Skip to content

CIB Mango Tree

FreeInstitutional

Publisher: github.com/civictechdc/cib-mango-tree | Type: Non-detector | Cost: Free

TL;DR

A free, MIT-licensed Python toolkit that runs Coordinated Inauthentic Behaviour tests over a fact-checker's own social-media dataset, with a "Copy Pasta Test" that surfaces near-identical synchronous posts across accounts. The open-source primary entry of the institutional CIB ladder.

What it does

CIB Mango Tree is an interactive Python terminal toolkit that ingests user-supplied social-media data and runs a sequence of behavioural tests against it. The most documented test is the Copy Pasta Test, which clusters near-identical text posts that appear across multiple accounts within tight time windows: the canonical signal of a coordinated amplification operation. Outputs are pattern files: lists of accounts, post clusters, and time windows that meet the test's similarity and synchrony thresholds. The tool does not pull from a platform API; the analyst is responsible for assembling the input dataset, which means the analysis runs against whatever source the analyst can access (Meta Content Library extracts for IFCN partners, Twitter API archives where still accessible, Telegram exports, custom scrapes for fringe platforms).

The toolkit's value sits inside Architectural Anchor 1's behaviour pillar: Mango Tree produces a non-detector behavioural-pattern signal that combines with claim-extraction (2B.2) and content provenance (1A.4) to support an institutional finding. It is not a detector and never produces an "AI-generated" verdict; it produces a "these N accounts posted the same string within M minutes" verdict that a human analyst then interprets in context.

When to use it

  • A regional research lab or fact-check newsroom has assembled a dataset of suspect posts (Meta Content Library, Telegram export, Twitter archive) and needs to test whether the accounts behind those posts show coordination signatures.
  • An election-cycle CIB investigation in Indonesia, the Philippines, or Thailand needs an open-source analysis layer that runs locally on the analyst's machine without sending data to a vendor cloud.
  • A research partner is training a regional newsroom in CIB methodology and wants a free, MIT-licensed tool that participants can install and reuse on their own datasets after the workshop.
  • An institutional response to a synthetic-content surge needs to test whether the amplification network behind the synthetic content shows coordination signatures separable from organic spread.

Limitations

Limitations

  • Requires Python proficiency; the interactive terminal is not a no-code tool.
  • Behavioural focus, not content: Mango Tree analyses posting patterns, not whether the underlying content is true, false, or AI-generated.
  • Input dataset is the analyst's responsibility. A Mango Tree run is only as representative as the data the analyst supplies; no Mango Tree result vouches for the completeness of the input.
  • No published SEA-specific deployment is documented in the toolkit's source base: the tool is regionally relevant in principle but the worked-example casebook for Indonesia, Malaysia, the Philippines, Thailand, Sri Lanka, or Laos has not yet been published.
  • Coordination signatures alone do not prove inauthenticity. A legitimate political party's social-media team can produce Copy Pasta Test hits; the tool surfaces patterns, not intent.

Privacy and threat model

Mango Tree runs locally on the analyst's machine. The dataset stays under the analyst's control; no upload is required and no vendor processing occurs. This makes Mango Tree a defensible choice for surveillance-environment work in Sri Lanka and Laos, where shipping behavioural data to a third-party cloud would itself constitute a disclosure event.

The threat model that matters here is downstream: a Mango Tree output naming specific accounts circulates as institutional research output, and that naming has consequences for the named accounts and for any source who supplied the underlying dataset. Treat the output file with the same care as a list of named informants in any other investigation; disclose to collaborators only on a need-to-know basis and apply the same source-protection arrangements that would apply to a verbatim quote.

Source-protection override (S4 — doxxing or vulnerable-community targeting)

A Mango Tree CIB report can name activists, journalists, or minority-community accounts caught alongside genuinely coordinated inauthentic accounts. The published naming is the risk; the analysis itself is local. Mitigation steps:

  1. Audit the candidate-account list for activists, journalists, ethnic or religious minorities, LGBTQ+ persons, witnesses, and migrants before publishing any naming output.
  2. Redact or pseudonymise identifiers in published network visualisations; consider a quiet response or institutional- partner alert in place of a public CIB call-out where the retaliation risk against the named operators is non-trivial.
  3. Preserve the source dataset under documented chain of custody on the analyst's machine; disclose to collaborators only on need-to-know basis.

