AI Training — suggestions held for review

All 21 events over the 7-day lookback window share an identical shape: intent='resolved', signal_kind='no-action-needed', walker_muted=true, detail=null, reason=null. The pattern holds uniformly across all contact types (unclassified, lead, member, business), with no conflicting signals where the operator chose a different action. The high volume and consistency strongly suggest this is a deliberate operator policy — contacts surfaced by the walker are being systematically dismissed without action — though the absence of any reason/detail field means the underlying cause of dismissal cannot be further differentiated.

All 22 events share an identical structure: intent=resolved, signal_kind=no-action-needed, walker_muted=true, detail=null, reason=null — across varied contact types, indicating this is a deliberate bulk-resolution pattern rather than contact-type-specific behaviour. The high volume (22 events across two days, with a cluster of 13 in under an hour on Jun 29) suggests a systematic review-and-dismiss workflow. The only counter-evidence is the absence of operator-supplied reasons, making it unclear whether this pattern is intentional policy or a UI default that skips reason capture.

All 50 events share identical structure: stale_days=30 and expires_in_days=90 across three distinct walker_kinds, fired consistently at ~05:15 each day — strongly indicating a scheduled policy rule. The pattern is walker_kind-agnostic (applies to post_apt_outcome, stalled_followup, and member_at_risk_attendance_drop equally), suggesting the trigger is purely the staleness threshold. Minor counter-evidence: one event has orq_row_count=5 and one has orq_row_count=2, but the auto-mute action is still applied identically, so count variation does not alter the consequent.

All 50 events share identical structure: stale_days=30 and expires_in_days=90 across three distinct walker_kinds, firing from the same source at the same daily scheduled time — this is a consistent, repeated automated policy decision, not a one-off. The pattern generalises across walker_kinds, suggesting the rule is walker-kind-agnostic. Minor counter-evidence: one event has orq_row_count=5 and one has orq_row_count=2, but the mute action was still applied identically, so row count magnitude does not appear to change the outcome.

All 3 events show identical pattern: operator records lead outcomes by scheduling specific appointments with datetime and type. The appointment types vary (intro, in_person, phone_call) but the core action is consistent - converting lead interactions into scheduled follow-ups. Strong evidence with no conflicting behaviors observed.

All 22 events show identical signal_kind='no-action-needed' paired with intent='resolved' and walker_muted=true, with zero exceptions across varying contact_types and timestamps spanning two days. This is a highly consistent, repeated shape (22/22 agreement) making it a strong candidate for a reusable auto-resolution rule, though contact_type diversity suggests the trigger is independent of contact classification, which should be confirmed by the operator before automating broadly.

All 22 events show an identical shape: signal_kind='no-action-needed' consistently pairs with intent='resolved' and walker_muted=true, across every contact_type variant, with zero exceptions or conflicting outcomes. This is a strong, uniform pattern with high volume in a short window, though the lack of variation (no counter-examples of unresolved 'no-action-needed' signals) means we can't rule out this is simply how the system always behaves rather than a discretionary operator choice — worth confirming this is intentional design vs. an artifact of the sampled window.

All 8 events share an identical shape: AI-sent WhatsApp replies triggered by chat_entry, all on leads with active status, no DND, no terminal state, and no resolved outcome. The pattern is highly consistent with zero conflicting events. Confidence is capped below 0.9 because all events appear to be from a single contact/thread (no outcome ever changes across 6 days), so this may represent one ongoing conversation rather than a generalised cross-contact rule — operator should verify if this fires across multiple distinct leads.

At least 10+ events (e.g. events 7, 9, 11, 22, 23, 25, 26, 27, 7, 9) show the same two-part structure: (1) ask about fitness background/experience, then (2) present the binary 1:1 vs free taster choice with a closing question. This is consistent across SMS, WhatsApp, and Instagram, and across named leads and unclassified contacts. The main counter-evidence is that some takeovers skip the qualification question and go straight to a link or booking (events 3, 6, 17, 21) — these appear to be cases where the contact has already progressed past the qualification stage, so they don't truly conflict with the pattern.

At least 10+ events (events 5, 7, 8, 9, 16, 17, 20, 21, 25, 26, 38 and others) show the operator overriding the AI to deliver the same two-option framing ('free taster or 1:1?') to leads and unclassified contacts at the first meaningful onboarding touchpoint. The pattern is highly consistent in structure and language across different channels (SMS, WhatsApp, Instagram) and different contact names. The main counter-evidence is that some take-overs are pure link-sends (events 4, 15, 37) or logistical scheduling (events 11, 22, 29) rather than the binary-choice pitch, suggesting the two-option pattern applies specifically when the contact's entry route hasn't yet been determined.

All 4 events share identical policy parameters (stale_days=30, walker_kind=stalled_followup, expires_in_days=90), indicating a consistent automated policy rule being applied. The only variation is orq_row_count (1 vs 5), which does not change the action taken, suggesting it is not a discriminating condition. Counter-evidence is limited: all events cluster on two days and could reflect a single batch run rather than independently recurring decisions, so true generalisability across time is not fully confirmed yet.

All 4 events share an identical shape: lead-stage active contacts being closed due to exhaustion, non-aborted, within a ~5-second window suggesting a batch operation. The pattern is consistent with zero conflicts across all events. Confidence is capped below 0.9 because the tight timestamp clustering (all within ~5 seconds) may indicate a single bulk action rather than four independently recurring decisions, which slightly weakens generalisability as a repeating causal rule.

At least 10–12 events (events 5, 7, 8, 9, 16, 18, 20, 21, 25, 26, 35) show the same structural pattern: operator takes over, explains two onboarding routes (1:1 vs taster/class), and closes with a binary question. This is highly consistent across leads and unclassified contacts on WhatsApp and SMS. Counter-evidence: some take-overs are purely logistical (scheduling, links, member check-ins — events 1, 2, 11, 12, 14, 30, 31) which don't follow this pattern, but those involve different contact types or stages, so the pattern is real but scoped specifically to early-stage onboarding conversations with leads/unclassified contacts.

All 15 events share the same source (inbox/dismiss) and event type (inbox_resolved), with cards_cleared always >= 1, indicating a consistent operator behaviour of dismissing inbox cards in the board_reconcile domain rather than acting on them differently. The variation in drafts_dismissed (0–2) across events is not conflicting — it reflects incidental draft state at dismiss time, not a different action choice. The pattern is strong in volume and consistency within a single day, though the 7-day lookback only covers one day, which slightly limits generalisability.

All 5 events share an identical payload shape (exhausted, lead→closed, active, aborted=false), indicating a consistent automated or operator-confirmed behaviour whenever a lead exhausts its sequence. The pattern is clear and uniform with zero conflicting outcomes across the sample. Confidence is not higher because the sample size is small (5) and all events cluster within ~21 hours, which could reflect a single batch flush rather than a truly generalised recurring rule — operator should confirm this is intended system-wide policy and not a one-off cleanup.

At least 10–12 events (e.g. events 7, 9, 11, 22, 23, 25, 26, 27, 28) show the identical two-step shape: assess experience level, then present the 1:1-vs-taster binary choice. Events 5 and 11 also show the operator editing AI drafts to add this pattern where the AI missed it. Counter-evidence: a minority of events (e.g. 1, 3, 6, 17, 21) skip the experience question and go straight to a payment link or booking, suggesting the operator only applies this pattern when the lead's background is unknown — the trigger is specifically the absence of known fitness context.

