Simplify dataset README

README.md

@@ -31,7 +31,7 @@ Changes from the original dataset:
 
 - Divide input and output lengths by **8**, rounding each output up. Hash IDs stay the same; logical blocks shrink from 64 to 8 tokens.
 - Divide active request time by **8** and cap ordinary whole-session idle gaps at **1 second**. Keep original gaps over an hour just above the harness's one-hour stream-separation threshold; the runtime idle cap skips those waits.
-- Keep complete sessions with **less than 10,000 requested output tokens**, counting all subagents and auxiliary calls. Total requested output: **962,859 tokens**.
+- Keep complete sessions with **less than 10,000 requested output tokens**, counting all subagents and auxiliary calls.
 - Preserve request grouping, spawn/join dependencies, shared prompt content, and reference overlaps. The agentx-mini traces are model-independent; actual replay overlap depends on server speed.
 
 ## Stats
@@ -46,26 +46,18 @@ total_input_tokens:   126,413,728
 total_output_tokens:      962,854
 ```
 
-These use the original dataset's counting convention: main turns are top-level
-calls, and subagent groups are explicit containers. Token totals are trace
-counts; rendered prompts also include tokenizer and chat-template overhead.
+Main turns are top-level calls, following the source dataset's convention.
 Five zero-output calls request one token each during replay, bringing requested
-output to **962,859 tokens**.
+output to **962,859 tokens**. Rendered prompts add tokenizer/chat-template overhead.
 
-Input totals count repeated prefixes on every call. With a cold, unlimited
-cache shared by all agents of each session on one model, the hash-based estimate
-is **5,499,568 uncached input tokens** (**95.65% prefix reuse**). Concurrent
-prefills, eviction, and tokenizer/chat-template overhead affect actual prefill.
+Input totals count repeated prefixes. The ideal estimate is **5,499,568 uncached
+input tokens** (**95.65% prefix reuse**), assuming a cold, unlimited cache shared
+by each session's agents on one model. Actual prefill depends on the server.
 
 ## Files and reproduction
 
 `traces.jsonl` contains the scaled Weka records; `stats.txt` contains their totals.
-Each record's optional `totals` field includes a modification notice and token
-totals; this metadata does not affect replay.
 AgentX Harness handles prompt synthesis and replay using the target model's tokenizer.
-All 200 sessions retain their original request grouping and execution dependencies
-with the tested harness. Seven agent/auxiliary labels change across six sessions;
-their execution graph and synthesized prompt content remain equivalent.
 
 To rebuild agentx-mini:
 
@@ -88,9 +80,6 @@ source .venv/bin/activate
 pip install agentx-harness==1.0.7
 ```
 
-The package is `agentx-harness`; the command is `agentx`. Its built-in `weka_hf`
-loader reads agentx-mini and synthesizes prompts using your chosen tokenizer.
-
 1. **Download agentx-mini** into a local directory.
 2. **Start your local LLM server** and load your model, for example using
    `llama-server`. It must expose an OpenAI-compatible API. Enable parallel
@@ -139,35 +128,13 @@ agentx profile \
 ```
 
 **Automatic download:** instead of downloading first, set `hf_weka_dataset` to
-the HF repository ID (`SocialLocalMobile/agentx-mini`) and run the same command
-from the directory containing `benchmark.yaml`. AgentX Harness downloads agentx-mini
-automatically. `distill.py` is only for rebuilding agentx-mini.
-
-With `--config`, AgentX Harness uses the settings in the YAML file. This configuration
-starts sessions at their first turn and runs one session tree at a time,
-including its parallel subagents, with a **20-minute profiling budget**.
-Startup and final request draining can add time; this does not guarantee that
-all 200 sessions finish. To replay all 200 once, replace `duration: 1200` with
-`requests: 9439`.
-
-Keep `timing_mode: agentic_replay` and the runtime idle cap: the retained long
-reference gaps help reconstruct separate agent streams, while the cap skips
-waits when the entire session tree is idle. Ordinary fixed-schedule replay
-would wait through those gaps. At the tested harness version, some joins can
-still leave extra idle waits (about 8–11 seconds observed). For a tighter bound
-with `concurrency: 1`, optionally add `system_idle_gap_cap_seconds: 1` to the
-profiling phase.
-
-All **200 sessions / 9,439 calls** were replayed against a mock server using
-unmodified agentx-harness commit
-[`e4bb2f8`](https://github.com/SemiAnalysisAI/agentx-harness/commit/e4bb2f835ffedbbe6b76252ea5c93caec68c5faf)
-with its default reconstruction thresholds. Call coverage, output budgets,
-prefix reuse, and dependency ordering were checked; mock timings do not measure
-real model performance.
-
-The released `agentx-harness==1.0.7` was also tested with Qwen3-0.6B on
-`llama-server`: a 20-minute run completed 26 sessions plus part of another,
-with **1,279 successful calls**. Output budgets and dependency ordering passed.
+`SocialLocalMobile/agentx-mini` and run the same command.
+
+Keep `timing_mode: agentic_replay` and the runtime idle cap to skip retained long
+reference gaps. The harness can occasionally exceed the idle cap around joins.
+
+Call coverage, prefix reuse, and dependencies were validated across all 200
+sessions. Also tested with `agentx-harness==1.0.7` and Qwen3-0.6B on `llama-server`.
 
 ## License