p75 (n≥8 comebacks) - WIDE band = erratic/unpredictable · solid rule = lifetime median · ○ = ongoing silence, not yet broken · log scale

reply = adjacent cross-author message within the 2h conversation gap · each length column normalised to 100% (conditional reply-time mix) · columns with <10 replies hidden · x tick counts = total pairs (both repliers) · marker = column median

each ridge = that month's reply-latency distribution (both people, capped at the 2h gap, log x) · ink overlay = session pickups (first reply of a conversation session) - the conversation-restart share of the tail · dot = month median · height normalised per month · months with <15 replies stay as tagged empty rows · newest on top

per month: kai latency density up, sam down (mirrored, log x, capped at the 2h gap) · dotted lines join each side's monthly median, shaded band = the gap between them · months where either side has <10 replies stay as tagged empty rows · newest on top

counted only when the opener got an answer (the run became a conversation) · latency = opening message → the other side's first · dots = individual answered calls (sub-15s drawn at the 15s line) · response_time measures in-conversation replies; this is arrival to a fresh thread

panels 2–3 = share of messages / words by kai (28-day average) around the 50% line · green 40–60% balanced, red one-sided · dashed = lifetime share (raw Σ) · peaks skip near-silent stretches (<25% of typical volume) · ▼ = a rupture conversation (hover for the model's summary, click → Discord) · faded = little volume in window

7 phrases (≥20 uses) compared · 346 total uses

cadence = mean hours between a side's OWN consecutive posts, uncapped - each gap lands on the day they RETURN · days with no posts carry no value (not zero) · log scale

LSM = 1 − |a−b|/(a+b) per function-word class, averaged over 9 classes, from per-user per-day rates · <60 distress · >80 aligned · dots = raw scored days (both sides ≥20 words) · ▼ = a rupture conversation (hover for the model's summary, click → Discord) · class panel colors drift ≥3pts vs lifetime, needs ≥8 scored days (this era) · this era = since Jun 04 2026 (58d, “intimacy, substance, and self”)

all panels 28-day average · we/I/you = share of that side's personal pronouns · blame index = "you"/"I" ratio · green/red bands = calibrated healthy/distress ranges

typed register from raw message text · lowercase = first letter after stripping :emotes:, links & leading markdown · rates per 100 of that user's own messages in the window · faded/blank = that stretch holds too few messages · title compares lifetime vs this era · this era = since Jun 04 2026 (58d, “intimacy, substance, and self”)

each month = sam's word distribution that month, TF-IDF cosine (words he uses EVERY month don't count as identity) · leaf color = the month's chapter · dot size = his words that month · leaf hover = the month's distinctive words · months with under 300 of his words omitted

words both sides use a lot, ranked by combined use, common conversational filler dropped · click a bar → that side's first use of the word

shared word = ≥3 lifetime uses by EACH side, stamped when the second side adopts it · adopter = the side whose first use came SECOND (completing the share) · faded bars = warm-up (first 90d of the loaded span or the first ~3 months of real two-sided text - mostly older vocabulary resurfacing) · overlap months need ≥200 content words from both sides - washed-out months had some text but too little to rate (dotted bridge)

each panel = one month's newly shared words (≥3 lifetime uses by EACH side), top 6 by total use · bars = lifetime uses per side · the shaded rollup pools the warm-up (first 90d or first ~3 months of real two-sided text - mostly older vocabulary resurfacing) · months minting nothing are skipped (“after N quiet mo”) · a quoted title = the chapter that month opened (hover it for the summary) · click a bar → that side's first use (the later side's first use is what made the word shared)

obituary = ≥25 uses across ≥90d & ≥3 months, then 56d silent (2× the smoothing window) · each bar spans the word's life, first→last use; dotted verticals = chapter cuts, ▾ = died within 14d of one · nursery = first seen in the last 150d, after a 90d burn-in · bar color = heavier user, grey = shared