All 8 events share the same shape: booking_confirmed + reason=appointment_upcoming + aborted=false → to_stage=booked, with no conflicting outcomes observed. 7 of 8 transitions are booked→booked (idempotent confirmations); 1 is member→booked, showing the pattern also applies when coming from a different prior stage. The consistent automated cadence (cluster at 10:15 and 08:15) suggests a scheduled trigger rather than ad-hoc operator action, which strengthens the pattern but also means it may already be automated — the operator should verify whether this is a system-driven rule or a manual override.

All 5 events share the same shape: stale_days=30, expires_in_days=90, and walker_kind is either 'stalled_followup' (4 events) or 'team_handler_idle' (1 event), suggesting a consistent policy trigger. The daily ~05:15 BST cadence further supports an automated scheduled rule rather than ad-hoc operator action. The only uncertainty is whether the orq_row_count threshold matters (values range 1–5 with no apparent filtering), and whether 'team_handler_idle' follows exactly the same rule or is a related but distinct sub-policy.

All 50 events share an identical structure (stale_days=30, expires_in_days=90) across three distinct walker_kinds, firing consistently at ~05:15 BST daily — strongly suggesting a scheduled policy sweep rather than ad-hoc decisions. The pattern is consistent with no conflicting behaviour observed (no cases where the same conditions did NOT produce a mute). Minor counter-evidence: orq_row_count varies (1-5), and it's unclear whether higher row counts ever suppress the mute, so the exact orq_row_count threshold is not fully resolved.

All 5 events share an identical shape: reason='exhausted', aborted=false, from_stage='lead', from_status='active', to_stage='closed'. The pattern is highly consistent with no conflicting outcomes observed. Confidence is capped below 0.9 because the dataset is small (5 events, clustered across only 2 days) and this may reflect a single automated system rule rather than a deliberate operator decision pattern — though the recurrence does confirm it generalises across multiple contacts.

49 out of 49 events share signal_kind=take_over with ai_draft_text=null, and across the operator's messages there is a highly consistent pattern: self-identification as Jack from CrossFit Bodmin, acknowledgement of the AI or a delay, and a clear CTA (link, call offer, or binary pathway choice). The pattern holds across all contact types and channels. The main variation is in the specific CTA offered (link vs call vs options), but the structural shape — humanise, apologise/acknowledge, give next step — is consistent enough to be actionable. Counter-evidence is minimal: event 17 (Instagram/business contact) deviates slightly as it is more social, but this is a single outlier.

The pattern is highly consistent across 30+ events: Jack takes over from AI, introduces himself, often apologises for delay or AI errors, asks about fitness background, and presents the binary choice of free taster vs. 1:1 (£45). This exact structure appears in events [10, 11, 13, 14, 15, 28, 32, 38, 43, 44, 48] among others. Counter-evidence: some take-overs are for existing members (check-ins, re-engagement) or are simple confirmations/closings, so the pattern is not universal across all contact types — it is primarily triggered by new lead/unclassified contacts at the discovery stage.

At least 15–20 events show the same structural shape: operator takes over (ai_draft_text=null), identifies himself by name, assesses experience level, and presents a binary 1:1-vs-taster choice with a closing question. This occurs consistently across WhatsApp and SMS channels for leads and unclassified contacts. Counter-evidence: a minority of take-overs are logistical (scheduling, links, member check-ins) with no binary-choice pattern (events 9, 10, 12, 28, 29), and two 'edit' events (1, 7) show a similar rewrite but from an existing draft, indicating the trigger isn't solely ai_draft_text=null — the operator also fires when the AI draft misses the personalisation/binary-choice format.

At least 10–12 events (e.g. events 7, 9, 11, 22, 23, 24, 25, 26, 27, 29) show the same two-part structure: ask about fitness background, then present the 1:1 vs free taster binary. This is highly consistent across different contacts, channels (SMS, WhatsApp), and days. The main counter-evidence is that some take-overs skip the background question and jump straight to logistics (e.g. events 1, 6, 14, 21) — but those appear to be contacts already qualified or further along, not first-touch leads. The pattern is strong enough for AI to propose this response shape when a new lead has not yet been asked about their background.

Across 30+ events, the operator consistently takes over with a personal introduction and a binary next-step offer whenever the AI has mishandled or stalled a lead/unclassified conversation — this pattern is highly consistent in structure (name intro + apology where needed + clear two-option CTA). Counter-evidence: a minority of take-overs (e.g. events 11-12, 35, 44-46) show no apology and no binary choice, just a simple confirmation or link, suggesting the pattern is strongest for cold/stalled leads but not universal across all contact states.

Events 14, 15, 17, 22 all show the same structure: operator takes over an SMS lead conversation, introduces as 'Jack from CrossFit Bodmin', often apologises for automations, and immediately asks about fitness/CrossFit background. Events 2 and 3 show a similar qualifying question pattern on leads mid-conversation. Counter-evidence: not all lead take-overs follow this exact template (e.g. events 4, 9, 13 skip the introduction and go straight to options/pricing), suggesting the intro+apology pattern is context-dependent — likely triggered specifically when the lead has received confusing automated messages rather than every new lead interaction.

All 4 events show the same shape: a closed record (from lead or member status) gets reactivated and is moved to 'follow_up' stage, with consistent flags (reason=reactivated, aborted=false). The pattern generalizes across two different from_stage values (lead, member), suggesting the destination stage is determined by the reactivation event itself rather than the origin. Confidence is moderated by small sample size (4 events, 3 identical + 1 variant) and short 7-day lookback, so robustness across longer time horizons is unconfirmed.

All 22 events show identical intent='resolved' and walker_muted=true whenever signal_kind='no-action-needed', with zero exceptions across a variety of contact_types, indicating a strong, consistent mapping. Confidence is capped below 0.8 because detail/reason are always null, so we can't confirm the underlying trigger logic beyond the signal_kind label itself, and all events come from a single dense burst (mostly one day), limiting evidence of generalization over time.

Strong pattern emerges across 8+ lead interactions where operator consistently offers specific time options (e.g., '8am, 10:30am or evening 4:30pm or 7pm') and provides booking links. Pattern is consistent across WhatsApp and SMS channels. Counter-evidence is minimal - no conflicting approaches seen for similar lead contexts.

All 3 events show identical pattern: operator records outcome for lead contacts by scheduling future appointments with specific times and types (intro→in_person→phone_call sequence suggests progression). Counter-evidence is limited sample size and unclear what triggers the specific appointment type selection.

22 of 24 events share an identical shape: intent=resolved, signal_kind=no-action-needed, walker_muted=true, with reason almost always null — strongly suggesting a consistent auto-resolution behavior independent of contact_type. The one event with reason='replied-call' still resolves identically, showing the pattern generalizes across trigger reasons, though the near-total absence of varying reason/detail fields means we can't yet confirm what differentiates this from cases needing action.

