Skip to content

CodedVibesX/trustline

Repository files navigation

Trustline

A cite-or-abstain consistency gate for startup data. It joins startup records, cross-checks each funding line against primary sources, flags where the same record contradicts itself, and abstains when it cannot get two independent sources. Built and run live on Harmonic's own public Hot 25.

The catch: DualEntry's Series A dated two different ways on Harmonic's own pages

The catch (real, in their own published data)

Harmonic's Q1-2026 Hot 25 dates DualEntry's $90M Series A as Mar 2025. Their Q2-2026 Hot 25 dates the same round as Oct 2025. The round was announced October 2, 2025, when DualEntry launched from stealth, and the company was founded in 2024, so a Series A in March 2025 is not possible. Two of Harmonic's own published pages disagree on one round, and a primary-source check settles it. Trustline finds this automatically by joining the two snapshots; it is not hard-coded.

Primary sources (all date the round to Oct 2 2025): DualEntry, PR Newswire, Built In NYC, Accounting Today.

Timestamped evidence (captured 2026-06-24, in case the page is later corrected, which would be the point of the gate): Q1 Hot 25 archive, Q2 Hot 25 archive.

What it produced

Run over both snapshots, the gate classifies every company on both lists and asserts a conflict only with two or more independent sources:

Company Verdict Why
DualEntry CONFLICT same $90M Series A dated Mar 2025 (Q1) vs Oct 2025 (Q2); 4 sources resolve it to Oct 2 2025
Dust UNVERIFIED (abstain) same round listed as $16M (Q1) vs $15M (Q2); under two sources to resolve, so it does not assert
Cogent Security PROGRESSION $11M Seed to $42M Series A is a real new round, correctly not flagged
5 others CONSISTENT identical across both snapshots

The scorecard, in Harmonic's brand

The restraint is the point. Two real discrepancies surfaced (DualEntry and Dust), and the gate only asserted the one it could fully source. It cleared a legitimate new round that naive diffing would false-positive. Wrong-but-confident is the failure mode a discovery platform cannot ship, and it is the one this avoids.

What is real and what is a proxy

  • Real: the snapshots are Harmonic's actual public Hot 25 pages, committed under data/raw/. The ingest, the entity join, and the consistency classification are fully deterministic and re-run with npm run audit. The DualEntry conflict and the Dust abstention fall out of the data automatically. The primary sources are real and linked.
  • A proxy, stated plainly: the quarterly Hot 25 is curated editorial, and the live product is behind a paid login that I do not have. The Hot 25 is a cheap public stand-in for the entity-and-field consistency a production discovery system must guard at 30M-company scale. This is not a claim about Harmonic's live data quality; it is a demonstration of the gate on the only public surface available.

Run it

npm install
npm run ingest        # parse the committed Hot 25 snapshots into typed records
npm run audit         # join, classify, cite-or-abstain -> data/report.json
npm run audit:check   # fail if the real catch disappears
npm run dev           # the scorecard at http://localhost:3001
npm run build         # static export (prebuild re-runs ingest + audit)

How it works

ingest    parse Harmonic's public Hot 25 (cheerio), only the fields the page exposes
join      match the same company across two quarterly snapshots
classify  same round + different value -> conflict candidate; new round -> progression
gate      assert CONFLICT only with 2+ independent sources; otherwise abstain (UNVERIFIED)
render    a scorecard in Harmonic's brand, every flag expandable to its sources

For a discovery team

The same field gate wires to an enrichment write path: before a record updates, check the new value against its prior snapshot and against independent sources, and route low-confidence or conflicting writes to review instead of asserting them. That is the difference between a database that is trusted and one that is turned off, and it is the layer the Staff Search and Discovery roles own.

Layout

engine/   ingest (cheerio), audit (the gate), schema
data/     raw Hot 25 snapshots, parsed records, primary-source evidence, report.json
app/      the Next.js scorecard, in Harmonic's brand
visual/   the self-contained catch card
scripts/  shoot.mjs, shoot-card.mjs (headless-Chrome capture)

Built by Lawrence Wolters. Public sources only, abstains under uncertainty. MIT licensed.

About

A cite-or-abstain consistency gate for startup data. Run live on Harmonic's own public Hot 25, it found a real self-contradiction (DualEntry's Series A dated two different ways across their own pages), cites the sources, and abstains when it cannot get two.

Resources

License

Stars

0 stars

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages