Find what matters most: Outcome-Driven Innovation
Break a job into steps, score each by how important it is and how badly it is served today, then fix the biggest gap first.
This page could not load its data. Reload; if it persists, the demo is down.
Signal Engine finds the credit unions that have a reason to talk, points you to the right people, and learns which first message earns a reply.
The public NCUA call report for lists credit unions. No team can write a thoughtful message to all of them, so the first move is to narrow the list.
This data is industry-wide. It covers every credit union that files a call report, not one institution.
What this does not show. These filters show a research method. They do not establish that a credit union wants to buy anything. The pre-registered test of whether signals predict later events was run and failed: they describe, they do not predict (see Proof panel, P1).
Three public facts, each with a cutoff, narrow the list. The cutoffs are screening choices made for this demo. They are not benchmarks and nothing here says they are the right ones.
The first panel below is still industry-level. The next panel zooms to one institution.
Each count is observed. Each cutoff () is illustrative.
We picked one from the to show what a single account looks like. It is an example, not a typical case.
Name withheld. One record from the signal set, shown as an example. It is not a typical or average case.
Inferred Likely buyer role: Chief Lending Officer (inferred from the pain, not observed)
A named illustrative person appears only at the later people-research step, as an example of what research produces, with its public sources.
Compared at the displayed 3-decimal precision; at full precision the last two counts are and .
A third-party analysis reports mergers announced for 2026 to date (through ), against for all of 2025 ().
NCUA's own merger activity report lags: newest posted file is ().
A merger is a change signal, not evidence anyone wants to buy anything.
What this does not show. These filters show a research method. They do not establish that a credit union wants to buy anything. The pre-registered test of whether signals predict later events was run and failed: they describe, they do not predict (see Proof panel, P1).
Each observed item about Credit union A (name withheld) is checked against the proposed offer ladder. Most do not match, and the page says so.
This page could not load its data. Reload; if it persists, the demo is down.
What this does not show. The delinquency trend is observed for Credit union A; it does not show that this or any credit union treats it as a pain, or wants the offer. The match column is our judgment, labelled Inferred.
Roles only. No person is named, and every role below is an inference to validate.
This page could not load its data. Reload; if it persists, the demo is down.
Who we would write to, what public sources could tell us, which channels the agent may use, and the checks a draft must pass before a human sees it.
This page could not load its data. Reload; if it persists, the demo is down.
Dana Ruiz, Chief Lending Officer Composite, not a real person
Illustrative. Not any named individual at any credit union.
LinkedIn: Manual, by human. The agent never touches LinkedIn.
What this does not show. It does not show that any real person fits this card, that any channel works, or that LoopInHuman meets any rule. Every row is a design choice or a third-party rule with its source.
An opener and two follow-ups for the composite persona. Every fact has a footnote; every gap is a labelled placeholder.
This page could not load its data. Reload; if it persists, the demo is down.
Composite, not a real person. Illustrative template. No message is sent. To: Dana Ruiz, Chief Lending Officer.
Sending is not part of this demo. Approve, Edit and Reject are previews that change this page only. Nothing is sent and nothing is stored.
Everything after this send is simulated.
Learning Lab · Act 1 of 3 · Find the bottleneck
Where this comes from: step 6, the draft
The loop looks at every stage of the pipeline, not just replies, and asks one question: which stage, if improved, would add the most meetings per month? It then proposes exactly one fix to test.
Synthetic scenario · each rate comes with a 90% credible range showing how sure we are
Not campaign results. This scenario plants the problem at one stage on purpose, so we can check whether the loop finds it (proof P2). Change the seed or the planted stage and run again; the same seed always gives the same funnel.
A demo seed, not a proof seed. 1 to 4,294,967,295.
The default matches the story below. Plant it elsewhere to see whether the loop still finds it.
| Rank | Stage | Score | Reference rate |
|---|
Eight steps of the operator's own job (Ulwick job map). Each step is scored for how much it matters and how well it is served today.
SIMULATED values; the BD operator scores his own steps later. Every importance, satisfaction and opportunity value in this table is simulated, never the operator's own rating. The importance and satisfaction inputs are illustrative numbers fixed in the demo; the opportunity column is computed only by the opportunity formula.
| Job step | In plain words | ImportanceSIMULATED | SatisfactionSIMULATED | OpportunitySIMULATED |
|---|
SIMULATED values; the BD operator scores his own steps later. The BD operator replaces them with his own ratings in a later step. Scoring one operator's own job is a prioritisation aid, not market evidence.
Credit union A (name withheld) · assets
This example's 60+ day delinquency rate (all loans)
Likely buyer role: Chief Lending Officer Inferred from the pain, not observed. No person is named here; a named example appears only at the people-research step, labelled as an example with its public sources.
| Quarter | 60+ day delinquency, % |
|---|
The writing-to-the-wrong-person problem in this Lab is a planted scenario, not a finding about this credit union.
Learning Lab · Act 2 of 3 · Run the experiment
Pick how many contacts you have and how big the true lift is. The page shows what that setup can and cannot tell you, next to a control run where nothing changed.
Champion = CIO-first. Challenger = CLO-first. Accounts split between them at random.
Range 100 to 500 in total, split between the two groups.