All 15 events share the identical trigger shape (stale_days=30, expires_in_days=90) and identical system action (auto-mute-inserted) across four distinct walker_kinds, showing this is a generalized threshold rule rather than a one-off. The pattern is consistent with no conflicting outcomes observed, though orq_row_count varies without visibly affecting the decision, suggesting it's not a causal factor — this was already the automated system's own consistent behavior rather than a discovered operator habit, so confidence is capped below 0.8 pending operator confirmation that this auto-mute logic should remain unchanged.

All 14 events share an identical trigger shape (stale_days=30 -> auto-mute-inserted with expires_in_days=90) across four different walker_kinds, showing this is a generalized system rule rather than a one-off, with orq_row_count varying freely (not a determining factor). Confidence is not higher because this looks like a deterministic automated policy rather than an operator decision, so it may just reflect a fixed threshold rule rather than a learned/discretionary pattern.

Strong pattern across 8+ lead interactions where operator consistently: (1) apologizes for delays/AI issues, (2) introduces himself as 'Jack', and (3) proposes specific call times (e.g. events 1,2,6,10,19,24). This appears to be a systematic approach to lead conversion. Counter-evidence includes some leads getting booking links instead of calls, but the call-scheduling pattern dominates lead interactions.

37 of 50 events (74%) are auto-stale sweeps with aa9_fired=false and no operator action, strongly indicating this is a fully automated, non-operator-driven resolution path that recurs constantly across the lookback window. The remaining events split between distinct patterns (live_lead_handover auto-approvals, and a couple of ownership-defer signals on the same entry), which are separate causal chains not conflicting with this one, but do mean 'auto-stale' itself has no single downstream operator action beyond the automated sweep — it's the sweeper's own decision, not a reactive operator behavior, which caps confidence.

All 4 events over 7 days show an identical shape: reason='exhausted', aborted=false, from_stage='lead', from_status='active', to_stage='closed', with no conflicting outcomes observed. This is a small but perfectly consistent sample, so confidence is above the noise threshold but not maximal given the limited event count and single domain.

Strong pattern seen in events 28 and 29 where operator explicitly apologizes for AI going 'off track' or 'technical bugs', tells contact to ignore problematic parts, then redirects. Event 28: 'apologies for the funny messages we have an ai assist that can go a little off track' and Event 29: 'ignore those bits, Cite is just a technical bug'. This is a clear recovery pattern when AI systems fail visibly to contacts.

All 4 events show identical to_stage='follow_up' on reactivation from a closed status, spanning two different from_stage values (lead, member), suggesting the transition target is driven by the reactivation event itself rather than the originating stage. Confidence is moderated by small sample size (4 events, 3 of which share the same from_stage='lead') and short 7-day lookback, so generalization across stages is only lightly tested.

All 7 events show an identical, uniform transformation: reason=exhausted, aborted=false, from_stage=lead/active → to_stage=closed, with no conflicting variants (e.g., no 'aborted:true' or different reason codes). This strongly suggests a deterministic system rule rather than case-by-case operator judgment. Confidence is capped below 0.9 because this looks like it may already be automated system behavior (not an operator decision) and the sample, while consistent, spans a short window with some near-duplicate timestamps (possible batch/dedup artifacts) that slightly reduce independence of evidence.

Across 15+ lead-classified events, the operator consistently follows a repeatable sales script: acknowledge/apologize, ask about fitness background or time preference, present taster-vs-1:1 options, quote pricing, and send a PushPress booking link. This structure recurs verbatim across multiple different leads (Jacob, Jo, Ross, Karen, Stuart, Kerry), showing strong generalization. Counter-evidence: some events are member/unclassified billing fixes or casual acknowledgments (e.g., events 9, 14-16) that don't fit this lead-onboarding shape, so the pattern is specific to new-lead intake, not all operator replies.

Across ~15 lead-tagged events, there's a consistent operator behavior: personalized greeting with name + operator name + gym name, acknowledgment of delay/automation issues, a qualifying question about fitness background, and an offer structured around taster vs 1:1 sessions, frequently closed with a PushPress booking link. This repeats across multiple different lead names (Jacob, Jo, Ross, Karen, Nicola, Stuart) with near-identical scripting, indicating a reusable operator playbook rather than one-off coincidence. Counter-evidence: some events (member/unclassified/casual_drop_in contact types) deviate from this exact script, showing the pattern is specific to 'lead' contact_type rather than universal across all contact types.

Clear pattern across 8+ lead interactions where Jack takes over, apologizes (delay/AI issues), introduces himself, and proposes calls with specific times. Consistent structure: apology + introduction + call scheduling. Counter-evidence: some variation in apology reasons and time specificity, but core pattern holds strong.

All 6 events show an identical shape: same event type, reason, stage transition, and status, indicating a consistent automated rule rather than operator discretion. Confidence is moderated rather than higher because this looks like a deterministic system behavior (not a judgment call) and the sample is clustered in time (4 events within seconds on one day), suggesting a single batch trigger rather than independently generalized cases.

All 11 events share identical event type, reason, aborted=false, to_stage='booked', and from_status='active', with from_stage being either 'booked' (9/11) or 'member' (2/11) — a consistent, non-conflicting shape. This looks like a recurring system-generated reconfirmation pattern rather than a one-off, though the repeated near-duplicate timestamps (e.g., 4 identical events at 15:15:00) suggest possible duplicate firing worth flagging to the operator rather than a distinct causal decision each time.

All 3 events share the same shape: affordance='fyi', contact_id=null, age_days == decay_days == 7, resolved by decay with no operator action. Two distinct event kinds (channel_outage x2, warmup_freshness_low x1) both follow the same resolution path, suggesting the pattern generalises across 'fyi' entry types rather than being kind-specific. Counter-evidence: small sample (n=3), and it's possible the operator simply never saw these rather than consciously delegating them to decay — confidence is moderate rather than high.

All 13 events share the identical trigger shape (stale_days=30, expires_in_days=90) across three distinct walker_kinds, and all consistently produce the same auto-mute action, suggesting a fixed policy rule rather than case-specific judgment. Confidence is not higher because orq_row_count varies (1-5) without visible correlation to any different treatment, and this looks like a deterministic system policy (likely a cron/batch job) rather than an operator decision, so 'reusable pattern' here mainly confirms existing automation rather than surfacing new operator intent.

At least 8–10 lead take-over events (events 6, 9, 10, 11, 12, 13, 18, 21, 23, 25, 27) follow the same shape: operator takes over from a stalled or automation-broken lead thread, re-introduces themselves, and either asks about fitness background or presents the taster-vs-1:1 binary choice. This pattern is consistent across SMS and WhatsApp channels. Counter-evidence: some take-overs for non-lead contacts (members, drop-ins, unclassified) follow different logic (pricing, links, casual closings), so the pattern is lead-specific and should not generalise to all contact types.

All 5 events share an identical signal shape: scheduler-initiated, AI-sent, SMS channel, tier_signoff intent, active lead with no terminal outcome and DND off. The pattern is internally consistent with zero conflicting actions observed. Confidence is capped at 0.72 rather than higher because events 1-4 occur within 5 seconds of each other on the same day, which suggests a potential burst/retry anomaly rather than 4 independent confirmations — effectively this may be 2 distinct scheduling occurrences (Jun 22 and Jun 28), not 5.