kai: 14k words · sam: 5.1k · 60 words both use 10+ times

share of each month's emotional mass (sum of GoEmotions score x messages, both people) · neutral excluded · the 27 emotions folded into seven families · fog = thin months (vivid from 200 msgs) · ▼ = a rupture conversation, deep red = unrepaired (hover for the model's summary, click → Discord)

each side vs its own baseline across 1 chats - sam vs their median elsewhere · kai vs the median partner · lifetime count-weighted means, bars clipped at 8× · faded bars = corpus-median mass under 0.5% (ratios built from trace amounts, skipped in the title)

conversations clustered by their emotion profile relative to the couple's baseline (what stands out, not raw levels), auto-labelled by signature emotions · top: monthly volume · middle: each mode's share at daily grain, 28-day average (raw daily is too sparse to read) · spans the full history via per-conversation emotion · bottom: among the 25 resolved vs 80 unresolved summarized conversations, which emotions lean each way · this era = since Jun 04 2026 (58d, “intimacy, substance, and self”)

net warmth = (Σpos − Σneg) GoEmotions mass over the day's messages, per side · gap panel = sam − kai, colored by the warmer side · dots = raw daily gaps · lines gap where a side sent nothing, fog = under 25 msgs/window · warmer-share counts only days both sent messages (n=87)

adjacent replies within 2h, dominant GoEmotions label folded to 8 buckets · text = % of answers · color = vs that side's usual answer mix · diagonal = mirroring · y counts = stimuli answered, rows with <20 blank · outlined cell = sharpest coupling (n≥10) · click a cell → a representative reply

top: how much more of each emotion kai (red) or sam (blue) brings than their overall share of the emotional expression (pp = percentage points) · ◆ = this era only (since Jun 4, 58 conversations) · mass = summed GoEmotions score behind a split - small mass reads thin · bottom: each side's share of the expressed (non-neutral) emotion over time, 28-day average - the lines crossing = a handoff of the emotional work · faded = few conversations in window · ▼ = a rupture conversation (hover for the model's summary + labor read, click → Discord)

themes auto-extracted per conversation; sized by conversations they recur in · single-mention themes trimmed from the cloud

recurring themes (≥3 conversations) · bubble size = conversations · valence = mean session sentiment · y = share of those conversations each side opened · ring = still live (raised ≤30d before the latest summarized conversation)

recurring themes (≥5 conversations) · x = median of (gap to next contact ÷ the couple's median gap within ±15 conversations of that moment) - immune to which era the topic lived in · red = slower than their pace then (withdrawal), green = quicker · band = q25–q75 · dot colour = next conversation's sentiment · size = conversations

each star = one conversation, placed by topic similarity (t-SNE) · near = talked about similar things · labels = each region's signature theme · colour = when it happened · size = conversation length · red ring = rupture (repaired or not) · click a star → that conversation in Discord

third parties the two talk ABOUT, from the model's per-conversation people read · counted once per conversation, names case-folded · the two speakers and generic non-names (someone, friends) dropped · quiet Nd = no mention in the last N days up to the latest summarized conversation (shown past 30d) · click a timeline dot → that conversation in Discord

each bar = share of the conversations naming that person that carried friction · red = a real rupture, gold = tension that held · dashed = the couple's baseline over all summarized conversations · names recurring in >=3 conversations · faded/italic = not mentioned in the last 6 months of the data · small n: read as a flag, not a verdict · click a bar → that person's latest friction conversation in Discord

each conversation flows from its dominant theme -> whether the model read the thread as resolved -> how its last stretch felt (warm / neutral / cold) · themes under 5 conversations pool into 'other themes'

each conversation placed in its month and assigned its dominant theme (band = that month's count) · 'other' = the long-tail of themes that don't fold into the named buckets · spans only the summarized window, widening as the topic cache backfills · the newest month is month-to-date · ▼ = a rupture conversation (hover for the model's summary, click → Discord)

each row = one theme's lifespan (first to last conversation), dot per conversation, coloured by coarse theme · ordered by when first raised · × = went quiet >30d before the dashed frontier · click a dot → that conversation in Discord

