METHODOLOGY · PUBLISHED RULES

Every score should be reproducible

This page documents the exact rules the code runs. The configuration it describes lives in src/lib/config and src/lib/ingest; changing those files changes this page.

1. Ingestion

Providers fetch from public endpoints on a schedule: RSS/Atom feeds and GitHub release lists every 30 minutes, the arXiv API every 6 hours, reviewed manual files with each feed run. The frontend never scrapes. Each fetch is recorded as an ingestion run with status, counts and a log; sources keep their last status and error so failures are visible on /sources instead of silently producing nothing.

Items are de-duplicated by a hash of the normalised source URL (tracking parameters removed). Reviewed rows are never overwritten by automation.

2. Classification

Classification is deterministic keyword scoring. The same text always yields the same result, and each record stores the matched terms (open “Why this classification?” in any evidence panel).

An AI gate must pass (terms such as “language model”, “LLM”, “frontier model”, lab or model names). Generic feeds (regulators) require it strictly.
Each category sums the weights of matched terms. The best category wins if its score reaches the threshold of 3; ties break in this priority order: RESEARCHER_WARNING › INCIDENT › CYBER › ALIGNMENT › AUTONOMY › GOVERNANCE › CAPABILITY.
Papers cannot be incidents; papers only count as capability or autonomy evidence if they mention frontier-model terms; “policy” in a paper is treated as RL jargon unless explicit governance terms appear.
Non-paper items that pass the gate but match no category are not indexed. Customer stories and marketing patterns in titles are dropped.
Timeline events derive only from first-party or government sources; policies derive only from official lab or government sources.

Category lexicon

CAPABILITY

introducing +1.5announcing +1new model +2state-of-the-art +2benchmark +1.5frontier model +2context window +2million-token context +2reasoning model +2outperforms +1.5capability +1coding agent +2named benchmark +3model launch +2multimodal +1scaling +1agent +1computer use +2

ALIGNMENT

alignment +2misalignment +3deception +3reward hacking +3goal misgeneralization +3situational awareness +2interpretability +2oversight +2alignment faking +4sycophancy +2safety research +2alignment technique +1.5honesty +1red teaming +1.5jailbreak +1

AUTONOMY

autonomy +3agentic +2.5AI agent +2long-horizon +3multi-step +1.5tool use +2self-improvement +3computer use +2task completion +1.5planning +1time-horizon eval +2

CYBER

cyber +3vulnerability +2exploit +2malware +3CTF +3pentest +2zero-day +3phishing +2prompt injection +2threat report +2misuse +1CVE +3

INCIDENT

incident +3outage +1postmortem +2anomaly +3bypass +2jailbreak +2leak +2breach +3rogue +2disclosure +1destructive action +2manipulation +2

GOVERNANCE

policy +2regulation +3safety framework +4safety commitment +3executive order +3legislation +3standards +1.5governance +3system card +3safety report +3export control +2AI safety institute +3deployment policy +3evaluation requirement +3

RESEARCHER_WARNING

resignation +4warning +2open letter +4catastrophic risk +3whistleblower +4safety team +2departure +3call for pause +4statement on risk +4

Severity

Severity terms: critical (+1), severe (+1), catastrophic (+2), exploit (+1), destructive (+1), urgent (+1), misalignment (+1). Total ≥ 4 → critical, ≥ 2 → high, ≥ 1 → moderate, else low. Research and capability items cap at high; papers cap at moderate.

3. Confidence

Confidence expresses provenance, not truth. It is the source-type weight multiplied by (0.8 + 0.2 × classification certainty), where certainty is the winning category score divided by 6, capped at 1.

Source type	Weight
Official lab publication	0.90
Peer-reviewed research	0.90
arXiv preprint	0.75
Major reporting	0.65
Government / regulator	0.85
GitHub release	0.60
Personal post (first-party)	0.55
Personal post (third-party)	0.30
Manual entry (reviewed)	0.60

Labels: HIGH ≥ 0.85, MEDIUM-HIGH ≥ 0.70, MEDIUM ≥ 0.50, otherwise LOW.

4. Incidents and researcher events

Both enter only through reviewed entries (data/manual/*.json or the admin console) and require at least one source URL. Quotes require attribution. Verification statuses: Unverified, Source confirmed, Multiple sources, Official report. Unverified incidents are displayed but excluded from the Risk Index.

5. Risk Index

The index is a weighted sum of six components, each scored 0–100 over a trailing window of 90 days compared with the 90 days before it. Weights are configuration and are stored with every snapshot. A snapshot is published only when the window holds at least 25 signals, 5 capability records in both windows, and 3 sources with data; otherwise the page shows RISK INDEX UNAVAILABLE · INSUFFICIENT SOURCE DATA with the reason.

Component	Weight	Method
Capability acceleration	30%	Ratio of capability-classified signals and capability timeline events in the current window versus the baseline window, weighted by source confidence. 1.0x ratio maps to 50; 2.0x or more maps to 100; 0.5x or less maps to 0. Requires at least the configured minimum of capability signals.
Autonomy & agency	20%	Share of capability evidence tagged autonomy, agentic, tool_use or long_horizon, blended 50/50 with the growth of that share versus the baseline window.
Safety gap	20%	100 minus the ratio of safety-side evidence (alignment signals, alignment/interpretability/evaluation papers, safety frameworks) to capability-side evidence in the window, scaled so parity maps to 50. A window where capability evidence is twice the safety evidence scores 100.
Incident severity	15%	Sum of severity ranks (low 1 … critical 4) of incidents with at least SOURCE_CONFIRMED verification inside the window, weighted by verification level (0.6 / 0.8 / 1.0), scaled with the saturating curve 100 × (1 − 2^(−points/10)): 10 severity points maps to 50, 30 maps to 87.5. UNVERIFIED incidents are excluded.
Governance weakness	10%	100 minus a saturating function of governance evidence (policies, regulations, commitments) in the window relative to the number of labs tracked. Zero governance evidence scores 100; one governance item per tracked lab scores 50.
Researcher warning signal	5%	Confidence-weighted count of researcher events of type resignation, public_warning or open_letter plus RESEARCHER_WARNING signals inside the window, scaled with 100 × (1 − 2^(−n/5)): 5 weighted events maps to 50, 15 maps to 87.5.

Levels: LOW < 25, GUARDED < 50, ELEVATED < 75, CRITICAL ≥ 75. The log-ratio mapping used by several components is 50 + 50 × log₂(ratio), clamped to 0–100, so parity scores 50 and a doubling scores 100.

6. Dogu Terminal

Queries are interpreted with rules (time window phrases, category keywords, lab aliases, verification words). Records are retrieved from the relevant tables and ranked by 0.5 × confidence + 0.3 × recency + 0.2 × keyword overlap. The structured answer is assembled from the ranked records. If an ANTHROPIC_API_KEY is configured, a synthesis pass may restate the numbered evidence in prose; it is discarded if it cites anything outside the list. Below the evidence minimum the terminal returns INSUFFICIENT EVIDENCE.

7. Daily brief

The brief lists counts and top records per section for the selected window, with the preceding window of equal length for comparison. No generated prose.

8. Known limitations

Keyword classification misfiles some records; reviewed corrections in the admin console are permanent and audited.
Labs without public feeds are under-represented in automated evidence.
Counts reflect what DoguLabs indexed, not everything that exists.

LOADING INSTRUMENTS…