Country and platform applicability

  • Indonesia: language-agnostic; relevant in principle for Bahasa-Indonesian CIB analysis, including the buzzer-network context that documented. No published Indonesian newsroom deployment.
  • Laos: language-agnostic; the offline local-machine execution is one of the few CIB analysis paths that does not require an external service. Honest gap: no Lao-language CIB worked example documented.
  • Malaysia: language-agnostic; relevant for the King / Anwar Ramadan-aid 2026 cluster context (regional case documentation) if a research partner assembles the underlying dataset.
  • Philippines: language-agnostic; relevant in principle for the #FactsFirstPH coalition's broader CIB work and the Brawner / Doc Willie Ong actor pool's amplification network. No published Mango Tree deployment in the Rappler / VERA Files pipeline so far.
  • Sri Lanka: language-agnostic visual modules apply; the local-machine execution model is the right privacy posture for Watchdog and Hashtag Generation's surveillance-environment work. No SL-specific deployment documented.
  • Thailand: language-agnostic; relevant for follow-up CIB analysis on the Anutin / Mauerberger image-distribution network referenced in the Thai PBS / SynthID case (regional case documentation).

Platform applicability: works on any platform whose data the analyst can extract. Works most cleanly with Meta Content Library extracts (Facebook, Facebook Groups) for IFCN-signatory partners and Telegram exports.

How to access

Free download from the civictechdc/cib-mango-tree GitHub repository. Clone the repo, install Python dependencies per the README, and run the interactive terminal. No account creation, no beta gate. The toolkit is licensed MIT.

Cost (current as of 2026-05)

Free. MIT licensed. No quota, no paid tier. The cost of running Mango Tree is the analyst's time to assemble the input dataset and the local compute to run the tests, both of which scale with dataset size. Civic Tech DC's maintenance is community-driven; the toolkit does not depend on a vendor business model and is not exposed to vendor-side pricing changes.

Quickstart

  1. Install Python 3.10+ on the analyst's machine.
  2. Clone civictechdc/cib-mango-tree from GitHub.
  3. Install dependencies via pip install -r requirements.txt per the repository README.
  4. Prepare the input dataset as the README specifies (CSV or JSON of post records with account ID, timestamp, text, and platform fields).
  5. Launch the interactive terminal entry point per the README.
  6. Configure the Copy Pasta Test parameters: similarity threshold (typical: 0.85+ string similarity), time window (typical: ten minutes for tight coordination signals; thirty for looser amplification networks).
  7. Run the test and inspect the output file (clusters of accounts posting near-identical content within the time window).
  8. Cross-check the named accounts against Maltego or Gephi for visualisation, and against Meta Content Library if the IFCN gate is open to your organisation.

In the toolkit's workflow

Behaviour-pillar non-detector tool per Architectural Anchor 1. Sits in 1C.1 Institutional CIB detection as the open-source primary entry, ahead of Coordination Network Toolkit (academic R or Python alternative) and CooRTweet (R, Meta Content Library access required), and ahead of Graphika as the enterprise escalation-option that frontline SEA newsrooms generally cannot reach.

Standard combinations:

  • With Maltego at 2C.2 to visualise the named-account clusters that Mango Tree surfaces.
  • With Gephi at 2C.2 for publication-grade network graphics on the same clusters.
  • With Meta Content Library at 2C.2 as the IFCN-gated data source for the input dataset, when the analyst's organisation is signatory.
  • With Yudistira (MAFINDO) at 2B.2 when the underlying posts need Bahasa-specific claim extraction before behavioural clustering.
  • With ABCDE Framework at 2C.3 for codifying the behavioural pattern in inter-org reporting.

This card carries no detector signal class: Mango Tree produces a non-detector behavioural-pattern signal. Per Anchor 2, that signal combines cleanly with claim-extraction (2B.2), reverse-image hits (2A.2), or source-history records (1B.1) to support an institutional finding. Coordination alone is not enough for a strong public claim; two non-detector signal classes are.

Decision-tree references: T-tree CIB sub-routine routes here from 1B.1 source-history hits when account-level pattern questions emerge; the upward bridge is 2C.3 codification under DISARM / ABCDE.

Override notes

Override notes

  • Behaviour-pillar declaration: Mango Tree serves the behaviour pillar only. Every output is a behavioural-pattern signal; it never produces a content verdict (true / false / AI-generated). The card frames Mango Tree accordingly.

  • No SEA-newsroom deployment yet: the toolkit's source base carries no published Indonesian, Filipino, Thai, Sinhala, Tamil, Malay, or Lao newsroom case where Mango Tree was deployed. The tool enters as primary on framework merit and open-access properties; the regional case for it is to be built rather than already available. Country pages flag Mango Tree as the recommended open-source path for any SEA research partner intending to build a CIB casebook.

Sources

  • Civic Tech DC. CIB Mango Tree — coordinated inauthentic behaviour detection toolkit. GitHub repository (MIT licence). civictechdc/cib-mango-tree.