cell = how much more (red) or less (blue) an emotion shows up in conversations about a theme vs the couple's baseline · rows and columns clustered so similar profiles sit together · per-message GoEmotion scores summed over each conversation · 'neutral' omitted

each emotion is an anchor on the rim; a theme is pulled toward the emotions it expresses most above the couple's baseline, with pull scaled by how far its top emotion runs over it · near the centre = weak or mixed emotional signature · colour = positive (green) / negative (red) dominant feeling · size = conversations

cell = avg msgs for that hour × weekday across the whole span · per-user panels share one color scale; combined has its own

each column = how that month's messages spread over the day (sums to 100%) · months with <20 msgs or <7 days of data blanked - hour shares of a sliver of a month are noise · both panels share one color scale

smoothing runs in hours here - a 7d window = 168 hourly points · peak callouts = raw unsmoothed days · lifetime rates count every elapsed hour, silent ones included · this era = since Jun 04 2026 (58d, “intimacy, substance, and self”)

petal = share of that person's messages in each hour (midnight top, clockwise) · shaded wedge = typical offline window (each day's longest 4-16h silence) · up triangle wakes / down triangle sleeps · rim band = hours both typically online · dotted ring = weekend-only share

each active hour profiled by share of all messages, words per message, who sends more, and its emotional palette (GoEmotions score per message, neutral dropped), then clustered · branch color = day part (tree cut at 0.7 of the deepest merge), named by the traits that set it apart · dot size = the hour's share of all messages · dot color = who leads it · hours with under 20 messages omitted

cell = days that month when both sent a message in the same clock-hour · blank = never · baseline per cell = kai-active days × sam-active days ÷ days in month, so the ×-lift measures schedule alignment beyond chance

balance = min/max of the two sides' daily counts · most-mutual = highest min(a,b), not highest ratio · median over days both spoke

active on 41% of days (120/290) · color capped at the 95th-percentile day (151 msgs) · current streak 5d active

messages per hour-of-day × date, gaussian-smoothed (≈1.5h × 3d) · color capped at the 97th-percentile smoothed hour (6 msgs/h)

line = circular mean send time, unwrapped across midnight · masked where messages spread too thin around the clock (resultant < 0.25) · gap = circular distance

active hour = ≥1 message · together = both sent in the same clock hour · talk-hour share = hours both spoke ÷ hours either spoke · lifetime avg counts silent days · ▼ = a rupture conversation (hover for the model's summary, click → Discord)

tempo = messages (top) / words (bottom) per minute within a session - in-session pace on session days only, not volume · gray bars = monthly median, that month's peak in the hover · dashed refs = lifetime medians · ▼ = a rupture conversation (hover for the model's summary, click → Discord) · faded = few sessions in window · this era = since Jun 04 2026 (58d, “intimacy, substance, and self”)

each row = one calendar year, columns Jan–Dec · color capped at the 95th-percentile day (151 msgs) · right ledger = year total over its days of data

band = the gap, shaded in the longer writer's color · dashed = lifetime average (raw Σwords/Σmsgs) · faded = few messages in window · peaks = each side's record window, gated on the fog's evidence bar · dotted verticals = chapter cuts

monthly bars = per-calendar-day mean (zeros included) - same scale as the daily line · rate = reactions given per 100 of the partner's messages in the window · faded/blank = partner barely messaging

emotes classified by name: pat/hug/heart→affection · kek/lol→laughter · nod→agreement · sob/😭→distress (sob-laugh compounds like kekwsob count as distress) · top: share of each month's reactions, both people, fog = thin months (<30) · ▼ = a rupture conversation (hover for the model's summary, click → Discord) · bottom: share of that user's own reactions, ● colored when moved ≥2pp vs lifetime over the this era · this era = since Jun 04 2026 (58d, “intimacy, substance, and self”)

monthly bars = per-calendar-day mean (zeros included) - same scale as the daily line · links solid, attachments faint, stacked per user

edit delay = message sent → last edit, capped at 2h upstream, log scale: <30s typo fix · 30s–10m rethink · >10m revisit · fog = few edits in the window (daily medians there rest on 2-3 edits) · rate = edited share of that day's messages