Reliability Engine
Reliability scoring for OpenClaw agents — scores claims by evidence and persuasion technique, not by the politics of the source. Explainable, non-partisan, auditable.
Install
openclaw plugins install clawhub:@laninthalesdran/openclaw-reliability-engineTwo ways to use the Reliability Engine — 🤖 OpenClaw agent plugin (this repo) · 🧩 Browser extension → Reliability-Web-Extension. Same engine, two front-ends.
Reliability Engine — an OpenClaw plugin
Scores how much a claim should move your beliefs — by evidence and persuasion technique, not by the politics of the source. Explainable, non-partisan, auditable. No censor in the basement.
Most "reliability" tools ship a list that says one outlet is good and another is bad. That bakes in a worldview and calls it objectivity. This one does the opposite: it scores the claim and its evidence, detects persuasion technique (not position), and earns source reputation from outcomes — and it shows its work every time.
The 10-second demo
Same fact. Same source. Same evidence. Only the wording changes:
"The agency reported Q1 GDP rose 0.7%." -> 9.29 / 10
"SHOCKING betrayal: the corrupt regime's so-called experts say a
catastrophic economy is an existential threat — everyone knows it,
wake up." -> 4.17 / 10
~ leading language 0.918 [loaded_language, appeal_to_fear,
us_vs_them_namecalling, moral_emotional_outrage]
The language alone cost it 5 points — and the tool names exactly which techniques did it.
The Onion: "Congress passes bill requiring all citizens to own a goose" -> 0.5 / 10
!! VETO: non-factual (source self-declares non-factual)
A conspiracy claim that can't be falsified -> 2.0 / 10
!! VETO: unfalsifiable (reputation cannot rescue this)
What it does (4 tools)
| Tool | What it answers |
|---|---|
reliability_score_claim | How much should this claim move my beliefs? (0–10 + breakdown) |
reliability_check_source | What's this source's validity / bias / salience / non-factual status? |
reliability_scan_language | What persuasion techniques is this text using? (0–1 + techniques) |
reliability_rate_bias | How slanted is this source? Rate its structural bias (0–1) from a batch of its claims — symmetric, kept separate from validity |
How it scores (no magic, all auditable)
8 evidence primitives, each 0–1, transparently weighted: provenance · verifiability · corroboration (independent only) · falsifiability · transparency · incentive-alignment · recency · source-prior.
Vetoes reputation can't override: an unfalsifiable claim or one with no checkable origin is floored. A source that self-declares as satire/parody (The Onion, etc.) is floored — but mixed outlets that also do real journalism (Private Eye, Cracked) are judged per claim, never blanket-zeroed.
Leading-language discount (up to −60%): detects persuasion technique, grounded in real research and politically symmetric by design (the us-vs-them lexicon carries left- and right-coded pejoratives equally). Categories cite: Da San Martino et al. 2019 (propaganda techniques), Brady et al. 2017 PNAS (moral-emotional diffusion), Tversky & Kahneman 1981 / Entman 1993 (framing), Ganter & Strube 2009 (weasel words), Cialdini 1984 (social proof), Blom & Hansen 2015 (clickbait), MPQA subjectivity lexicon.
Anti-laundering trust: a low-validity source cannot vouch another up — endorsements are gated by the endorser's own validity squared, cliques collapse, and validity is earned from resolved-claim outcomes, not assigned. (InfoWars vouching for a source moves ≈0.0025.)
State-media control graph: outlets sharing a controlling/funding entity (e.g. RT + Sputnik + TASS = Russian state) collapse to one corroborator so coordinated state outlets can't fake independent confirmation; state control raises the incentive-conflict floor and attaches an honest, control-type-scaled flag (a direct state instrument is framed differently from a publicly-funded independent broadcaster like the BBC). Applied symmetrically across all governments — it's a documented funding/control fact, not a political judgment. See data/state_media_registry.jsonl.
Bias rating — structural, symmetric, firewalled from validity: every source carries a bias (direction + magnitude) that the engine never lets bleed into validity — a source can be accurate and slanted, and the two are always reported separately. Magnitude is measured from persuasion-technique density (the same symmetric leading-language detector), so it counts technique, not which side it serves; direction can be fed by the Coverage Engine's measured selection/omission asymmetry, never assigned by hand. reliability_check_source returns it. Honest caveat: because magnitude rides on the leading-language detector, a calm-but-slanted source can read low — see KNOWN_LIMITATIONS.md §3.
Install
openclaw plugins install clawhub:tntholley/reliability-engine
Verify it locally (no build needed; Node 24+ runs TypeScript directly)
node test/demo.ts # reproduces every number above — 20/20 checks
# (claim-scoring, anti-laundering, leading-language, satire,
# calibration, state-media, and bias rating)
Layout
index.ts OpenClaw plugin entry (registers the 3 tools)
openclaw.plugin.json plugin manifest
src/ the engine (zero OpenClaw deps — reusable, testable)
engine.ts claimScorer.ts rhetoric.ts trustGraph.ts stateMedia.ts calibrate.ts registry.ts types.ts
data/ the knowledge: leading-language lexicon (cited), self-declared-
non-factual registry (71 outlets) + patterns, state-media control
graph, seed source memory
test/demo.ts verification harness
Known limitations — read them
This is a confidence-weighting aid, not a truth oracle and not a censor. Its hardest adversary is state media (accurate enough to keep its validity high while distorting through framing, coordinated corroboration, and omission), and omission/agenda-setting is structurally beyond a per-claim scorer. The control graph above mitigates the corroboration-gaming and flags state control, but the residual is real. The full, honest list is in KNOWN_LIMITATIONS.md — read it before trusting a score. An engine that hides its blind spots is worse than one that names them.
License & ethos
Apache-2.0. Built by Travis Edward Holley (TNT Holley Inc.) and Claude Opus 4.7 (1M context, Anthropic). The engine refuses to be a partisan referee: it scores evidence and technique, keeps bias (which way it leans) separate from validity (whether it's accurate), and every score is explainable so you can argue with it. That's the point.
