CONSERVATIVE FORECAST DIAGNOSTICS FOR THE FINAL PAPER
25 October 2026

Scope and inputs
These calculations use the preserved event-derived paired_scores CSVs and
phase_events.csv in ERL_Full_Evaluation_Analysis_20261023.zip. They do not reread
the original raw JSONL streams or bind the inspected source to the historical
binary. The prior event audit supplies raw integrity/projection provenance;
the new validation JSON records hashes of the projections used here.

New exclusion, applied before computing either ERL or baseline error
The original projection contains 161,717 scored prediction/outcome pairs across
six missions. Sixty-six pairs have an outcome timestamp at or after the recorded
exit of the phase that owns their prediction: exactly one for every workload
switch. Two further moderate-replicate 1 pairs have prediction/application stamps
before that phase's recorded workload_enter. Excluding all 68 leaves 161,649 pairs.
All retained pairs satisfy prediction <= apply <= outcome and fit inside their
recorded phase bounds. The final phase has no workload_exit in this projection;
its terminal bound is left open, not invented. Its observations remain those in
the preserved recorded stream. This phase-based exclusion is conservative and
does not certify underlying source-frame or sensor interval membership.

The source review independently explains the transition risk: GA clears previous
action learning credit when committing the new workload, while the collector's
pilotPrevEpoch survives and may score the first incoming-workload observation
against the outgoing ticket. Thus an event called admission is not a verified
model-admission acknowledgement. The two pre-enter rows are conservatively
excluded because the GA can resume before the launcher records workload entry;
their exclusion is not a claim that they necessarily executed the wrong workload.

Recomputed baseline method
Use only retained valid throughput outcomes as baseline history. For each scored
query, the persistence baseline is the latest same-workload retained outcome
available strictly before the recorded prediction timestamp. The EMA baseline
uses alpha 0.30, starts from the first retained valid outcome, and carries its
workload-specific history across visits. Every baseline value and target is
selected without future observations. Four first-workload observations per run
have no prior baseline, leaving 161,625 common pairs. The same target rows are used
for all three predictors. The script asserts positive baseline age and that no
excluded observation is used in baseline history. No state from discarded
boundary outcomes is allowed to propagate into the reconstructed EMA.

Table-ready results: median absolute FPS error on identical comparable pairs
Profile      Rep  Pairs  ERL    Persistence  EMA 0.30
Aggressive    1  26935  3.091     5.449       4.722
Aggressive    2  26929  2.836     3.705       3.186
Inactive      1  26891  3.221     3.889       3.363
Inactive      2  26956  3.027     3.724       3.164
Moderate      1  26958  3.806     4.759       3.944
Moderate      2  26956  3.367     4.177       3.549

Suggested Methods paragraph
We treated forecast scoring as a retrospective diagnostic of the collector's
recorded target rather than as an intervention on the controller. We excluded
prediction/outcome pairs crossing a recorded workload-exit boundary and pairs
whose prediction/application preceded recorded workload entry (68 of 161,717
pairs). We then compared the logged ERL forecast with workload-specific
last-observation and exponential-moving-average forecasts (alpha=0.30), using
only clean observations whose availability timestamp preceded prediction.
The comparison contains 161,625 identical targets across the three predictors;
the baseline histories persist across workload visits. We report descriptive
median and mean absolute errors per mission without treating individual
controller periods as independent experimental replicates.

Suggested Results paragraph
The logged ERL forecasts had lower median absolute throughput error than both
retrospective baselines in each of the six missions (TableX). ERL MedAE ranged
from 2.836 to 3.806 FPS, compared with 3.705–5.449 FPS for persistence and 3.164–4.722 FPS
for the workload-specific EMA. This is evidence that the selected forecasts were
useful relative to these two simple predictors of the recorded target; it is not
evidence that retaining memory improved controller decisions. ERL and EMA
predictions coincide within 1e-6 in 7.8–49.0% of comparable pairs, consistent with
their related exponentially weighted mechanisms. Visit 3 MedAE was lower than
visit 1 in only one mission (aggressive replicate 2); there was no consistent
recurrence-associated reduction. At visits 2/3, recorded exact-bucket source
status covered 99.54–99.88% /99.69–100% of retained pairs. This describes the
reported prediction source, not per-target measured support or causal benefit.

Required limitations in or beside the above paragraphs
1. Every original outcome window is labeled assigned, and settings-change state
   is unknown(code 2). Structural pairing plus boundary exclusion does not prove
   that a later aggregate consists only of work performed after the assigned
   action, nor that power/temperature belong to that action interval.
2. Prediction values are frozen selected evaluations, while the issue timestamp
   is reconstructed from the selection boundary in the inspected source; it is
   not an independently instrumented prediction-completion instant. Historical
   source-to-binary identity remains unverified.
3. This is an exploratory post-hoc comparison on the actions the controller
   selected. The baselines were not deployed policies and cannot establish a
   scheduling advantage, counterfactual configuration accuracy, learning effect,
   positive transfer or memory-retention effect.
4. Only throughput is recommended as the main forecast result. Power/temperature
   diagnostics are included in CSVs for completeness but their freshness and
   smoothing dependencies are unresolved. All processing-latency scores were
   explicitly excluded(code 3): the predictor target is the historical E 2E
   channel, whereas the collector outcome is processing latency. Do not claim
   processing-latency predictive accuracy.
5. Exact-bucket support is shared across targets in the inspected source. It is
   not proof that every target was independently observed, calibrated or fresh.
6. There are only two independent missions/profile. Point estimates are
   descriptive; no frame-period p-values, significance or independent-sample
   confidence intervals are justified here.

Sensitivity
Requiring predictions to occur at least 1,5 or 10 seconds after recorded phase
entry changes each mission's full retained-set FPS MedAE by at most 0.0107 FPS.
This is a descriptive sensitivity check, not a guarantee of settled actuation.
All full-precision outputs are in final_forecasts_phase_margin_sensitivity.csv.

Files
final_forecasts_recompute.py: portable recalculation; --input and --output.
final_forecasts_validation.json: exact counts, projections' SHA 256 and limits.
final_forecasts_excluded_pairs.csv:68 individually identified exclusions.
final_forecasts_run.csv: retained-set FPS/power/temp summaries.
final_forecasts_visit.csv / workload_visit.csv: recurrence diagnostic details.
final_forecasts_recurrence.csv: visit 1/2/3 FPS MedAE per mission.
final_forecasts_baseline_run.csv / baseline_visit.csv: comparable-row summaries.
final_forecasts_coverage.csv: counts without same-workload prior observations.
final_forecasts_table.tex: compact main FPS table for direct inclusion.
final_forecasts_*_clean_pairs.csv.gz: retained baseline-comparable row evidence.

No production source or parent manuscript was modified.
