# What the Elisa experiments show

These results describe the supplied mock file, not real customers or population
risk. The source checksum is
`36be13cea8c2ea9448317fb026c09de967242df06e7c00b8e6256fd89c48a214`.
The reviewed evidence below is available on main `6c95595`. This is an evidence
handoff for AIH-8, not the finished interactive demo.

## A three-minute story supported by evidence

**Removing names does not remove all distinguishing detail.** In two tested row
releases, adding exact throughput and latency to the original quasi-identifiers
raises measured unique-row fractions from approximately 0.006% and 0% to about
81.85%. That assumes an attacker knows those exact measurements. It is neither
successful recovery of real identities nor an estimated real-world attack rate.
See the [QI sensitivity probe](../experiments/mikael_qi_stress/probe.py) and
[review/provenance record](experiments/mikael-backlog.md).

**Keeping almost every row can still destroy the useful signal.** Traffic-medium
rounding with full-field suppression keeps 96.60% of rows on the full input but
only 16.48% of traffic volume. In three fixed subscriber-composition samples it
keeps about 96.2% of rows but only 14.66–15.02% of volume. These are overlapping
same-data sensitivity fixtures, not confidence intervals or held-out validation.

**Overall utility hides uneven losses.** Coarse rounding with original-QI
suppression keeps about 89.6% of sampled volume overall, but 2G volume becomes
zero, 4G retention is 35.24–36.42%, and 5G retention is 93.62–93.68%. Zero volume
does not necessarily mean zero retained rows. See the complete before/after
application and RAT tables in the
[reviewed composition experiment](../experiments/mikael_subsample_stress/RESULTS.md).

**A narrower aggregate product can preserve traffic summaries, at a cost.**
Aggregating by hour, province, application and RAT with at least five observed
contributors keeps 99.9737% of volume. Yet a single contributor can supply 99.9818%
of a released cell. Suppressing cells whose largest contributor exceeds 20%
keeps 87.6970% of volume. This replaces individual rows with sums and counts;
quality distributions and subscriber/session analysis are not preserved.
See the [reviewed aggregate experiment](../experiments/mikael_aggregate_traffic/RESULTS.md).

The aggregate cap also has uneven costs: at five contributors and the 20% cap,
2G retains only 14.42% of volume, versus 92.96% for 4G and 87.30% for 5G.
File-transfer volume is entirely suppressed; IoT retains 0.76%. These are
fractions of each group's own original volume, not shares of the whole file.
The coordinator independently reconstructed all app/RAT marginals from the source;
local evidence is `out/coordinator-stress/aggregate-subgroup-reproduction.json`.
An aggregate-volume improvement must therefore be shown with subgroup losses.

## Comparisons that can be shown together

| Product / fixed configuration | Retained volume on full input | What the measurement establishes | What it does not establish |
| --- | ---: | --- | --- |
| Traffic rows, medium rounding, full-field suppression | 16.4821% | No released full-field row classes below five in this fixture; substantial volume loss | General anonymity, robust linkage resistance or good utility from high row retention |
| Traffic + quality rows, medium rounding, full-field suppression | 0.5126% | Very restrictive measured row release with severe volume loss | A usable quality product merely because quality columns remain |
| Aggregate traffic, five contributors, no dominance cap | 99.9737% | Strong preservation of tested traffic totals and app shares | Protection from a dominant contributor or known-others reasoning |
| Aggregate traffic, five contributors, 20% dominance cap | 87.6970% | Measured cell-level contributor-share bound after suppression | Differential privacy, safe repeated releases or preservation of individual-row uses |

Rows and aggregate cells are different products. The volume column is a shared
descriptive utility quantity, not permission to rank incompatible privacy metrics
or declare a universal winner. All percentages use the full observed source-volume
denominator for the corresponding full-input run. Physical units and business
validity of the mock measurements have not been independently established.

## Useful chart data and remaining work

- M-EXP-03 `out/mexp03/*.stress.json`: row/volume coverage and full-cell diagnostics.
- M-EXP-04 `out/mexp04/chart-summary.json`: aggregate traffic and marginal series.
- M-EXP-05 `out/mexp05/sensitivity-summary.json` plus individual results:
  same-sample app/RAT input and released volumes, coverage and sensitivity ranges.