All 4 events share an identical shape: booking_confirmed with reason=appointment_upcoming, a self-loop from booked→booked, active status, and aborted=false. This is consistent across 3 separate days (Tue, Wed, Fri twice), suggesting a recurring automated or operator-triggered reminder confirmation pattern. Counter-evidence: the identical payload across all events could indicate a single systemic loop rather than distinct operator decisions, and the lack of stage change means it may be a no-op notification rather than a deliberate action requiring future automation.

All 4 events share an identical payload shape (cards_cleared=1, drafts_dismissed=0, source=journey) and occurred within a 4-second burst, strongly suggesting a consistent automated or semi-automated reconcile action rather than noise. There is no conflicting outcome — every event results in contact_closed under the same conditions. The tight time clustering (4 events in ~4 seconds) slightly reduces confidence that these represent independent operator decisions versus a single batch operation being logged multiple times, which could overstate generalisability.

All 24 events share identical intent='resolved', signal_kind='no-action-needed', and walker_muted=true, forming a highly consistent shape across a 7-day window and multiple contact types — strong evidence of a stable, repeated operator behavior. Confidence is capped below 0.8 because 'reason' and 'contact_type' vary without clear differentiation in outcome, and the events are clustered heavily on one day (Jun 29), suggesting possible batch/bulk-clearing behavior rather than a per-case decision pattern.

All 3 events share the same shape: AI-sent, WhatsApp channel, follow_up intent, active/non-terminal/non-DND contact, fired by a time-based or schedule-based caller_source. The pattern generalises across two distinct journey stages (lead and booked) and three distinct caller_sources, suggesting the rule is 'send follow-up whenever a scheduled trigger fires and the contact is reachable and non-terminal.' Confidence is moderate rather than high because 3 events is a small sample and all have null outcomes, so there is no evidence yet that the outbound message achieves the expected re-engagement.

Both events share an identical shape: channel_outage kind, fyi affordance, null contact_id, and age_days exactly matching decay_days (7), both resolved by the same decay mechanism at the same timestamp. This consistency across two independent workboard entries (2427 and 2409) suggests a deliberate policy — fyi-only outage notices with no contact are left to decay rather than actively closed. Counter-evidence: only 2 events and both fired simultaneously, which could reflect a batch decay run rather than a robust ongoing policy; confidence is moderate until more temporally distributed examples are observed.

All 7 events share an identical shape: AI sender, WhatsApp channel, chat_entry caller, active lead with no terminal outcome and DND off. This is consistent across multiple days and times, suggesting a deliberate routing rule rather than coincidence. The main caveat is that all events appear to be from a single contact's journey (no counter-examples of this trigger producing a different action), so the pattern may be contact-specific rather than fully generalised — operator should confirm this applies across multiple contacts.

All 4 events share identical shape: booking_confirmed, reason=appointment_upcoming, from_stage=booked, to_stage=booked, aborted=false — indicating a deliberate no-op transition pattern when an appointment is imminent for already-booked contacts. The pattern is consistent with zero conflicting signals. Confidence is moderate rather than high because 4 events is a small sample, all may belong to the same automated scheduled job (events at 08:15 on two days suggest a cron), and it's unclear whether this represents operator intent or a recurring system artefact worth reviewing.

Both events share identical shape: booked+active contact, dnd=false, no outcome yet, AI sender, WhatsApp channel, and the two distinct caller_sources represent a deliberate two-stage reminder sequence (day_before → day_of). This is a consistent automation pattern rather than noise. Confidence is moderate rather than high because there are only 2 events and no variation to test against — the pattern may be a single appointment's reminder pair rather than a truly generalised rule, though the structured caller_source naming strongly implies intentional sequencing.

All 4 events share an identical shape: AI sender, WhatsApp channel, chat_entry source, active lead with no terminal state and DND off. This consistency across multiple events on the same day suggests a stable automated pattern rather than one-off operator actions. Counter-evidence: all events occur on a single day (Jun 22) for what could be a single contact thread, so generalisability across contacts is unconfirmed — operator should verify whether this reflects multiple distinct leads or one conversation.

All 11 events share identical event type, reason, aborted flag, to_stage, and from_status, showing a highly consistent, repeatable pattern of appointment-upcoming reconfirmations. The only variation (from_stage 'booked' vs 'member') doesn't change the outcome, suggesting the trigger generalizes across entry states. Confidence is moderate-high rather than very high because this looks like an automated system heartbeat/reconfirmation rather than a distinct operator decision, so its value as an actionable 'pattern for operator' is limited beyond confirming expected automation behavior.

All 5 events share an identical shape: orq_row_count=1, stale_days=30, expires_in_days=90, across exactly two walker kinds, fired at consistent early-morning scheduled times. The repetition across distinct walker kinds and multiple days suggests a deliberate policy rule rather than a one-off. Counter-evidence: the sample is small (5 events), both walker kinds always appear with orq_row_count=1 so it's unclear if higher row counts would still trigger a mute, and we cannot rule out that this is purely automated system behaviour rather than an operator-authored policy.

All 4 events share an identical shape: AI sender, WhatsApp channel, chat_entry caller, active lead with no terminal outcome and DND off. This is consistent enough to suggest a standing rule that chat_entry triggers on WhatsApp from active leads always result in an AI reply. The main caveat is that all 4 events occurred on the same day (Jun 22), so this could reflect a single conversation thread rather than a generalised cross-contact pattern — confidence is moderate rather than high for that reason.

All 27 events share the same shape: workboard_review caller, team sender, lead journey_stage, dnd=false, and an operator-approved outbound send — this is a highly consistent pattern across two working days and three channels. The pattern is slightly weakened by one anomalous event (event 5: isTerminal=true, journey_status=closed, outcome=exhausted) where the operator sent to an exhausted-closed lead, suggesting the rule is not strictly enforced by the system and is human-overridable. No conflicting behaviour exists (operators always send, never skip), so the pattern is real but the anomaly prevents full confidence.

All 8 events share an identical signal shape: AI sender, WhatsApp channel, chat_entry caller, lead/active journey snapshot with no DND or terminal flags — this is highly consistent with no conflicting cases. The pattern is clear but confidence is moderated because all events map to a single contact/thread (no cross-contact variety is visible), so it may reflect one conversation rather than a generalised rule. Still, the uniformity warrants proposing this as a reusable trigger-action rule for operator review.

All 5 events share an identical shape: AI sender, WhatsApp channel, chat_entry caller source, lead stage, active status, DND false, non-terminal, no outcome — making the pattern structurally consistent across multiple events and two separate dates. The pattern is strong in uniformity but limited by the fact that all signals come from a single contact or a very narrow scenario, so it may reflect one ongoing conversation thread rather than a fully generalised rule. No conflicting cases (e.g., operator overrides or human replies) are present in the dataset to contradict it.

Across 20+ lead-classified events, there's a consistent operator playbook: friendly personal intro, ask about experience level, present taster-vs-1:1 choice, negotiate times, then send a pushpress plan link — this repeats across many different named leads (Jacob, Jo, Karen, Ross, Nicola, Stuart, Danielle). All events share ai_draft_text=null and signal_kind=take_over, meaning the AI never handled these and the operator fully manually crafts responses, suggesting this is a well-established human workflow ripe for AI templating. Counter-evidence: member and casual_drop_in contacts get different treatment (billing fixes, free week nudges), so the pattern is specific to 'lead'/'unclassified' contact_type at top-of-funnel, not universal across all contact types.

