// Proof

Not a claim. A test run.

Every Beta tool in the toolkit, run against real test data before it earned that label. Here's exactly what came back.

Signal Accuracy Auditor

27 synthetic flagged-account events across 5 signal types, including two signal types with an identical 66.7% raw false-positive rate, one with enough resolved volume to judge and one deliberately too small, plus 2 still-pending flags used to check they get excluded from every rate.

1/1

High false-positive signal caught

1/1

Identical-rate small sample correctly withheld

2/2

Pending flags correctly excluded

→Correctly flagged one signal type at a 66.7% false-positive rate across 6 resolved flags, while correctly withholding judgment on a different signal type at the exact same 66.7% rate, because that one only had 3 resolved flags on record.
→Correctly excluded 2 still-pending flags from every rate calculation rather than letting an unresolved batch drag a signal type's accuracy down artificially.
→Correctly computed a 40% false-positive rate on a mid-tier signal and left it unflagged against the 50% default threshold, avoiding an overly aggressive call on a genuinely mixed signal.

Two signal types looked identically unreliable by raw rate alone. Only one of them had enough evidence to actually say so.

Buyer Research Brief Generator

4 synthetic accounts covering balanced research, single-threaded deep research, a boundary case just under the single-threaded volume threshold, and an account whose technical research never once touched the pricing page.

1/1

Single-threaded account caught

1/1

Below-threshold account correctly cleared

1/1

Pricing gap caught despite deep engagement

→Correctly flagged an account with 6 events from one contact as single-threaded research, while correctly clearing a different single-contact account with only 1 event, below the 5-event volume threshold.
→Correctly caught an account that covered security, implementation, ROI, and integrations in real depth across 3 distinct contacts, yet never once touched pricing or product, the exact pattern a generic discovery script would miss.
→Correctly excluded an unmatched content view ("Blog Post") from every specific topic's coverage instead of letting it inflate a real topic.

The account with the most technical depth was also the one about to get a pricing conversation it had never actually asked for, and the brief caught it.

Revenue Infrastructure Builder · Phase 1: CRM Data Audit

A synthetic CRM export seeded with realistic messy data: 12 contacts, 11 companies, 10 deals.

93.3/100

Completeness score

Duplicate clusters found

Stale deals flagged

→Caught 3 contact duplicate clusters, including a same-email different-casing pair and a last-name typo (Vasquez/Vazquez), plus 2 company duplicates matched on shared domain.
→Flagged 5 open deals with no activity in 45+ days, worst at 173 days idle, sorted so the coldest deals surface first.
→Caught 1 deal missing a pipeline stage entirely and 1 closed-won deal missing its deal amount, both invisible to standard reporting until flagged.

Zero false positives on the genuinely distinct records in the same test set.

Revenue Infrastructure Builder · Phase 4: Pipeline Forecast

10 deals scored against a $300K quarterly quota.

$262K

Best case

$173.2K

Likely case (weighted)

0.85x

Coverage ratio

→Closed-won revenue ($52K) derived directly from the CRM export, not entered as a self-reported number.
→Correctly downgraded a Negotiation-stage deal out of the commit tier because it had gone 68 days without activity, stage alone would have called it a safe bet.
→Flagged a 0.85x coverage ratio against remaining quota, below the 2x threshold that signals real pipeline risk rather than just weak prioritization.

Every number reconciles: best case, likely case, and worst case all tie back to the same underlying deal list.

ICP List Builder

7 raw candidate companies scored against an ICP: software, SaaS, or fintech, 50-500 employees, $5M-$50M revenue, US or Canada.

Strong fits (P1)

Good fit (P2)

Disqualified

→Correctly tiered a 650-employee SaaS company as P2 rather than P1, oversized on headcount and undersized on revenue, with partial credit rather than a hard zero for the ambiguity.
→Zeroed out a non-profit candidate entirely on the disqualifier rule, regardless of how well it scored on every other dimension.
→Every score ships with a full per-criterion breakdown, so a fit tier is never a black box.

A ranked, explainable list, not just a yes/no filter.

Coach Card Generator

Three buying-committee personas (CRO, Head of RevOps, CFO) for our own RevOps Support offering.

