Run a small, controlled A/B experiment testing the data-organization technique pasted at the end after "TECHNIQUE:". Build it, run it, analyze it, and write the report. Don't ask me questions — make sensible choices and document them.

1. HYPOTHESIS. State one testable hypothesis and ONE primary metric (usually accuracy). Note the technique's Type: if "representation" or "access", both conditions MUST contain identical facts — only the organization differs. If "augmentation", state exactly what was added.
2. GROUND TRUTH. Write a seed-fixed generator script that simulates entities with hidden traits interacting over time, using INVENTED names and numbers so nothing is answerable from world knowledge. It must take a difficulty level (1, 2, 3) and output canonical-facts.json plus the raw data. Higher levels mean: bigger corpus (roughly 50k / 200k / 500k tokens), more near-miss distractors, stale versions, and conflicting values with dates, and more multi-hop facts. Declare these three levels now; do not change them later.
3. CONDITIONS. Write ONE transformation, once: it takes the raw data and produces ./A/ and ./B/ (randomly assign which is which; record the mapping, never reveal it to sub-agents). Human condition: how a competent team would actually store this today. No strawman — a real team should defend it as normal. Agent condition: the technique applied to the human condition. Apply this same transformation at every difficulty level. Never revise either condition after seeing any results. No per-file size limits — let each condition organize files the way it naturally would in terms of formats (e.g. PDF, DOCX, JSON, XLSX, databases, ...), folder/table structures and file/table counts.
4. TASKS. From canonical-facts.json ONLY (never from the folders), write 8 questions per level (24 total), each with one exact expected answer a script can check: 6 on-target (situations the technique targets) and 2 off-target (ordinary questions it wasn't built for).
5. RUN. Do 2 unscored runs to check the harness works, then run all 24 tasks × 2 conditions. Each run is a FRESH sub-agent with a clean context (Task tool or a separate non-interactive CLI process — never this conversation), using a fast, cheap model, in parallel batches. Every sub-agent gets the SAME verbatim prompt — only the folder path differs — with the question, read access to one folder, identical tools (read, grep, bash), and a cap of 15 tool calls (hitting the cap = failure). Never reveal the hypothesis, the technique, the expected answer, or the other folder. Record: answer, tool calls, bytes of tool results (as a token proxy), and whether the cap was hit.
6. SCORE. Grade with a deterministic script (normalized exact match — no LLM judging). Primary test: paired sign test / McNemar pooled across all 24 pairs. Then break results down by difficulty level and by on/off-target, and describe the trend: does the gap grow as difficulty rises? Report median tool calls and result-bytes as secondary metrics.
7. REPORT. Write RESULTS.md: the technique card, hypothesis, the three predeclared difficulty levels, exact models and isolation method, the generator and transformation scripts inline (runnable via `uv run` with inline deps, fixed seeds), both folder trees with sample files, all 24 tasks + expected answers, the sub-agent prompt verbatim, raw results table, statistics, and a conclusion in this vocabulary: "difference appears from level N" / "no difference at any tested level within budget" / "B hurts" — for this model and task family only. State limitations (one synthetic world, one model, n=24). This one file must be enough for anyone to reproduce everything.

Budget guard: stay under ~50 sub-agent calls total.

OPTIONAL EXTENSION (only if budget remains): test whether an agent can IMPROVE the organization. Split 12 new dev tasks / 12 held-out tasks. Let a fresh agent iterate on condition B using only dev-task feedback, then score the revised B once on the held-out tasks. Report separately.

TECHNIQUE:
{paste one ### T{n} card here}