Traffic up, conversion collapsed

MTI-002 · graph metric-truth · 15 ledger records

What this run caught

No violationsFabricated citations, unapproved effects and tool violations are all zero

Adjudication

Agent verdict
instrument_failure
bot_traffic_mixed
Answer key
instrument_failure
bot_traffic_mixed
Outcome
correct
12 of 12 citations resolve

Graph execution

NodeKindStateRecordsToolsms
intake transform passed 1 0
map_terrain agent passed 1 10 582493
hypothesize agent passed 3 0 184480
gather_witnesses agent passed 3 26 3858993
witness_gate gate passed 1 0
adjudicate agent passed 1 0 126992
citation_gate gate passed 1 0
propose_remediation agent passed 1 5 597785
human_approval write human passed 1 0
write_back write tool passed 1 0
report transform passed 1 0

Witness ledger

idClaimSourceSupportsRefutes
e11 Aggregated to weekly buckets, signup_conversion_daily shows sessions rising 1665→3628 from week0 (2026-06-01) to week4 (2026-06-29..07-05) while conversion fell 0.3562→0.2679, and absolute signups did NOT stay flat but rose 593→707→803→895→972. warehouse_query
urn:li:dataset:(urn:li:dataPlatform:dbt,witnessgraph.marts.signup_conversion_d
- h2
e12 Every web_sessions row in every week carries is_automated=0; zero sessions are flagged automated, so the entire session surge sits in is_automated=0 rows. warehouse_query
urn:li:dataset:(urn:li:dataPlatform:postgres,witnessgraph.raw.web_sessions,PRO
h2 h1
e13 Segmenting the denominator and conversion by channel shows the SAME four channels throughout with no new/low-intent channel, each growing ~2.1–2.3x wk0→wk4 (mix ~constant), and each channel's OWN conversion falling ~1/4–1/3 rather than staying stable. warehouse_query
urn:li:dataset:(urn:li:dataPlatform:postgres,witnessgraph.raw.web_sessions,PRO
- h2
e14 distinct customer_code count is flat (~393–397) across weeks 0–4 while sessions rose 1665→3628, so sessions-per-customer more than doubled (4.19→9.16); the surge is repeated rows from a stable customer population, not new visitors. warehouse_query
urn:li:dataset:(urn:li:dataPlatform:postgres,witnessgraph.raw.web_sessions,PRO
h3 h2
e15 The signup_conversion_daily denominator is a raw COUNT(*) of web_sessions rows per day with no filter on is_automated, channel, or customer_code, so any rise in rows-per-customer inflates the denominator directly with no new visitors required. warehouse_query
urn:li:dataset:(urn:li:dataPlatform:dbt,witnessgraph.marts.signup_conversion_d
h1, h3 h2
e1 The signup_conversion_daily view computes its `sessions` denominator as `(SELECT COUNT(*) FROM web_sessions w WHERE w.day = ...)` with a filter on day only and no predicate on is_automated, channel, or browser_version, so every session row for a day is counted in the denominator. config
urn:li:dataset:(urn:li:dataPlatform:dbt,witnessgraph.marts.signup_conversion_d
h1 -
e2 In raw.web_sessions the is_automated flag equals 0 for 100% of rows in every week (is_automated=1 count is 0 in weeks 0-5), so the session surge from 1701 (week 1) to 3628 (week 4) lies entirely within is_automated=0 rows and adding an is_automated filter to the denominator would remove nothing. warehouse_query
urn:li:dataset:(urn:li:dataPlatform:postgres,witnessgraph.raw.web_sessions,PRO
- h1
e3 Between pre-surge weeks 1-2 and surge weeks 3-4, browser_version=chrome-126 sessions rose 3.6x (1027 to 3746) while every other browser_version stayed flat (firefox-128 898->887, safari-17 826->882, chrome-127 848->848), so the surge clusters on a single browser signature rather than spreading across user agents. warehouse_query
urn:li:dataset:(urn:li:dataPlatform:postgres,witnessgraph.raw.web_sessions,PRO
h1 h2