Tailored cards

Interchangeable sections

→The CFO card leads with cost predictability and bounded scope. The CRO card leads with pipeline and rep-productivity impact. Same offering, translated for what each person actually asks about.
→Every card includes a real proof point pulled from an actual completed engagement, not a generic value-prop line.

Swap the headers between any two cards and they stop making sense, which is the actual test of whether a card is tailored.

AEO Agent

Audited this exact site (telemeterstrategy.com) against a synthetic unoptimized single-page app with no llms.txt, no schema, and a JS-only shell.

87/100

This site scored

25/100

Unoptimized SPA scored

Checks run

→Correctly gave this site full marks for AI crawler access, llms.txt quality, and raw-HTML content, the exact infrastructure built earlier in this same engagement.
→Correctly zeroed out the unoptimized comparison on llms.txt and structured data, and caught its raw HTML holding only 9 characters of real content behind an empty script shell.
→The one gap the audit honestly flagged on this site: only 2 of 6 valuable schema types on the homepage specifically, since FAQPage and Service schema live on other routes it didn't check.

The scoring differentiates a real site from a fake one instead of returning the same reassuring number regardless of input.

Brand Citation Scanner

Ran a real scan on our own name across review sites, press, podcasts, and communities, no synthetic data this time.

Categories checked

Citations confirmed

Known citation missed

→A broad search on the company name alone found nothing in press. A targeted follow-up search found a real, confirmed citation we already knew existed.
→Correctly reported zero results for review sites and podcasts rather than assuming a citation exists somewhere unseen.
→Did not surface a separate, independently-known citation, a video panel appearance, even with a fairly targeted search. We reported that gap honestly instead of hiding it.

A tool that admits what search cannot find is more useful than one that quietly assumes coverage it does not have.

GTM Initiative Audit

12 LinkedIn ads across 7 campaigns and 4 topics, checked against a stated strategy (planned topic ratio, budget, and audience guardrails).

Underperformers flagged

7 vs 3

Active campaigns vs max

3 topics

Ratio drift caught

→Correctly flagged 5 ads exceeding 1000 impressions at under 0.4% engagement, and correctly left low-impression ads unflagged since there wasn't enough data yet to judge them.
→Correctly excluded a paused ad from every calculation instead of letting it skew the active portfolio's numbers.
→Caught real messaging-ratio drift: one topic running 22.7 points over its planned share, two others running 15.9 points under, both past the threshold worth flagging.

Every flag traces back to a stated rule, not a gut call on which ads look tired.

Pipeline Velocity Monitor

A synthetic deal stage-history export: 5 deals, including one healthy re-qualification, one unexplained regression, two stalled deals, and one closed-won deal used to check the tool correctly excludes closed deals from open-pipeline flags. Extended separately with 3 deals testing the false-reactivation check: one that went quiet then resumed with no evidence, one that resumed with a real economic-buyer update, and one clean backward move used to confirm the new check never fires on a move already classified as a regression.

1/1

Healthy re-qualifications caught

1/1

Unexplained regressions caught

2/2

Stalled deals caught

1/1

False reactivation correctly caught

→Correctly classified a deal that moved from Proposal back to Discovery as healthy re-qualification because the economic buyer changed, and a separate deal that moved from Negotiation back to Evaluation as an unexplained regression because the buyer field stayed identical.
→Flagged both deals sitting past the 30-day stale threshold, one at 41 days, one at 36, while correctly leaving a deal at 21 days in its current stage unflagged.
→Left the closed-won deal out of every stalled and missing-deadline flag despite it having no decision deadline on record and its last stage entered 60 days ago, since closed deals aren't open-pipeline risk.
→Correctly flagged a deal that sat quiet for 50 days and then moved with no economic buyer or decision deadline attached, while correctly leaving a second deal unflagged that went quiet the same way but resumed with a real buyer update on record.

Zero false positives across a clean deal, a closed deal, and both regression types in the same test set. The deal that came back from silence with real evidence and the one that came back with nothing but a reply looked identical on the stalled-deal list, and only the false-reactivation check told them apart.

Attribution Builder

14 synthetic deals tagged with a primary trust signal (Referral, Customer Proof, no signal recorded, and a deliberately tiny Expert Content sample), run twice: once at the default 5-deal reliability threshold, once at 3.

