time estimation and experiment details in README.md

.gitignore

@@ -3,3 +3,4 @@ venv/
 *.parquet
 *.zip
 *.csv
+__pycache__/

README.md

@@ -1,3 +1,44 @@
-# Budget Allocation in Differential Privacy
+# Experiment: PREDICTIVE METRIC FOR OPTIMAL BUDGET ALLOCATION IN DIFFERENTIAL PRIVACY
 
-O experimento está em experiment.py
+Based on the [original experiment](https://github.com/conseg/TheImpactofDifferentialPrivacyondatautilityinfundamentalmathematicaloperations), but using a real life dataset.
+
+Key difference: sensitivity calculation to satisfy Differential Privacy.
+
+## The Dataset
+
+PNAD Contínua (IBGE - Brasil) - 2026 (primeiro trimestre)
+
+## The Experiment
+
+Objective: evaluate the metric on a t-test equation to compare Northeast and Southeast income (informal vs formal workers)
+
+Steps:
+
+1. Clip and normalize dataset (C = 46_366, the constitutional salary cap, which is set at the salary of Supreme Court justices). Must be a value taken from the real world, not from the dataset.
+2. Calculate necessary statistics and their sensitivities
+3. For each possible budget allocation with total epsilon=12, granularity=0.5, epsilon > 0 (1,352,078 possibilities): evaluate the metric for the statistics and the budget allocation sequence
+4. Save all possible scores and budget allocation sequences to csv for further analysis, and print the best metric score.
+
+## Estimated Experiment Execution Time
+
+To reproduce (in a Linux shell):
+
+1. `time python ./experiment.py real`
+2. Send an interrupt (Ctrl+C) when N exceeds 500.
+3. To find the number of hours, based on the last N and the elapsed time (t): `((1352078*t)/n)/3600`
+
+Laptop (Intel Core i5 1135G7 - 8GB RAM)
+PC (AMD Ryzen 7 8700G - 32GB RAM)
+
+Laptop: 514 iterations in 12.13s
+PC: 493 iterations in 8.11s
+
+Estimated total time on the laptop: 8h51m
+Estimated total time on the PC: 6h10m
+
+## Tasks
+
+1. Provide a way to run the experiment in parts, to allow parallel execution on multiple CPUs.
+2. Run the experiment to evaluate the metric.
+3. Combine the resulting CSVs into one.
+4. Analyze the csv to find notable metric scores for specific sequences.

experiment.py

@@ -126,7 +126,9 @@ print(formal_sen)
 results = []
 best_metric = sys.float_info.max
 
+n = 0
 for bud in sequences:
+    n += 1
     metric = 0
     sta = informal_sta + formal_sta
     sen = informal_sen + formal_sen
@@ -135,7 +137,7 @@ for bud in sequences:
     metric += us_result
     metric += ue_result[0] + ue_result[1]
     metric = metric / 14.0
-    print(metric)
+    print(f"n={n}; metric={metric}")
     results.append([*bud, metric])
     if metric < best_metric:
         best_metric = metric