Context Retention & Attribution Flow
Filter Role:
Analyze context retention, token loss zones, and claim attribution lineages in AI agent executions.
Paste raw agent context trace logs and summary outputs to run local retention scoring:
In long context agent runs, thousands of tokens flow into a prompt, but only a fraction survive into the final summary claims. Each chunk receives a retention score. When the retention threshold is set to zero point twenty, Chunk 3 falls below with score zero point fifteen. Because Chunk 3 holds fourteen thousand two hundred tokens, dropping it leaves ten thousand four hundred retained out of twenty-four thousand six hundred. Selecting a claim or sliding the retention threshold highlights the supporting lineage and updates coverage metrics instantly.
How do retention thresholds and attribution scores quantify which parts of an agent context trace actually support its final claims?
In complex agent loops, thousands of tokens from tool executions, system prompts, and conversation history pass into the language model context window. This tool inspects synthetic token traces, evaluating which chunks exceed a user-selected retention threshold and how generated summary claims map back to source blocks. Dropping low-weighted chunks (such as redundant tool stdout or dump files) increases compression while highlighting potential attribution gaps if essential claims lose their underlying evidence.
The preset token counts, retention scores, and attribution confidence values in this inspector are hardcoded illustrative fixtures, not live neural attention weights or mechanistic attribution probes. In custom audit mode, token counts are estimated by character length heuristics (length / 4) and retention scores rely on simple word-set lexical intersections rather than semantic embeddings or model KV-cache metrics.
Click preset 'Multi-Turn Research' to populate four context chunks totaling 32,000 tokens with two summary claims. At the default 0.20 retention threshold, all chunks meet the threshold (chunk-3 has a score of 0.22), showing 100.0% coverage (32,000 / 32,000 tokens retained) and 0 tk dropped. If you slide the retention threshold up above 0.22, chunk-3 is dropped, reducing retained tokens to 17,000 tk (53.1% coverage) and flagging 1 context block as unmapped.
Large language model agent frameworks frequently ingest large raw tool outputs, such as database logs, DOM trees, or file payloads. While modern models accept large context windows, attention across long inputs can degrade or selectively focus on specific spans, leaving substantial fractions of raw context functionally unused in subsequent generation steps.
In this page's script, context coverage calculates the fraction of total prompt tokens belonging to chunks whose retention score matches or exceeds the threshold slider. Chunks below the slider are marked as unmapped dropped tokens, showing the trade-off between retaining raw context and distilling concise agent outputs.
The interactive D3 flow links generated summary claims on the right to their source context chunks on the left. The attribution gap score displayed in the metrics bar represents (1 - average claim confidence), where a lower score indicates that generated claims are strongly supported by mapped input chunks.
When auditing custom text inputs in this demo, the client script splits text into paragraphs, calculates vocabulary intersection between the input and summary sentences, and assigns synthetic retention values. It serves as an educational heuristic model rather than an authoritative cryptographic or mechanistic attribution guarantee.