75%

Referral win rate

25%

Untagged deal win rate

Deals missing a signal

→At the default threshold, correctly called every signal too small to trust yet, including a 75% win rate, rather than presenting a 4-deal sample as a real finding.
→At a lower threshold, correctly ranked Referral (75% win rate, 4 deals) above Expert Content (100% win rate, but only 1 decided deal), the exact case where a naive "best win rate wins" ranking would get it wrong.
→Cleanly separated the 4 untagged deals into their own list as a CRM hygiene gap rather than folding them into a misleading "Not recorded" performance score.

The tool's job is knowing when a number is too small to trust, not just computing the number.

Deal Handoff Friction Monitor

10 synthetic late-stage deals covering all four friction types plus clean baselines: an early-stage deal, a fast mover, and a closed deal with old dates used to check the tool correctly excludes closed deals from live checks.

1/1

Stalled deals caught

1/1

Missing signing authority caught

1/1

Broken promises caught

1/1

Stale buyer updates caught

→Correctly flagged a deal 20 days into Procurement against a 10-day threshold, while leaving a deal 2 days into Legal Review unflagged.
→Correctly matched a paraphrased onboarding brief entry ("priority support tier") against the original sales promise ("Priority support") without a false positive, then correctly caught a separate deal missing two full promises ("Free onboarding support for 60 days" and "Custom integration with Slack") from its brief.
→Left a deal still at Verbal Agreement out of every check since none of the late-stage gates apply yet, and excluded a Closed Won deal with months-old dates from both the stale-update and stalled checks despite the old timestamps.

Zero false positives across four distinct flag types, two clean late-stage baselines, and two exclusion edge cases in the same test set.

Enrichment Reliability Auditor

An 18-row synthetic enrichment log across 4 sources: one clean source, one genuinely noisy source with enough volume to judge, one stale-data source, and a deliberately small-sample source with a 100% override rate used to test that the sample-size guard actually holds.

1/1

High-override sources correctly flagged

1/1

Small-sample source correctly withheld

2/2

Stale fields caught

→Correctly flagged a source with a 50% override rate across 6 enrichments, and correctly left a source at 16.7% across the same sample size unflagged, both above the 5-enrichment minimum.
→Correctly withheld judgment on a source with a 100% override rate, the highest raw rate in the entire test set, because it only had 2 enrichments on record. The rate was still reported, just not flagged as a problem.
→Correctly separated missing-source-lineage fields from stale-but-sourced fields into two distinct lists rather than treating both as the same kind of gap.

The one source most likely to look alarming on a raw override-rate ranking was the one the tool correctly refused to judge.

Stage Advancement Auditor

6 synthetic deals covering justified and unjustified advances, a deal with zero buyer evidence across every forward move, a deal with a backward move mixed in to check it gets excluded, and a single-stage deal too early to judge.

5/5

Justified advances caught

4/4

Unjustified advances caught

1/1

Zero-evidence deals caught

→Correctly flagged 3 straight forward moves on one deal with no buyer evidence recorded at any of them, distinct from a deal with just one unjustified hop among otherwise-justified moves.
→Correctly excluded a backward stage move from every list entirely, neither justified nor unjustified, since judging forward-move legitimacy is a different question from the backward-regression check Pipeline Velocity Monitor already handles.
→Correctly left a single-stage deal with no transitions yet out of every list, since there is no advancement yet to judge.

One evidence signal was enough to clear a move as justified. Zero, even across a deal with real seller activity logged at every stage, was not.

Handoff Acceptance Monitor

19 synthetic handoffs across 4 owners and 3 handoff types: one owner with too little resolved volume to judge despite a low raw rate, one owner with enough volume and a genuinely low rate, one fully clean owner, and one handoff not yet due used to check it does not count against anyone.

8/8

Stalled unaccepted caught

1/1

Low-acceptance owner caught

1/1

Insufficient sample correctly withheld

→Correctly flagged an owner at a 33.3% acceptance rate across 6 resolved handoffs, while correctly withholding judgment on a different owner at a 25% rate, the lower of the two, because that owner only had 4 resolved handoffs on record.
→Correctly excluded a handoff routed hours earlier, well inside its 24-hour response window, from every stalled list and from the owner's acceptance-rate denominator entirely, rather than counting it as a miss before it was even due.
→Correctly separated a record with no owner assigned at all into its own list instead of folding it into the stalled-unaccepted count, a different failure mode from a routed-but-unacknowledged handoff.