All 4 events share an identical shape: booking_confirmed triggered by appointment_upcoming, with no stage change (booked→booked) and aborted=false. The pattern is consistent with no conflicting cases. The main uncertainty is that this looks like a system-generated confirmation loop rather than a deliberate operator action — the repeated no-op transition (same from/to stage) may indicate an automated re-confirmation signal rather than a meaningful operator decision point, which limits confidence somewhat.

All 3 events share the same shape: AI-sent WhatsApp follow-ups triggered automatically (scheduler or appointment reminder), with the contact in an active, non-terminal state and no resolved outcome. The pattern holds across two different journey stages (lead and booked), suggesting it generalises to the active/non-terminal condition rather than being stage-specific. Counter-evidence: the small sample (3 events) and differing caller_sources mean this could reflect three distinct configured automations rather than a single unified rule — confidence is moderate rather than high.

25 of 26 events share the same shape: workboard_review-triggered, team-sender, lead stage, active status, dnd=false, isTerminal=false — across multiple channels (sms, whatsapp, instagram), indicating a consistent manual review-and-send workflow for active leads. The one exception (event #4) had outcome=exhausted and isTerminal=true with journey_status=closed, suggesting the operator also sends terminal/exhausted contacts occasionally — this is a mild conflict but represents only 1/26 cases and may be an edge-case or error. The pattern is robust enough to propose but should be reviewed for whether terminal/closed contacts should ever be included.

All 6 events share the same core shape: AI-initiated reply via chat_entry, active non-terminal lead, DND false — and in every case the outbound send was executed. The pattern holds across two channels (SMS and WhatsApp) and two outcome states (null and exhausted), suggesting the trigger is broad. The main uncertainty is whether 'exhausted' outcome on SMS (events 1–2) is intentional or an edge case that should be blocked; the operator may want to review whether sending on an exhausted outcome is correct behaviour.

All 5 events share an identical payload shape (cards_cleared=1, drafts_dismissed=0, source=journey), which is a consistent, repeatable signal rather than noise. The clustering of 4 events within a 4-second window on Jun 28 may indicate a burst/retry mechanism rather than 4 independent human actions, which slightly weakens confidence that each represents a truly independent operator decision. Counter-evidence: the small sample size (5 events, 2 distinct timestamps) and possible de-duplication failure mean the true independent-event count could be as low as 2, keeping confidence moderate rather than high.

All 4 events share an identical shape: same event type, same reason, same from/to stage, same status, and aborted=false — suggesting this is a deliberate, repeated pattern where an upcoming-appointment signal causes a booking re-confirmation in place rather than a stage transition. The lack of any conflicting events (e.g. aborted=true or a different to_stage) strengthens consistency. Counter-evidence: the small sample (4 events, all structurally identical) could reflect a single recurring automated trigger rather than a true operator decision pattern, so confidence is moderate rather than high.

All 3 events share the same shape: AI-sent WhatsApp follow-up, active/non-terminal contact, null outcome, triggered by automated sources (scheduler or appointment reminders). This suggests a deliberate automation pattern where the system sends follow-ups at key journey moments. The main caveat is the small sample (3 events) and that two different journey stages (lead, booked) are present — though both result in the same action, so stage appears not to differentiate the consequent. No conflicting signals observed.

All 15 events share an identical structural shape: stale_days=30 and expires_in_days=90 trigger an auto-mute-inserted decision, consistently across four different walker_kinds and varying row counts, suggesting a uniform system rule rather than a one-off. Confidence is moderated because this looks like an automated policy rather than an operator-driven choice (no evidence of operator override or exception), so the 'pattern' may simply reflect a fixed threshold rule already encoded in the system rather than a discretionary behavior to generalize.

All 5 events share an identical shape: AI sender, WhatsApp channel, chat_entry caller, lead/active journey state, DND off, non-terminal. This strongly suggests a standing rule that active leads messaging via WhatsApp are handled by the AI automatically. Counter-evidence: the sample is small (5 events, single contact likely), there is no variation in conditions to confirm the rule boundaries, and we cannot rule out that these are all the same conversation thread rather than independent trigger-action pairs.

All 16 events share the same source (inbox/dismiss) and event type (inbox_resolved), showing a consistent operator behaviour of dismissing inbox items one-by-one during board reconciliation sessions — two distinct sessions visible (Mon 29 Jun ~11:24–12:40, Tue 30 Jun ~11:57). The pattern is strong in action shape but the payload varies (cards_cleared ranges 0–2, drafts_dismissed ranges 0–2), suggesting the operator handles whatever is present without a fixed rule on composition. The main uncertainty is whether dismissal is always the right action or whether some items should instead be acted upon (no 'accept/commit' events observed, which could mean this is the correct resolution path or simply that those events are captured elsewhere).

All 5 events share an identical payload shape (cards_cleared=1, drafts_dismissed=0, source=journey), which is consistent enough to suggest a repeatable pattern rather than noise. The clustering of 4 events within 4 seconds on Jun 28 is slightly unusual and could indicate a batch/replay scenario rather than 5 independent operator actions, which slightly reduces confidence. No conflicting shapes are present, but the rapid burst warrants operator review before promoting this as a fully generalised rule.

All 3 events share the same shape: AI sender, WhatsApp channel, follow_up intent, dnd=false, isTerminal=false, outcome=null, journey_status=active — fired by distinct but schedule-based caller_sources. This suggests a consistent operator policy of allowing the AI to send follow-ups whenever a scheduled trigger fires and the contact remains reachable and unresolved. The main counter-evidence is the small sample size (3 events) and the variation in journey_stage (lead vs booked), which could mean different message content applies per stage — but the send-or-not decision appears uniform.

Pattern shows operator scheduling in-person follow-up (June 15) after intro appointment (June 10). The 5-day gap and progression from 'intro' to 'in_person' suggests a systematic lead nurturing approach. However, only 2 events limits confidence - need more instances to confirm this isn't coincidental timing.

All 7 events with entry_kind=live_lead_handover and entry_source=walker:live_lead_handover:auto-sent (events 4,6,12,17,18,32) consistently show aa9_fired=true and signal_kind=approve, a clean and repeated shape. This is distinct from the much larger auto-stale/sweeper cluster which is a separate, non-approval pattern (system marking cards stale for various reasons like contact_deleted, aged_out, dormant_handoff) with no consistent single operator action beyond the automated staleness marking itself — that cluster is background housekeeping rather than a decision pattern worth flagging as actionable.

All 3 events show consistent pattern of scheduling appointments with leads after recording outcomes. Appointment types vary (intro, in_person, phone_call) but core behavior is consistent - operator always schedules follow-up within reasonable timeframe. Pattern generalizes across different appointment formats, suggesting robust operator practice for lead nurturing.

