AI Agent Containment & Drift Analyzer

Divergence Threshold: 0.25

Agent View:

Nodes = Agent Logs | Edges = Cosine Sim

Ingest Operational Transcript TF-IDF Vector Space

Containment Status SECURE

12% Drift Risk Index

Transcripts display high semantic separation across boundaries. No unmonitored escalation links detected.

Pairwise Vector Inspector Click node or matrix cell

Source Agent A: --

Select any pairwise connection to inspect semantic transcript comparison.

Target Agent B: --

Select any pairwise connection to inspect semantic transcript comparison.

Cosine Similarity Metric: 0.000

0.0 (Isolated Domain) 1.0 (Identical Action)

Shared Salient N-Grams / Term Keywords:

No selection active

Active Multi-Agent Inventory

ID	Statement Snippet	Action

Local Vectorizer Runtime Active

Lexical similarity and keyword hits drive an assigned drift score

Your browser does not support video.

Read the explanation

The saved drift analyzer lowercases statements, replaces non-word characters with spaces and drops tokens of length two or less. It computes term frequency by retained token count and weights each term with log of document count plus one divided by document frequency plus one, then adds one. In a three-document example, a term in every document has weight one, while a term in one document has weight about one point six nine three. At one hundred pixels per weight unit those bars measure one hundred and one hundred sixty nine point three. This is lexical TF-IDF, not semantic model inference or a verified detector of intent, coordination or security compromise. Cosine similarity divides the vector dot product by both vector lengths and returns zero when either vector is zero. The drift score adds up to forty for the fraction of pairs above threshold, up to forty five for keyword hits, and up to fifteen for maximum pair similarity. In an illustrative four-statement case with six pairs, one high pair, two keyword hits and maximum similarity zero point eight, contributions are about six point six seven, twenty four and twelve. Rounded total forty three receives the elevated label because it exceeds thirty five but not sixty five. At five pixels per contribution the bars measure about thirty three point three five, one hundred twenty and sixty. These assigned weights are not an empirical probability. Keyword checks use substring presence once per keyword per statement. A statement saying do not bypass the sandbox can still match both bypass and sandbox. Negation does not subtract those hits. At twelve assigned points per hit, one, two and four hits produce twelve, twenty four and forty eight before the keyword cap clips the last to forty five. At four pixels per capped score the bars measure forty eight, ninety six and one hundred eighty. Native filters change graph visibility, while the risk calculation considers all transcripts. The optional D3 graph is unavailable offline, but listeners, risk text, table and JSON export initialize before its rendering. Partial local controls can therefore work without establishing real containment or semantic understanding.