The owner with the worse raw number was the one the tool correctly declined to judge.

ROI Proof Generator

A synthetic task run log across 6 automation batches, mixing measured and estimated basis, an unlabeled row to check the default, and one row with a deliberately implausible time-saved claim.

1,865

Tasks resolved

$2,200

Measured cost avoided

1.47x

ROI multiple (measured only)

→Correctly split hours and cost avoided into measured ($2,200 / 55 hours) versus estimated ($1,090 / 25 hours) buckets rather than blending them into one number.
→Correctly defaulted a row with no basis recorded to "estimated" rather than assuming it was confirmed.
→Correctly excluded a row claiming 600 minutes saved per task from every total, listing it separately for review instead of letting it inflate the receipt.

Every dollar in the headline ROI multiple traces back to a measured row. Nothing estimated or flagged made it into that number.

Call Scorecard Builder

Two synthetic Discovery-call transcripts scored against the same 5-behavior checklist, one call hitting most of the checklist, one call missing most of it.

4/5

Strong call score

2/5

Weak call score

6/6

Behaviors cited correctly

→Every observed behavior came back with the exact timestamp and quote where it happened, not just a checkbox.
→Correctly ignored buyer lines entirely when searching for rep behaviors, including a buyer line that used the word "budget" on the strong call, which did not get miscounted as the rep asking about budget, the citation still pointed to the rep's own earlier question.
→Correctly identified the same missing behavior, confirming decision process, on the strong call as the one gap worth coaching, instead of a vague overall score.

Two calls that might get the same subjective "felt fine" review scored differently, and differently, on specific, checkable behaviors.

Sequence Fatigue Monitor

10 synthetic contacts covering all three fatigue patterns plus clean baselines: a contact with only one touch (too early to judge), a paused sequence that correctly stopped after a stage change, and a contact triggering two flags at once to check they don't interfere with each other.

2/2

Dead sequences caught

3/3

Repeated-asset contacts caught

2/2

Stale automation caught

→Correctly computed a trailing unanswered count of 3 for a contact whose full history had 4 unanswered touches with one response buried earlier, proving the count is a trailing streak, not a lifetime tally.
→Correctly left a sequence unflagged after it was manually paused following a stage change, while flagging a near-identical sequence that kept firing after the same kind of stage change.
→Correctly flagged one contact for both a repeated asset and stale automation at once without either flag suppressing the other, and correctly left a single-touch contact unflagged as too early to judge.

Zero false positives across three distinct fatigue patterns, five clean baselines, and a compounding-flag edge case in the same test set.

Buyer Readiness Score

5 synthetic accounts covering all four combinations of engagement and readiness, including one boundary case scored exactly at the threshold on one axis.

2/2

Engagement without readiness caught

1/1

Genuinely hot caught

1/1

Fast movers caught

→Correctly flagged a heavily-engaged account (100/100 engagement) with no budget confirmed, no timeline, and only 1 of 5 objections resolved as engagement without readiness, exactly the pattern a blended score would have called "hot."
→Correctly classified a low-activity account with budget, timeline, and the competing alternative all confirmed as a fast mover, the shape a referral or prior-relationship deal actually takes.
→Handled a boundary case scored at exactly the engagement threshold and one point under the readiness threshold correctly, sorting on the right side of the line rather than rounding generously.

Two accounts that looked identically "engaged" on a single dashboard score turned out to need completely different next steps once readiness was scored separately.

Beachhead Segment Selector

4 synthetic candidate segments covering a clean strong fit, a genuinely weak fit on nearly every criterion, a segment scoring well on paper but under the 10-conversation validation minimum, and a second strong fit deliberately included to check the positioning-drift warning actually fires.

1/1

Under-validated segment correctly capped

1/1

Weak-fit segment correctly ranked last

Yes, 2 segments

Positioning-drift warning fired

→Correctly capped a segment at "needs more validation" despite a 77.0 raw score, higher than one of the two segments that actually qualified as beachhead-ready, because it only had 3 of the required 10 completed conversations.
→Correctly scored a long-cycle, low-relationship, low-pain segment at 24.3 and ranked it last, driven mainly by a 270-day average conversion time falling well outside the 3-6 month window the framework targets.
→Correctly flagged a positioning-drift warning when two segments both cleared the beachhead-ready threshold, naming both by name instead of silently picking the higher-scoring one and hiding the tie.