Those ignored local outputs live in the implementer worktree; committed experiment
READMEs provide reproduction commands and RESULTS files record immutable artifact
hashes. Keep original artifacts and their executed-code provenance. Prepare a
separate display copy if the UI needs another format; never silently relabel
previously inspected metrics as held-out.

The UI should lead with paired traffic bars and a row-versus-volume comparison,
then show subgroup losses and the aggregate-product trade-off. Add the existing
quality measurements and an explicitly scoped candidate-narrowing illustration;
do not invent linkage counts or show source identifiers. Unsupported analysis
must be visibly unavailable, not zero risk. Definitions, provenance and coverage
belong in accessible detail beside the story.

The reviewed [M-EXP-06 clipping comparison](../experiments/mikael_contributor_clipping/RESULTS.md)
retains 95.7637% volume at five contributors / 20% maximum share, versus 87.6970%
for whole-cell suppression. This is not equal-privacy evidence: more aggregate
cells are exposed, and repeated-release/known-others reasoning remains unprotected.
P2P still retains only 4.1021%, VoIP 13.5947%, and 2G 58.1106% of original volume.
All six settings and both methods were independently reproduced in fresh review.
The reviewed [window comparison](../experiments/mikael_window_aggregation/RESULTS.md)
retains all six original window labels at k5/cap0.2, but only 79.5194% of volume,
versus 87.6970% for hourly cells. Hourly output cannot answer the original-window
question. Retaining labels is also not enough: File transfer, IoT and Remote
access volume disappears entirely in the original-window release. All six
settings were independently reproduced. The reviewed
[clipping sensitivity experiment](../experiments/mikael_clipping_sensitivity/RESULTS.md)
retains 94.7546–95.3983% volume at cap0.2 across three fixed subscriber samples,
versus 95.7637% full input. P2P still retains only 1.9334–4.1838%. Two different
sample/cap settings were independently reconstructed by coordinator and reviewer;
full baselines match exactly. These overlapping samples do not provide confidence
intervals or held-out validation. The reviewed
[original-window clipping experiment](../experiments/mikael_window_clipping/RESULTS.md)
retains 92.3702% volume at k5/cap0.2 versus 79.5194% for whole-cell suppression,
with original-window TVD 0.0156875 versus 0.0360683. Both retain six labels, not
exact window shares. P2P retains only 1.1228%, Speedtest 1.8824% and VoIP 5.4326%.
Coordinator and reviewer independently reconstructed different settings with
bisection; no equal-privacy claim follows from the shared cap. The reviewed
[quality histogram experiment](../experiments/mikael_quality_histograms/RESULTS.md)
preserves coarse quality distributions as a different aggregate product, not raw
measurements or within-bin detail. At k100, throughput retains 887,413 of 887,515
valid rows, but 211,825 source values were already missing. RTT has 508,833 missing
values; 2G contributes only 115 valid values out of 3,236 rows and all 115 are
suppressed at k100. Its distribution is unavailable, not zero distortion. The reviewed
[paired-release sensitivity study](../experiments/mikael_aggregate_differencing/RESULTS.md)
finds changed contributor counts and volumes in all six tested whole-subject
removals while min5/maxshare0.2 bounds still hold. The observer is assumed to see
both releases and know removal timing; this is not measured real-world membership
accuracy. Coordinator and reviewer independently reconstructed two cases. The
result demonstrates why single-release bounds cannot promise repeated-release
protection. These findings motivate showing missingness and suppression alongside charts.
AIH-8 still needs one agreed UI owner, a complete
chart-led interactive story and browser acceptance. AIH-10–12 retain the narrower
time-loss, pseudonym-policy and presentation-provenance follow-ups. No result here
is a release approval, differential-privacy guarantee or general anonymity claim.

The reviewed [quality coarsening study](../experiments/mikael_quality_coarsening/RESULTS.md)
shows the cost of recovering sparse measurements. Merging pairs or groups of four
RTT bins still leaves2G unavailable at k100. Merging all eight recovers115 valid
values but destroys within-RAT distribution detail: every original-boundary CDF
interval spans0–100%. Even paired merging leaves maximum ambiguity of about68
percentage points in4G/5G. This is release-resolution ambiguity, not a statistical
confidence interval. Selective contiguous partitions are now being tested under
a fixed objective; no optimized result is accepted yet.