Both events share an identical signal shape: a closed lead stage being reactivated and routed to follow_up, with no aborts, occurring on the same day within ~90 minutes of each other. This consistency across two events supports a genuine pattern rather than noise. However, the low event count (2) and the tight temporal clustering (same day) limit confidence — this could reflect a one-off batch operation rather than a recurring operational rule, so operator review is warranted.

Both events share identical shape: channel_outage kind, fyi affordance, null contact_id, and age_days exactly matching decay_days (7), resolved by the same decay mechanism at the same timestamp. This suggests a consistent policy of letting informational, contact-free channel outage entries expire automatically rather than requiring manual resolution. Counter-evidence: only 2 events sampled, both captured at the same instant which may indicate a batch run rather than independently recurring behaviour, slightly limiting generalisability.

Both events share identical shape: channel_outage, affordance=fyi, contact_id=null, age_days=decay_days=7, resolved by the same decay mechanism at the same timestamp. This suggests a consistent policy of letting FYI-level channel outages with no contact self-expire after 7 days. Counter-evidence is thin — only 2 events, both captured at exactly the same time, which could indicate a batch run rather than repeated independent decisions, slightly limiting generalisability.

There is a clear, repeated multi-step script across 8+ lead-facing events: greet by name, mention automation issues, ask about experience level, then offer taster vs 1:1 with pricing and a booking link. This is consistent enough (events 3,4,10,13,14,15,16,17,22) to be a reusable template for lead onboarding takeovers. Confidence is moderated because member/unclassified contact types show different, less consistent behavior (billing fixes, casual acknowledgments) that shouldn't be conflated into the same pattern.

Across 30 events, a clear recurring script emerges for lead/unclassified contacts: personalized greeting apologizing for automation issues, a qualifying question about experience level, then routing to taster/1:1 options with pricing and a PushPress link — this repeats in at least 8-10 near-identical instances (e.g. events 10,11,12,20,21,22,23,25,30). Counter-evidence: member and casual_drop_in contacts get different treatment (billing fixes, simple link sends), and the exact sequencing/wording varies enough that this is a soft conversational pattern rather than a rigid rule, so confidence is moderate rather than high.

All 3 events share the same shape: stale_days=30, orq_row_count=1, expires_in_days=90, and the auto-mute-inserted action — two different walker_kinds both trigger the same response, suggesting the pattern generalises across walker types rather than being kind-specific. The consistent parameter values (30-day staleness threshold, 1-row count, 90-day expiry) across events support a policy-driven rule rather than coincidence. Counter-evidence: only 3 events in 7 days is a small sample, and the two distinct walker_kinds could indicate separate policies that happen to share parameters rather than a single unified rule.

All 17 events share identical source/action shape (inbox/dismiss clearing 0-2 cards and 0-3 drafts), and they cluster into two clear batch sessions (Jun 29 morning/midday, and isolated events Jun 30/Jul 4), indicating a repeatable 'bulk dismiss during reconciliation' behavior. However, the pattern is purely mechanical (same action repeated) with no distinguishing antecedent signal beyond 'inbox has items' — there's no evidence of what specifically triggers a session (time-based, volume-based, or manual habit), so confidence is moderate rather than high.

All 5 events show an identical shape: same from_stage, from_status, reason, to_stage, and aborted=false, indicating a consistent system/operator behavior rather than a one-off. However, all events are system-generated (not clearly operator-decided) and cluster within a very short window (4 within 5 seconds, plus 1 a day later), suggesting they may stem from a single automated batch process rather than distinct causal decisions — this limits confidence that it generalizes as a discretionary operator pattern.

All 3 events share the same shape: an 'fyi' affordance signal aging exactly to its 7-day decay_days threshold, uniformly resolved by the same automated decay mechanism, across two different kinds (warmup_freshness_low, channel_outage) and separate workboard entries, suggesting a generalizable decay rule rather than a one-off. Confidence is moderate rather than high because the sample is small (n=3), all events come from a single automated source (orq-decay/auto-stale) rather than reflecting a distinct operator decision, and there's no evidence yet of how non-fyi affordances or contact-linked entries behave under decay.

Both events share an identical shape — closed leads being reactivated and moved to follow_up — occurring twice within roughly 90 minutes on the same day, suggesting a deliberate operator behaviour or process. However, with only 2 events and no variation in conditions, it is difficult to rule out a coincidental batch operation or a one-off bulk reactivation rather than a repeatable general rule. Confidence is modest because the sample is minimal, both events occurred on the same day (possibly a single operator action), and there is no counter-evidence but also no diversity of context to confirm generalisation.

All 7 events share identical event/reason/aborted values and consistently transition to_stage='booked' from either 'booked' (5 events) or 'member' (2 events), with no conflicting outcomes — this is a stable, repeatable shape. Confidence is moderate rather than high because this looks like a system-driven notification/reconfirmation rather than a distinct operator decision, and the pattern is narrow (single reason code, single domain) with limited variation to stress-test generalisability.

All 7 events are identical in shape (same reason, from_stage, from_status, to_stage, aborted=false), showing a consistent, repeatable system behavior rather than an operator decision. However, this looks like a deterministic system/automation outcome rather than a discretionary operator action, and the tight clustering of 4 events within 2 seconds on Jun 28 suggests a batch/bulk process rather than distinct causal instances, slightly limiting generalizability.

All 4 events show identical structure: same intent, channel, sender, and scheduler-driven trigger with no outcome/terminal state, recurring every 2-4 days over the lookback window. This looks like a consistent scheduled nurture cadence rather than a one-off action, but sample size is small (4 events, single lead/thread inferred) and lacks outcome data to confirm effectiveness.

All 3 events share the identical shape: age_days == decay_days (7), affordance='fyi', contact_id=null, resulting in auto-resolution by the decay mechanism — consistent and non-conflicting. Confidence is moderate rather than high because the sample is small (3 events, 2 distinct kinds) and this reflects an existing automated system behavior (source=orq-decay/auto-stale) rather than a discovered operator decision, so it's more a confirmation of known logic than a novel inferred pattern.

All 3 events show an identical shape: a closed lead reactivated and moved to follow_up, with no conflicting outcomes (aborted always false, to_stage always follow_up). This is consistent enough to suggest a deterministic transition rule, though the sample is small (3 events, all within a 4-day window) and could reflect a single automated rule rather than a generalizable operator judgment call.

All 7 events share identical event/reason/aborted values and the same to_stage='booked', showing a consistent, non-conflicting transition triggered by 'appointment_upcoming'. The only variation is from_stage (booked in 6/7 cases, member in 1/7), which is a minor branching detail rather than a contradiction, but the small sample and single non-booked source stage limit full generalisability.

All 3 events share identical parameters (stale_days=30, orq_row_count=1, expires_in_days=90), occurring across two different walker_kinds, which suggests a generalised policy rule rather than a one-off. The pattern holds consistently with no conflicting values observed. However, 3 events over 7 days is a small sample, and it's unclear whether the 90-day mute duration or the orq_row_count=1 threshold are hard rules or coincidental defaults — operator should confirm these are intentional policy constants.