The segment with the better raw score on paper was not the one the tool called ready, because it had not yet been validated with enough real conversations to trust that score.

Pipeline Momentum Monitor

An 84-row synthetic stage-history export: a strong 10-week early period across two segments, then a recent 3-week window where Mid-Market held steady but Enterprise stalled and slipped backward, plus a thin-history case and a genuinely stable case run separately to confirm the tool never flags a trend that isn't there.

Yes

Deceleration correctly caught

Yes

Divergence correctly flagged

Enterprise

Declining segment isolated

→Correctly computed a blended forward-conversion rate of 87.9% across the full export, while the most recent 21-day window on its own had already fallen to 64.7%, the exact "headline still looks fine" pattern the tool exists to catch.
→Correctly isolated Enterprise as the segment losing momentum, a drop from 100% to 33.3% forward-conversion between its own prior and recent windows, while Mid-Market stayed healthy in the same period.
→On a separate thin-history export (3 transitions), correctly reported insufficient_history instead of fabricating a trend, and on a separate genuinely stable export, correctly reported no deceleration and no divergence.

The full-export number said conversion was fine. The most recent three weeks, on their own, said otherwise, and the tool caught the gap between the two instead of reporting only the reassuring one.

Expansion Signal Monitor

5 synthetic closed-won accounts covering a clean multi-signal expansion, a capacity-plus-volume expansion, a same-department volume-only case, a single-new-contact case, and an account closed too recently to judge.

2/2

Multi-signal expansion caught

1/1

Single-contact case correctly held back

1/1

Too-early account correctly excluded

→Correctly flagged one account expansion-ready on three clustered signals (3 new contacts, a VP joining, and 3 distinct new departments) and a second on two signals (new-contact volume plus 92% seat capacity), naming the actual evidence behind each.
→Correctly held an account back from expansion-ready status despite 3 new contacts joining, because all 3 landed in the same department and none were senior, only one signal type ever fired.
→Correctly reported no signal at all for an account with exactly one new contact since close, and correctly excluded an account closed only 5 days ago as too early to judge rather than scoring it "no signal."

Four accounts added new contacts since close. Only two of them actually looked like expansion once judged by more than headcount alone.

Churn Precursor Monitor

5 synthetic churned accounts used to calibrate three precursor patterns, plus a deliberately clean churned account to keep the hit rate honest, then applied forward to 4 active accounts covering a full cluster, a single signal, a genuinely healthy account, and a thin-history account.

3/3

Precursors validated

1/1

At-risk account correctly flagged

1/1

Single-signal account correctly held at "watching"

→Correctly validated all three precursor patterns at an 80% hit rate across 5 judgeable churned accounts, and correctly left one deliberately clean churned account out of the "fired" count instead of forcing a match.
→Correctly flagged one active account at-risk with all three validated precursors firing at once (champion silent 90 days, engagement down, distinct contacts contracted), naming each one individually.
→Correctly held a second active account at "watching" despite a real engagement decline, because only one of the three validated precursors fired, and correctly cleared two other active accounts showing no validated precursor at all.

The same three patterns that preceded four out of five past churns are now being tracked live on accounts that have not churned yet, and only one of four active accounts actually matches the full pattern.

Deal Runway Auditor

4 synthetic closed-won deals used to calibrate historical stage benchmarks, applied forward to 4 open deals covering a deal with real runway, a deal genuinely out of time, one missing a close date entirely, and one sitting in a stage with no closed-won history to benchmark against.

4/4

Stages benchmarked from real history

1/1

Out-of-runway deal correctly flagged

2/2

Missing-data cases correctly separated

→Correctly computed a 30-day historical median remaining time for deals reaching Proposal, then correctly flagged an open deal in Proposal with only 14 actual days left before its expected close date, a 16-day shortfall against its own stage's real history.
→Correctly cleared a Discovery-stage deal with 105 days left against that stage's 73.5-day historical median as having real runway.
→Correctly separated a deal missing an expected close date entirely from a deal sitting in a stage ("Legal Review") with no closed-won history to benchmark against, two different kinds of unknown, reported as two different statuses instead of one vague "can't tell."