e4 From pre to surge weeks all four channels roughly doubled in sessions (referral 906->1615, affiliate 906->1598, paid_social 920->1578, organic 867->1572) while each channel's distinct customer_code count stayed flat (~345-367 both periods), so there is no new or scaled acquisition channel and no rise in distinct visitors. warehouse_query
urn:li:dataset:(urn:li:dataPlatform:postgres,witnessgraph.raw.web_sessions,PRO
h1 h2
e5 Distinct customer_code per week is flat (~393-397 in weeks 0-4) while the maximum sessions-per-customer explodes from 13 (week 0) to 1003 (week 3) to 1915 (week 4) and the average rises 4.19->9.16, so the added sessions come from a tiny tail of customer_codes replaying up to ~1900 sessions each, not from additional visitors. warehouse_query
urn:li:dataset:(urn:li:dataPlatform:postgres,witnessgraph.raw.web_sessions,PRO
h1 h2
e6 Excluding chrome-126 from the denominator leaves non-chrome-126 sessions flat (~1280-1315/week) and yields a conversion rate that is stable-to-rising (0.463->0.547->0.628->0.681->0.747 across weeks 0-4) while the blended rate falls (0.356->0.268), so the entire conversion decline is attributable to the chrome-126 segment inflating the denominator. warehouse_query
urn:li:dataset:(urn:li:dataPlatform:dbt,witnessgraph.marts.signup_conversion_d
h1 h2
e21 In raw.web_sessions the weekly session-row count rose from 1665 (days 1-7) to 3628 (days 29-35), a +118% increase, while distinct customer_code stayed flat at ~393-397 per week, so sessions-per-distinct-visitor doubled from 4.19 to 9.16; SUM(is_automated) was 0 in every week. warehouse_query
urn:li:dataset:(urn:li:dataPlatform:postgres,witnessgraph.raw.web_sessions,PRO
h3 h1, h2
e22 The signup_conversion_daily view defines its denominator as (SELECT COUNT(*) FROM web_sessions WHERE w.day = ...), counting raw session rows with no de-duplication by customer_code and no filter on is_automated, so any rise in session-rows-per-visitor deflates the conversion rate one-for-one even when distinct visitors and signups are unchanged. config
urn:li:dataset:(urn:li:dataPlatform:dbt,witnessgraph.marts.signup_conversion_d
h3 -
e23 From week 2 (days 15-21) to week 4 (days 29-35) sessions roughly doubled in all four channels (referral 476->921, organic 462->913, affiliate 474->902, paid_social 486->892) while distinct visitors per channel stayed flat (referral 273->267, organic 261->262, affiliate 265->285, paid_social 264->270), so the session inflation is uniform across channels rather than concentrated in any single (e.g. pricing-page-originated) entry point. warehouse_query
urn:li:dataset:(urn:li:dataPlatform:postgres,witnessgraph.raw.web_sessions,PRO
- h3, h2
e24 release_deploy_log contains only two shipped deploys in the window - D-0012 'quick_reorder' (day 12) and D-0020 'price_alerts' (day 20) - and no entry describing a pricing-page redesign or any change to how web_sessions rows are written. warehouse_query
urn:li:dataset:(urn:li:dataPlatform:postgres,witnessgraph.raw.release_deploy_l
- h3
e25 metric_definition_history records exactly one metric-definition entry (version 1, day 1, metric 'active_users_daily') and no recorded change to the session or conversion unit of measurement anywhere in the window. warehouse_query
urn:li:dataset:(urn:li:dataPlatform:postgres,witnessgraph.raw.metric_definitio
- h3
e26 schema_migration_log contains only one migration in the window - M-0001 (day 2) 'CREATE INDEX idx_orders_customer ON orders(customer_code)' - and no migration altering web_sessions or any session-generating write path. warehouse_query
urn:li:dataset:(urn:li:dataPlatform:postgres,witnessgraph.raw.schema_migration
- h3

This page is built from the run ledger alone. It never reads the agent's own summary of what it did.