Both events share an identical shape: booking_confirmed with reason=appointment_upcoming, same-to-same stage transition (booked→booked), aborted=false, and active status — suggesting this is a deliberate re-confirmation pattern triggered by an upcoming appointment rather than a stage advancement. The pattern is consistent across two consecutive days, implying a recurring scheduled trigger. Counter-evidence: only 2 events in 7 days is a thin sample, and same-stage transitions could also indicate a misconfigured workflow rather than intentional behaviour — operator should verify intent.

All 5 events are identical in shape (same source, cards_cleared=1, drafts_dismissed=0), suggesting a consistent, repeatable trigger-action mapping. However, 4 of 5 events are clustered within seconds of each other on one day, which may reflect a single burst/backfill rather than independent generalizable occurrences, so robustness across varied contexts is unproven.

All 3 events show identical structure: a closed lead reactivated and transitioned to follow_up, with no conflicting outcomes observed. However, the sample is small (3 events, all within a 4-day window) and comes from a single domain/cell, so generalisation beyond this narrow slice is uncertain — this looks like a system-default transition rather than a discretionary operator choice, which slightly limits its value as an 'action pattern' but still holds as a reliable detectable rule.

Both events share an identical shape: booking_confirmed fires with reason=appointment_upcoming on an already-booked/active contact, causing a self-loop (booked→booked) with aborted=false. This suggests a deliberate pattern where an imminent appointment triggers a confirmation re-check rather than a new stage transition. Counter-evidence: only 2 events in 7 days is a thin sample, and the self-loop could equally be a system artefact rather than an intentional operator action — confidence is kept low accordingly.

Both events share an identical snapshot shape (active lead, no DND, no terminal outcome, no resolved outcome) and were both triggered from workboard_review within 8 seconds of each other, suggesting a deliberate operator send action during a review session. The two events differ only in channel (SMS vs WhatsApp), which may indicate multi-channel outreach rather than two independent decisions, slightly weakening the generalisation. Confidence is modest because only 2 events exist, the near-simultaneous timing could mean a single review interaction rather than two separate pattern instances, and channel selection logic is not yet clear.

All 3 events share identical structural shape: affordance='fyi', age_days equals decay_days (7), contact_id=null, and the same resolution mechanism (orq-resolved-by-decay). This consistency across different 'kind' values (warmup_freshness_low, channel_outage) suggests the decay behavior generalizes across signal types rather than being kind-specific. However, confidence is moderate because this is a system-automated decay process rather than an operator decision, sample size is small (3 events, 2 distinct kinds), and there's no evidence of operator override behavior to confirm this is desired versus merely default system behavior going unchallenged.

All 5 events share an identical signal shape: scheduler-triggered, AI sender, SMS channel, tier_signoff intent, active lead with no DND or terminal state. This strongly suggests a repeating automated pattern. However, events 2–5 occurred within 2 seconds of each other on the same day, raising a concern about duplicate firing (a burst anomaly) rather than 5 independent confirmations — which limits confidence. Event 1 occurring ~20.5 hours later provides one genuine independent data point supporting recurrence.

All 16 events share the identical source (inbox/dismiss) and event type, forming a consistent batch-clearing behavior over a short session (11:24-12:40 on Jun 29, plus one follow-up Jun 30), which supports a real habitual pattern of rapid-fire dismissal. However, the antecedent is weak as a 'detectable trigger' since we don't see what preceded each dismiss (e.g., what the card/draft content was), so this is more a description of a recurring action shape than a true if-X-then-Y causal rule tied to external context — confidence is kept moderate accordingly.

All three events share the identical shape: distinct appointment IDs reaching the outcome-loop terminal node with outcome='unknown', spread across three separate days, suggesting this is a recurring gap rather than a one-off. The pattern is consistent with no conflicting signals (no events show a different outcome at terminal state). Confidence is modest because three events is a small sample and there is no visibility into whether an operator action was actually taken in response, or what the intended resolution workflow is — this may simply be a system logging artefact rather than an actionable operator pattern.

All 5 events share an identical payload shape — same intent, channel, sender, snapshot fields, and caller_source — confirming a consistent automated pattern rather than manual variation. However, events 2–5 fired within 2 seconds of each other on the same day, suggesting possible duplicate/burst firing from the scheduler rather than 5 independent trigger instances; this reduces confidence that each event represents a truly distinct causal occurrence. The pattern itself (scheduler + active lead + no terminal state → AI sends tier_signoff SMS) is detectable and actionable, but the burst behaviour warrants operator review to rule out a scheduling bug inflating the apparent frequency.

Both events share an identical shape — appointments reaching 'outcome-loop-terminal' with outcome='unknown' on consecutive days — suggesting a recurring workflow gap where appointments exit the automated loop without a resolved outcome. The pattern is consistent across two events with no conflicting signals, but the sample size is small (n=2) and neither event shows what the operator actually did in response, only that the terminal state was reached. Confidence is modest because we cannot confirm the consequent action from these signals alone — the causal link is inferred, not observed.

15 of 16 events share an identical shape (source=inbox/dismiss, small integer counts of cards_cleared/drafts_dismissed, clustered in tight bursts within minutes), indicating a repeated manual triage habit of clearing inbox items individually. Confidence is moderate rather than high because the pattern is really just 'operator uses inbox/dismiss repeatedly' — there's no clear distinguishing antecedent condition (e.g. specific card type or count threshold) that predicts *when* this triage happens versus other resolution paths, and all 16 events come from one clustered session on Jun 29, limiting generalisability across days.

All three events share an identical structure: the outcome-loop reached its terminal state but the outcome value is still 'unknown', across distinct appointment IDs on three separate days. This consistent shape suggests a systemic gap — appointments are closing without outcomes being captured — rather than a one-off edge case. Confidence is moderate rather than high because we only observe 3 events and cannot yet distinguish whether this is an operator-driven corrective action pattern or simply a recurring data quality failure with no established response; the consequent action is inferred from the likely operator intent, not directly observed.

Both events share the same signal_kind ('record_outcome') and both result in a scheduled appointment being logged with a concrete datetime and appointment type, suggesting a consistent operator workflow of capturing booking outcomes. The pattern holds across two different contact types ('lead' and 'unclassified') and two appointment types ('phone_call' and 'intro'), which adds mild breadth but also introduces variability. Confidence is moderate because the sample size is only 2 and the differing contact/appointment types mean the trigger conditions are somewhat broad — more events would be needed to confirm this generalises reliably.

Both events share identical context shape — active lead, no outcome, not terminal, not DND, triggered from workboard review — and both result in operator-review sends, just across different channels (SMS and WhatsApp) within 8 seconds of each other, suggesting a multi-channel send pattern for the same contact or cohort. The sample size is only 2 events, which limits confidence, and the channel variation (SMS vs WhatsApp) may reflect a deliberate multi-channel strategy or simply two independent contacts reviewed in sequence — this ambiguity cannot be resolved from the data alone. Pattern is plausible but should be validated against a larger window before automation.

Both events share identical stale_days (30) and expires_in_days (90) values, and both are auto-mutes inserted by the walker-policy decision engine, suggesting a consistent policy rule: mute walkers that have been stale for 30 days for a 90-day window. The two walker_kinds differ (stalled_followup vs post_apt_outcome), which could indicate the rule applies broadly across walker types rather than being kind-specific. Counter-evidence: only 2 events exist within 7 days (orq_row_count differs: 5 vs 1), which is a thin evidence base — confidence is kept modest until more events confirm generality across additional walker kinds.