The CRM close date said this deal was on track for the quarter. Its own stage's real closing history said it needed twice the time it actually had left.

Loss Reason Pattern Monitor

26 synthetic closed-lost deals: a prior 90-day window with an even spread across four loss reasons, a recent 90-day window where Pricing losses spike, plus a separate segment-clustering batch where Enterprise loses almost exclusively to Competitor.

Pricing, +41.8 pts

Rising reason correctly flagged

Enterprise / Competitor

Segment cluster correctly isolated

SMB

Unflagged segment correctly cleared

→Correctly computed the overall distribution across all 26 losses (Pricing 42.3%, Competitor 30.8%, Timing 15.4%, No Budget 11.5%) and correctly gated the trend comparison on both windows clearing the 5-loss minimum before trusting a percentage.
→Correctly flagged Pricing as rising, 60% of recent losses versus 18.2% of prior losses, a 41.8-point swing, once both the 10-loss recent window and the 11-loss prior window cleared the sample gate.
→Correctly isolated Enterprise as clustering hard on Competitor losses (83.3% of that segment's losses versus 30.8% book-wide, a 52.5-point gap on 5 qualifying losses), while correctly leaving the SMB segment unflagged since its mix never cleared the clustering threshold.

The book-wide loss reason looked like Pricing. One specific segment was actually losing almost exclusively to a named competitor, a pattern the overall distribution alone would have buried.

Win Concentration Auditor

Two synthetic closed-won quarters, one built from 8 deals where a single rep and two enterprise contracts drove almost all of the total, one built from 10 deals spread evenly across reps and segments, plus a separate open-pipeline export run against a quota to test the coverage decomposition.

Top 3 deals, 93.7%

Concentrated quarter correctly flagged

Top 3 deals, 31.2%

Distributed quarter correctly cleared

3.67x to 1.63x

Coverage ratio correctly adjusted

→Correctly flagged the concentrated quarter on both thresholds at once, its top 3 deals were 93.7% of bookings against a 40% threshold, and one rep alone was 70.6% of bookings against a 35% threshold, while correctly leaving the distributed quarter unflagged at 31.2% and 20.6% on the same two checks.
→Correctly computed that the concentrated quarter's bookings reached only 2 of 3 present segments at a meaningful share, Enterprise and Mid-Market, while SMB never cleared the 10% contribution bar despite 5 deals landing there.
→Correctly decomposed a raw 3.67x pipeline coverage ratio down to 1.63x once rep-created pipeline (47.3% of the open total) and pipeline stale past 180 days (55.5% of the open total) were excluded, the same shape as a headline coverage number collapsing once it is actually broken apart.

The concentrated quarter and the distributed quarter had similar deal counts and similar totals. Only one of them would have survived losing its top rep.

Narrative Asset Gap Mapper

7 synthetic closed-won call transcripts seeded with three recurring narratives at different frequencies, one one-off line planted to check it does not get mistaken for a pattern, run against a 6-asset content inventory with deliberately uneven coverage across those narratives.

3/3

Recurring narratives correctly surfaced

1/1

One-off line correctly excluded

Yes

Lower-frequency, bigger-gap narrative ranked above higher-frequency one

→Correctly surfaced all three planted narratives at their true frequencies, fast deployment at 5 of 7 calls, dedicated CSM at 4 of 7, and no-rip-and-replace integration at 2 of 7, while correctly excluding a one-off line about the founder's background that appeared in only 1 of 7 calls from the recurring-narrative list entirely.
→Correctly read asset descriptions rather than titles alone, marking the existing support one-pager "Needs Work" against the dedicated-CSM narrative because its own description noted it predated the current CSM model, instead of crediting it as coverage just because a support asset existed.
→Correctly prioritized the integration narrative above the CSM narrative despite its lower raw frequency (2 of 7 versus 4 of 7), because it was fully missing from the content inventory rather than just weakly covered, and two separate deals cited it as "the deciding factor," the kind of judgment call a frequency-only ranking would have gotten backwards.

The narrative that showed up in the most calls already had the most content built for it. The narrative that actually closed two deals on its own had nothing written about it at all.