1x means no real difference. Act 3 uses the 2x case.
The champion's true rate in this simulation.
Our rule always looks monthly, at most 3 times.
A demo seed. The A/A control uses this seed plus one.
Champion vs challenger posteriors from the seeded run, with the A/A control beside it. Both use the same setup; only the challenger's true rate differs.
The A/A control shows how often noise alone looks like a win. Expected loss is the reply-rate points we would give up if the challenger turns out worse.
| Run | Group | Replies / contacts | Posterior mean | 90% credible range |
|---|
Rule written and dated 2026-11-19, before any scoring: promote only if the chance the challenger is better is above 99%, the expected loss is at most 0.05 points, there are at least 150 contacts per group in total at that look (counted cumulatively), and looks are monthly for at most 3 months. Expected loss = how many reply-rate points we would give up if the challenger turns out worse.
| Month | Champion | Challenger | P(win) | Expected loss (pt) | Rule says |
|---|
Sample-size calculator · alpha 0.05, power 0.80, two-sided
Check, computed live here: per group.
Fixed reference table from the published proof results. It does not change with the sliders.
| True lift | Promoted | Runs |
|---|
When nothing really changed, how often does each checking habit declare a false win?
False wins when nothing changed. The official figures come from the published peeking run and proof P3; the live line re-simulates only the naive weekly-peeking habit at your settings with demo seeds.
Learning Lab · Act 3 of 3 · The human gate
No real-world conversion result exists yet. The rule can recommend; only the human gate can promote a challenger. Every decision is written to a ledger designed to be append-only and hash-chained (this preview keeps entries only until the page reloads).
LEDGER (append-only within this page session, hash-chained SHA-256) · PREVIEW ONLY: nothing is written anywhere and no evidence exists. Entries below are added, never replaced; they last until you reload the page.
No preview entries yet.
Expected meetings per month, champion vs promoted challenger, with intervals
This is a simulation of a planted 2x lift. It shows how the page will look. It does not say real outreach improved; no real outreach data exists yet.
An optimizer will over-fit to whatever it is told to maximize. The person checks the measure is still the right one (Goodhart guard) and owns the risk.
Trust · How it works and why you can trust it
Business view says what each method is for. Technical view adds the formula. Every card links to its source and says whether it is a published method or our proposed use of it. None of them proves our system works; that is the proof panel's job.
Break a job into steps, score each by how important it is and how badly it is served today, then fix the biggest gap first.
Pick one main measure up front, check the split came out as planned, run a no-change control, and give tests time.
Primary metric (OEC) · SRM check · A/A test · at least 2 weeks. We borrow the habits, not their scale.
Ordinary p-values break if you keep checking and stop when it looks good. Always-valid methods stay honest whenever you stop.
Instead of one yes/no p-value, ask how likely the challenger is better and how much we lose if we are wrong. Stopping early still needs a pre-set rule.
Bandits send more traffic to the apparent winner, but that makes the measured effect less trustworthy. We keep them as an option, not the default.
A method where the model reflects on what failed and rewrites prompts. We use it only to draft challengers in the background; a person and a test decide.
Reflective prompt evolution (arXiv 2507.19457), run offline against an eval set plus shadow review. Never live.
The current approach stays in charge until a challenger earns its place, and a person signs off. This copies a pattern from bank model risk. We do not claim to meet any regulation.
Pattern borrowed from model-risk practice. SR 26-2 supersedes SR 11-7 and puts generative/agentic AI out of scope. We call this a pattern, not compliance.
Replies and losses get tagged by what pushed or held the buyer back: the pain, the pull of something new, habit, and worry.
Moesta four forces: push, pull, habit, anxiety. Used as reply and loss tags, not as a score.
When a number becomes the target, people game it. So the main measure is positive replies that lead to meetings, not raw replies.
Trust · Proof panel
Each block shows the bar, the result with its interval and a plain PASS or FAIL (P4 grades one case; its other lift is reported, not graded). Every figure below is read from the stored results when the page loads; if a result is unavailable the card says so and shows no number.
Held-out AUC
interval
AUC = vs ≥
Planted stage named
of runs · interval
Hit rate = vs ≥
| Planted location | Stage | Hits | Runs | Hit rate |
|---|
False promotion
of runs · interval
False promotion = vs ≤
Official run
of runs · interval
Promoted = vs ≥
of fresh runs · interval
Promoted = vs ≥
Pre-registered separately, reported next to the official result, never replacing it. Bar fixed before the run output was recorded.
| True lift | Status | Promoted | Interval | Runs | Verdict mix (promote / keep / not yet known) | Month promoted (first / second / third look; promoted runs only) |
|---|
False wins when nothing changed, from the published peeking output.
Reproducible: the full method and seeds are published with the proof (available on request).
Tuning seeds, before the official run. It shows why the rule text was amended. Official results are in the blocks above.
This page could not load its data. Reload; if it persists, the demo is down.
A merger is a proxy for a change event, not for buying intent.
SIMULATED
SIMULATED
The loop named the planted stage in of runs (bar , interval ). A planted answer in a simulation, not field data.
SIMULATED
The gate promoted noise in of runs (bar at most , interval ). One base rate and one volume only.