Experiment record

NORITZ one-minute raw audit: 588 daily-field comparisons with zero mismatches

A read-only audit reconstructed all 21 NORITZ daily aggregate columns for 28 completed JST days from retained one-minute raw evidence. All 588 date-by-field comparisons matched, so no production values were changed.

Summary

AIEL-2026-0009 audited whether retained one-minute NORITZ water-heater observations could reproduce the stored daily aggregate table after collector changes raised concern about stale or inconsistent historical values.

The final read-only audit covered 28 completed JST days, from 2026-08-09 through 2026-09-05, and all 21 daily columns. Every date-by-field value was independently reconstructed from retained raw evidence and the documented derivation rules.

The result was:

The suspected historical inconsistency was not observed in the audited range. Because the repair rule permitted mutation only for demonstrated mismatches, the evidence-preserving outcome was to leave the production daily table unchanged.

Background

The NORITZ collection path stores one-minute raw observations and derives daily aggregates from them. After changes to the collector, historical sample_count values were questioned because a nominal one-minute cadence can suggest an expectation near 1,440 observations per ordinary 24-hour day.

That expectation alone is not a valid correction rule. Missing observations must remain missing, and the stored count must represent admitted observations rather than a value forced to 1,440.

The experiment therefore treated the retained one-minute layer as the lower-level Evidence source and asked a narrower question: do the stored daily values actually disagree with what can be reconstructed from Raw?

This experiment extends the data-quality use of the Raspberry Pi/NORITZ collector documented in AIEL-2026-0003. It does not reopen that experiment’s conclusion about network access to heat-pump electrical power or energy.

Methods

The audit was performed without writing to the production daily table.

Retained one-minute observations were taken from the operational Raw store and a pre-archive backup that preserved the earlier part of the audit interval. Records were normalized by unique source key and observation time, then grouped by JST calendar day. The stored daily row was compared against an independently reconstructed row field by field.

All 21 daily columns were included, covering timestamps, sample count, observed duration, cumulative gas and water boundaries and deltas, operating-state sample counts, residual-hot-water references, tank capacity, status, and note.

The current Apps Script implementation established the exact daily state rule used for the two text fields:

• invalid gas delta: 要確認;
• current JST day: 集計中;
• completed past day: 日次確定 only when observed_hours >= 23.5 and sample_count >= 1200;
• otherwise: 観測不完全.

note is deterministically generated from that state and from whether both first and last residual-hot-water values are available.

The stored sample_count itself is not replaced by 1,440. It is the admitted unique valid one-minute observation count. Missing observations are not converted to zero-valued measurements.

The still-accumulating 2026-09-06 daily row was excluded from the final comparison because it was not a completed day at the audit snapshot.

Experiment Log

The first reconstruction pass established that 19 of the 21 daily columns could be derived directly from the retained Raw values and that those reconstructed values matched the stored daily rows.

At that stage, status and note were not treated as reconstructable from Raw alone because their policy thresholds and text-generation branches were not yet established from retained data. Inspection of the current Apps Script then supplied the missing derivation rules: the 23.5-hour and 1,200-sample completeness thresholds, the current-day and invalid-gas-delta branches, and the deterministic note construction.

With those rules included, the audit covered all 21 columns. Across the 28 completed days, all 588 date-by-field comparisons were MATCH. There were zero mismatches, zero non-reconstructable fields, and no duplicate source keys in the audited reconstruction.

The planned selective repair therefore had no target. No production daily field was rewritten.

Conclusion

For the retained completed-day range from 2026-08-09 through 2026-09-05 JST, the stored NORITZ daily aggregate table was reproducible from retained one-minute Raw evidence plus the current derivation rules. All 588 audited values matched.

The hypothesis that historical daily values in this audited range had remained stale or inconsistent after the collector changes was not supported by the audit. This is a negative result, not a failed experiment: the audit established that the evidence-preserving action was a no-op.

A sample_count below 1,440 is not intrinsically erroneous. The implemented completeness rule for a completed past day uses both observed_hours >= 23.5 and sample_count >= 1200, while the stored count remains the actual admitted Raw observation count.

This result is bounded to the audited retained range. It does not establish that future daily rows can never become inconsistent, and the mismatch-repair branch was not exercised because no mismatch existed.

Evidence summary

The Evidence States for the information obtained in this Experiment are as follows.

See Evidence State for the shared definitions.

Experiment data

The one-minute Raw observations and stored daily household values remain retained but non-public. They are not required to be exposed in order to publish the aggregate audit result.

A sanitized machine-readable audit artifact contains only the audit period, aggregate counts, reconstruction thresholds, no-op outcome, privacy declarations, and limitations. It contains no household telemetry, spreadsheet identifiers, source keys, credentials, or installation-specific device/network identifiers.

AIEL-2026-0009 sanitized audit summary

The public artifact does not include the per-day gas, water, residual-hot-water, or operating-state values used in the private reconciliation.

Related experiments

AIEL-2026-0003 — NORITZ RC-G057MPW-2 hybrid water heater: investigating network access to heat-pump power and energy documents the collector environment used as the basis for this later data-integrity experiment.

Machine-readable experiment record

A machine-readable canonical record of this Experiment is published as JSON.

AIEL-2026-0009 experiment.json