Resource Contention Auditor

10 synthetic open deals seeded with one solutions engineer required by three deals inside a 2-day window, one legal reviewer required by two deals close together and a third far apart, one executive sponsor required by two deals spread months apart as a clean control, plus two deals missing required fields and one deal already past its own last-feasible date.

1/1

Tight 3-deal contention cluster caught

1/1

Clean non-contending resource correctly cleared

1/1

Past-feasible-date deal caught

→Correctly clustered three deals all requiring the same solutions engineer within a 2-day window (Nov 10, 11, 12) into one contention flag, while correctly splitting a legal reviewer's four deals into one 2-deal cluster 2 days apart and one deal left alone 29 days out, rather than treating every deal sharing a resource as one undifferentiated conflict.
→Correctly left an executive sponsor required by two deals nearly two months apart entirely unflagged, confirming the buffer window doesn't manufacture contention out of a resource simply being reused across the pipeline.
→Correctly separated three different problems into three different lists: a deal missing its required-resource field, a deal missing its milestone date, and a deal whose milestone date already fell 10 days past its own last-feasible date, instead of collapsing them into one generic "data issue" flag.

Three deals looked independently healthy on every normal measure. All three needed the same solutions engineer inside a two-day window nobody had checked for.

Qualified CPL Auditor

160 synthetic leads across three channels with known spend: one channel seeded to look cheap blended but mostly anonymous or unqualified, one channel priced consistently on both measures as a clean control, and one channel with real spend but zero qualified leads. A separate 4-lead batch tested the ICP-fallback scorer directly, industry and employee count against an ICP definition, with one lead deliberately high-scoring but anonymous.

10x gap ratio

Hidden-cost channel correctly flagged

1.25x gap ratio

Clean channel correctly left unflagged

1/1

Zero-qualified channel correctly isolated

→Correctly computed a 100 blended CPL on the seeded channel while its Known Qualified CPL came back at 1000, a 10x gap, against a clean control channel whose blended and qualified CPL landed at 200 and 250, a 1.25x gap correctly left unflagged at the 2x threshold.
→Correctly isolated a third channel with real spend and 20 total leads but zero that were both identified and ICP-qualified, reporting it separately from the gap-ratio flags rather than folding a total-failure channel into the same list as a partially-inflated one.
→Correctly excluded a lead from the ICP-fallback test that scored a perfect 100 on industry and employee count but was marked not identified, confirming the identified gate is enforced independently of ICP fit rather than letting a high score alone count as qualified.

The cheapest-looking channel in the spend report was actually the most expensive channel once anonymous and unqualified leads were removed from its denominator.

AI Workflow Rework Auditor

A synthetic 42-row run log across six AI-assisted workflows with known time costs: one workflow whose drafts are quick to generate but slow to correct, one clean control, one dominated by manual copying into another system, one that works on one input type and fails on another, and one with too few runs to trust. Three malformed rows were added to test exclusion.

28 min vs 25 min baseline

Net-negative workflow correctly stopped

70% vs -40% by input type

Split workflow correctly narrowed

3/3

Malformed rows excluded

→Correctly caught a workflow claiming 20 minutes saved per task that actually took 28 minutes all-in against a 25 minute baseline, and returned a stop verdict, while a clean control workflow saving 13 minutes per task returned a keep.
→Correctly separated a lead-research workflow into one input type saving 70% of baseline and another losing 40%, returned a narrow verdict naming which to keep and which to pull, rather than averaging the two into a misleading middle.
→Correctly identified manual transfer as the dominant cost in a workflow that saved only 1.5 of 10 baseline minutes, flagged a correction repeated across all three reviewers as a source-data problem, and excluded three malformed rows rather than counting them.

The portfolio implied 108.9 hours saved on a gross basis and delivered 26.1 hours once review, correction, and handoff time were counted.

Why publish Beta tool results instead of only finished ones?

Because the proof is in what a tool actually produces, not its status label. Real output at every stage says more than a polished demo of a finished one.

Is this test data or a live client's data?

Test data, seeded with known scenarios so the output can be checked against a known-correct answer before a tool ever touches a live CRM.

Can I see a tool run against my own CRM data?

Get in touch and we'll walk through it live against a sample of your data.