Tools · Personal research tool

Undertow

An experiment in hearing a shift in public language before it becomes the headline. We plan to test it in the open. It hasn’t earned your trust yet.

ExperimentUnprovenOctober 12 target · start unverifiedPersonal research tool

October 12, 2026 was the target for first unscored alerts; the start is unverified here. October 19 remains the recorded target for scored alerts, not a confirmed milestone. No forward-test results are published here.

Request a walkthrough
Read the method and scorecard

What it is

Undertow is designed to look for shifts in public writing: new phrases, changes in tone, and the same idea appearing in places that don’t usually agree.

The bet is that some shifts may appear weeks before a topic reaches the front page. That’s a hypothesis, not a demonstrated capability. The proposed test is how we find out.

Why build it in public

A useful test needs more than a list of wins. The plan is to publish confirmed flags, misses and false alarms. Core test rules stay fixed, and permitted human-review changes are recorded.

If Undertow turns out to be noise, the verdict should say so. It is an experiment, not a proven forecasting system.

Intended flow, not results

How it works

  1. Step 1

    Selected public sources

    A selected mix of public feeds and open data. Permissions and retention are under continuing review.

  2. Step 2

    Source collection

    Public items from the selected sources are collected for analysis.

  3. Step 3

    Language-shift detection

    Designed to look for new and rising phrases and ideas spreading across sources.

  4. Step 4

    Weekly flags

    Flags are recorded weekly with the time they were raised.

  5. Step 5

    Scored outcome review

    Each scored alert gets a sixty-day outcome window before it is counted.

Around every step

Core test rules stay fixed. Permitted human-review changes are logged, so they can be told apart from what the frozen rules flagged on their own.

Historical plan: October 12, 2026 was the target for first unscored alerts; the start is unverified here. October 19 remains the target for scored alerts. This diagram shows the design, not results.

Real pixels, historical view

Inside Undertow

Undertow runs privately. This is a trimmed view from a run before the forward test: newly emerging topic clusters in three of its communities, with two weeks of mentions per day. It is not a forward-test result.

Undertow's private app header from the October 5, 2026 run, with its section tabs.Newest emerging clusters by community on October 5, 2026: Tech, Finance and Health columns, each listing three topic clusters with two-week mention sparklines and source counts. A side note is blurred.
Private dashboard snapshot, October 5, 2026 (sensitive details trimmed). Shown as it was on that date; counts are from that run only. Only Tech, Finance and Health appear in this view.Scroll sideways to read the full image.

The proposed test

How we’ll judge it

  1. 01

    Rules frozen first

    Core test rules stay fixed. Permitted human-review changes are recorded. The fingerprint below identifies the recorded rules definition, not proof that the test has run.

  2. 02

    Weekly, timestamped flags

    Record each flag and when it was raised. October 12, 2026 was the target for first unscored alerts; the start is unverified here. October 19 remains the recorded target for scored alerts.

  3. 03

    Checked against reality

    Compare flags with later observations after their outcomes have time to mature. How those observations are linked to a flag is part of the method still under review.

  4. 04

    Results before a verdict

    A proposed twelve-week observation period is followed by outcome review. Each alert has a sixty-day outcome window, so a final verdict cannot be promised at week twelve. The results will inform whether to keep it, change it or shelve it.

Illustration

What a flag looks like

A simplified picture of the idea: a term rising clearly above its own usual range. The frozen test rules are more specific than this sketch.

Example dataEXAMPLE DATA: illustrative weekly mentions for a made-up termInvented values. A line stays inside a shaded usual range for ten weeks, then climbs above it from week eleven. Not a real Undertow result.EXAMPLE DATA0255075Usual rangeFirst week above rangeW1W5W9W14Example term, mentions per week (invented)
Illustration only. Not a real Undertow result.

Results must earn the verdict

Scorecard

Planned updates · monthly
Weeks complete of 12
—
Early flags confirmed
—
Misses
—
False alarms
—
Unscored alertsScored alerts, not yet measuredCounted later as: early flag confirmed · miss · false alarm
Historical target dates, with the test start unverified here. Proposed weekly cohorts, all empty. Bars stay empty until outcomes are reviewed. Each scored alert needs sixty days to mature, so later weeks fill in well after week twelve. Updates are planned monthly.

No forward-test results are published here. The test start is unverified here. Dashes mean unmeasured, not zero. The twelve-week observation window is proposed; counts will appear as reviewed outcomes become available, misses and false alarms included.

Recorded rules fingerprint

695a8b6ee78aae791aa3d7e899e422e7d53c55c77e1637de2b5e783456af6ac5

This identifies the rules definition. It does not establish a deployed test or a completed result.

Last scorecard update: no results published

Plans, not completed milestones

Roadmap

  1. Fall 2026

    Foundation

    Checks continue

    Foundation in place; reliability and recovery checks continue. Source permissions and retention remain under review.

  2. Fall 2026

    Forward test

    Start unverified

    A proposed twelve-week observation period under fixed core rules. The historical October 12 target does not establish a completed start. Record permitted human-review changes and allow outcomes time to mature.

  3. Winter 2026

    Deeper data, smarter reading

    Planned

    Explore broader coverage and ways to read tone and convergence alongside the test, without replacing its frozen rules.

  4. Winter 2026–27

    Public verdict

    Contingent decision gate

    Publish the scorecard and assess the results after outcomes mature. Timing depends on the test and the unresolved method checks; a useful result has to survive the misses and false alarms too.

  5. Later, if earned

    Wider use

    Conditional

    Consider a public tool or a regular brief only if the results justify it. There is no product for sale or release commitment.

Sources

A selected mix of public feeds and open data is the intended input.

Source permissions and retention are under continuing review. No verified source count or blanket terms-clearance claim is published here.

What it is not

  • Not financial, investment or legal advice.
  • Not a trading signal or a crystal ball.
  • Not a product for sale. It is a personal research experiment.

Other tools

Beta

RADAR

AI platform changes, explained, with corrections

Explore RADAR Public alpha

CCP Study Lab

Learn to investigate a security claim

Explore CCP Study Lab Prototype

PROBE

Adversarial AI testing, step by step

Explore PROBE

Discuss the evidence you need.

The same discipline behind these tools, applied to your systems.

Start an advisory conversationAll tools →