# Undertow method (`undertow-1.0`)

## Purpose
Undertow records influence operations that platforms and governments have publicly disrupted or named: who
published the takedown, when, where the publisher says the operators worked, whom they targeted, at what scale, and
which DISARM techniques the publisher's own description shows. It is the suite's influence-operations layer.
**Operation level only.** It never stores indicator values, account handles, URLs or domains, or the names of
individuals. It states no probability.

## Lanes (first live slice, 14 Nov 2026)

| Lane | Source | Read how | Cadence | Kept |
|---|---|---|---|---|
| `meta` | Meta threat-research repository on GitHub (MIT) | README table and each network's Markdown file; one GitHub API call for the dates files were added | Twice a year | Title, period, origin, targets, Meta's counts, Meta's network summary (scrubbed, quoted under MIT), the indicator **count** by type |
| `google` | Google Influence Operations Bulletin | Quarterly pages found from blog.google's security RSS and the URL pattern, Q1 2026 onward | Quarterly | Month, counts of channels, blogs, accounts and domains, the linked-to country, languages, countries named in the stance. No bulletin text. Unparsed entries kept and counted |
| `openai` | OpenAI news RSS feed only | `openai.com/news/rss.xml`; report pages never fetched (OpenAI's terms bar programmatic extraction) | Irregular | Operation name, origin wording, the one-line summary (scrubbed, quoted with credit) |
| `tiktok` | TikTok's reports to the EU Code of Conduct on Disinformation | Transparency Centre JSON, `disinfocode.eu/reports/tiktok/<n>/json`; never tiktok.com (TikTok's terms bar automated extraction) | Semiannual | Origin, target audience, accounts, followers, the month where the elections chapter gives it, TikTok's description (scrubbed) |
| `eeas` | EEAS 4th report on FIMI threats (March 2026) | Named Information Manipulation Sets entered once into `config/register/eeas-fimi-4.json`, each row cited to its printed page; the FIMI page is read each run for a newer report | Annual | Name, aliases EEAS gives, attribution wording, operators EEAS names, alignment, targets, counts, short behaviour quotes with pages |

## What an item is
One item per disclosed network or operation: kind `observation`, tier `routine`, `status_class: reported`, source
grade **A2** (the publisher is reliable on its own enforcement actions; the takedown is probably true and cannot be
confirmed from outside). No anomaly rule yet: there is no baseline of disclosures.

## Three questions kept apart
1. **Did the takedown happen?** The publisher's own report of its own action: A2, `reported`.
2. **Where were the operators?** The publisher's judgment, kept verbatim (`origin_wording`) with the publisher's
   stated level (`origin_level_stated`: high, moderate or none; see `confidence-mapping.md`).
3. **Who directed it?** Recorded only when a publisher names an operator (`sponsor`, as reported actors).
   **Origin is never read as sponsorship.** "Aligned with Russia" (EEAS on Storm-1516 and Operation Overload, which
   it calls non-attributed) is recorded as alignment, never as origin.

Our confidence follows tic-confidence-1.0 and is about question 2 only: see `confidence-mapping.md`.

## Text guards (`undertow/text.py`)
Every sentence kept from a publisher passes `scrub()`: URLs, e-mail addresses, @handles, defanged values and bare
domains become "[link withheld]"; a run of two or more capitalised words that are not known places,
organisations, platforms, operation names or ordinary words, or one after a title ("Prime Minister X"), becomes
"[name withheld]". The rule leans towards withholding. A single surname with no title before it is not caught; that
is assumption UT-ASM-4, marked strained. The count of withheld spans is on every item. The build's `check` scans
every published file for defanged values, handles and URLs outside the publishers' own disclosure pages.

## DISARM (`config/disarm_rules.json`)
DISARM Framework v1.6.1 (CC BY-SA 4.0). A technique is attached only when a rule in the published table matches the
publisher's own words; the matched phrase is kept (for Google, not kept: facts only). `basis: mapped` means it is
Undertow's reading, not the publisher's claim; `basis: reported` is for a publisher that cites an id for one
operation (none in this slice). A behaviour no rule matches stays as text. **The rule table is proposed and not yet
reviewed by Tom**; every item says so (`disarm_rules_status`).

## Outputs
- `data/feed.geojson`: tic-feed/2. Geometry on the target countries the publisher names; Google and OpenAI name
  none, so their geometry is on the countries named in the stance or summary (`countries_named`), and null when none
  is named.
- `data/stix.json`: a STIX 2.1 bundle with deterministic UUIDv5 ids. Report per disclosure, campaign per network,
  location with `originates-from` and `targets`, DISARM's own attack-pattern objects (unchanged, with DISARM's CC
  BY-SA marking) with `uses`, intrusion-set with `attributed-to` for named operations, TLP:CLEAR and a statement
  marking per source licence. Never: indicator, observed-data, cyber-observables, an identity for a person.
- `data/events.json`: expected next disclosures for Verdict's calendar, on each publisher's past pattern. Dates
  only, no probabilities; none is a published schedule.

## What the method does not do
- It does not judge whether an operation worked, how far it reached, or who sponsored it.
- It does not read OpenAI's report pages or tiktok.com, so OpenAI's Breakout Scale and TikTok's monthly list are
  missing. Permission emails to both are drafted for Tom.
- TikTok's Code of Conduct reports cover only networks that touched Europe or a crisis, 3 to 8 months behind TikTok's
  own monthly list.
- Google names no target; the countries its stance names are not targets.