Both events share an identical shape: active leads with no outcome, no DND, non-terminal status, triggered from workboard_review, with the team as sender — differing only in channel (SMS vs WhatsApp). This suggests the operator sends outbound messages across multiple channels to the same or similar leads when reviewing the workboard. Counter-evidence: only 2 events within 8 seconds makes it plausible these are two channel attempts for a single contact rather than two independent cases, which limits generalisability. Confidence is moderate rather than high due to the small sample size and near-simultaneous timing.

All 3 events share an identical shape: distinct appointment IDs producing outcome='unknown' via 'outcome-loop-terminal', spread across 3 separate days — suggesting a systemic loop-termination failure rather than a one-off. The pattern is consistent with no conflicting signals. Confidence is modest because the event count is low (exactly at threshold), there is no visibility into what the operator actually did in response, and 'unknown' could reflect legitimate no-shows or cancellations rather than a system failure requiring action.

15 of 17 events occurred in a single tight session (Jun 29, 11:24-12:40), each dismissing 1 card and 0-2 drafts, showing a consistent per-item dismissal behavior rather than a bulk clear — this is a real repeated shape. However, the pattern is purely descriptive of operator habit (dismiss small batches serially) rather than a clear trigger-condition causing a distinct decision; there's no evidence of what specifically caused dismissal (e.g. staleness, irrelevance) versus just routine inbox grooming, so confidence is moderate.

All 16 events share the identical stale_days=30 and expires_in_days=90 thresholds, consistently triggering the same auto-mute action across four distinct walker_kinds, suggesting a fixed policy rule rather than case-specific judgment. Confidence is moderated because the sample is dominated by a fixed system parameter (stale_days/expires_in_days never vary) with no visible alternative outcome (e.g. no observed cases where staleness didn't trigger mute), making it hard to confirm the trigger threshold is causal versus simply the only value ever logged.

14 of 16 events occurred in a tight burst on Jun 29 (11:24-12:40), all using the same source (inbox/dismiss) and shape (cards_cleared 0-2, drafts_dismissed 0-2), suggesting a repeatable batch-clearing behavior rather than isolated incidents. However, the pattern is essentially 'operator dismisses inbox items' which is somewhat tautological with the event type itself, and there's no clear antecedent signal distinguishing when a card is cleared vs. just drafts dismissed vs. both — the values vary without an obvious external trigger, so confidence is capped as moderate rather than high.

Both observed events show identical from_stage/from_status/to_stage values, suggesting a consistent transition rule. However, with only 2 events occurring within under 2 hours of each other (possibly the same batch or automated process), there's limited evidence this generalises across different times, operators, or lead contexts.

All 5 events show an identical shape (same source, cards_cleared=1, drafts_dismissed=0), which is consistent evidence of a stable reconciliation behavior. However, 4 of the 5 events are clustered within 4 seconds on the same day, suggesting they may be duplicates or a single batch operation rather than 5 independent generalizable occurrences, so confidence is capped below high certainty.

All three events share a consistent surface pattern (high diff_ratio, sent_len much shorter than canonical_len, same tree, facts_resolved=false), suggesting canonical texts are long instructional/meta prompts while actual sends are short conversational replies — a structural mismatch rather than true policy violation. However, the underlying causes differ across events (off-topic deflection, exact-text non-compliance, and legitimate paraphrased discovery-answer handling), so the 'why' driving divergence isn't uniform, and only 3 events limits generalisability — this should be reviewed as a detector calibration issue rather than a single actionable operator behavior.

4 of 6 events show a consistent shape (cards_cleared=1, drafts_dismissed=0) clustered on Jun 28, suggesting contact_closed reliably triggers a card clear. However, 2 more recent events (Jun 29, Jul 03) diverge — one clears nothing but dismisses a draft, breaking the pattern's recency and consistency. This conflicting recent behavior lowers confidence and suggests the rule may be conditional on an untracked factor (e.g., whether a draft vs. card exists).

All 7 events share an identical shape (same intent, sender, channel, caller_source, snapshot), showing this is a stable, repeatable automated behavior rather than a one-off. However, this is a machine-generated scheduler action, not an operator decision — there's no evidence of operator judgment or a detectable pre-condition an operator set; it may simply reflect a fixed cron/scheduler cadence rather than a causal 'operator does Y when X' pattern, and 4 of the 7 events are near-duplicate timestamps (same day, seconds apart), suggesting possible retry/batching noise rather than distinct decisions.

44 of 50 events share the identical shape (signal_kind=auto-stale, aa9_fired=false, source=stale-card-sweeper) across varied stale-reasons, which is a strong repeated signature. However, this is not really an 'operator action' pattern — aa9_fired=false and null entry_source indicate this is a fully automated sweeper process with no human/operator decision point, so framing it as a causal operator behavior pattern is questionable; the only true operator-driven signal in this dataset is the small aa9_fired=true 'approve'/live_lead_handover cluster (5 events), which is too thin and homogeneous (all auto_sent) to establish a conditional operator choice pattern rather than another automation.

All 6 events share an identical shape (same intent, sender, channel, caller_source, and snapshot state), which supports a consistent automated send pattern. However, this is a homogeneous, low-variance signal set with no contrasting cases (no examples of suppression, DND-true, or non-lead stage) to confirm the trigger is truly discriminating rather than just 'this is what the scheduler always does' — so it's more a confirmation of existing automation than a newly discovered causal rule.

4 of 29 events show outcome='exhausted' on sms replies (events 7, 21, 25, 26), all with isTerminal=false and journey_status still 'active', suggesting a recognizable trigger state. However, the consequent is ambiguous: there's no clear, consistent follow-up action visible in this data (no subsequent event shows escalation, channel switch, or closure directly tied to these exhausted events), so this is more an observed state pattern than a confirmed action-response pattern. Flagging for operator review rather than asserting a strong causal rule.

Both events share the same signal_kind ('record_outcome') and both include a 'scheduled_for' timestamp pointing to a near-future appointment, suggesting this is the operator's consistent workflow for logging that a booking has been made after a contact. However, the two events differ in contact_type ('unclassified' vs 'lead') and appointment_type ('intro' vs 'phone_call'), meaning the trigger conditions are somewhat varied. With only 2 events the confidence is modest — the shape is consistent but the small sample size means this could be coincidence rather than a generalizable rule.

Both events share an identical shape: appointments reaching the terminal state of the outcome loop with an unresolved 'unknown' outcome, occurring on consecutive days. This suggests a recurring failure mode rather than a one-off anomaly. However, with only 2 events and no visibility into whether the operator actually took a specific action in response, confidence is modest — the pattern may reflect a systemic data issue rather than a deliberate operator response trigger.

Both events share the same signal_kind ('record_outcome') and both result in an appointment being scheduled roughly one day after the booking is captured, suggesting a consistent same-day-booking-for-next-day pattern. However, the contact_type and appointment_type differ between the two events (lead/phone_call vs unclassified/intro), making it unclear whether the trigger is truly generalised or whether these are coincidentally similar outcomes. With only 2 events and notable variation in contact classification, confidence is low — this is plausible noise rather than a robust pattern.