Eventdisplay CTAO Pipeline — Issue Report

# Eventdisplay CTAO Pipeline — Issue Report

**Repositories reviewed:** `Eventdisplay`, `Eventdisplay_AnalysisScripts_CTA`, `Eventdisplay_AnalysisFiles_CTA`, `Eventdisplay-ML`, `logFiles`

**Scope:** code correctness, algorithm consistency, and parameter propagation for the CTA IRF Monte Carlo production chain. Large auxiliary data products are not part of this checkout; physics-validity items are therefore assessed from code and log evidence rather than numerical output.

---

## Pipeline overview

| Step | Mode key | Entry script | Executable |
|------|----------|-------------|-----------|
| 1 | `EVNDISP` | `CTA.EVNDISP.sub_convert_and_analyse_MC_VDST_ArrayJob.sh` | `evndisp` |
| 2 | `DISPBDT` | `CTA.DISPTRAINING.sub_analyse.sh` | `trainTMVAforGammaHadronSeparation` |
| 3 | `MAKETABLES` | `CTA.MSCW_ENERGY.sub_make_tables.sh` | `mscw_energy -filltables=1` |
| 4 | `ANATABLES` | `CTA.MSCW_ENERGY.sub_analyse_MC.sh` | `mscw_energy` + `smoothLookupTables` |
| 5 | `XGBSTEREOTRAIN` / `XGBSTEREOANA` | `CTA.XGBSTEREO.sub_train/analyse.sh` | `eventdisplay-ml-train/apply-xgb-stereo` |
| 6 | `PREPARETMVA` | `CTA.prepareTMVA.sub_train.sh` | `WriteTrainingEvents` |
| 7 | `TRAIN` / `TRAIN_RECO_*` | `CTA.TMVA.sub_train.sh` | `trainTMVAforGammaHadronSeparation` |
| 8 | `ANGRES` / `QC` / `CUTS` | `CTA.EFFAREA.sub_analyse_list.sh` | `makeEffectiveArea` |
| 9 | `PHYS` | `CTA.WPPhysWriter.sub.sh` | `writeCTAWPPhysSensitivityFiles` |

Dispatcher: `Eventdisplay_AnalysisScripts_CTA/CTA.runAnalysis.sh`, invoked from `CTA.mainRunScriptsReduced.sh`.

---

## Severity definitions

| Label | Meaning |
|-------|---------|
| **Critical** | Logic inversion or memory error that corrupts results unconditionally |
| **High** | Bug that silently produces wrong physics output, crashes, or causes silent job loss under realistic inputs |
| **Medium** | Fragile code or parameter mismatch that can produce wrong results in specific but plausible conditions |
| **Low** | Metadata / diagnostic error with no direct physics impact |

---

## 1. Environment and orchestration

### ENV-1 · High — Hard-coded DESY AFS path makes the pipeline non-portable

- [x] No an issue and now changes required (we run these scripts only at DESY)

`setSoftwarePaths.sh` hard-codes a single ROOT installation path and derives all other paths from the current shell directory:

```sh
export ROOTSYS=/afs/ifh.de/group/cta/cta/software/root/root_v6.30.02.Linux-almalinux9.3-x86_64-gcc11.4/
TDIR=$(pwd)
export EVNDISPSCRIPTS="$TDIR"
export WORKDIR="${CTA_USER_WORK_DIR%/}/analysis/AnalysisData/${DSET}"
```

Source: `Eventdisplay_AnalysisScripts_CTA/setSoftwarePaths.sh:20-27`, `:36-45`.

**Impact:** running the same dataset on any other site or from a different checkout directory silently selects different software or fails late in the job.

### ENV-2 · High — Dataset-to-array mapping uses cascading `ARRAY=()` reassignments, leaving dead configurations

`CTA.runAnalysis.sh` maps dataset names to subarray lists via repeated unconditional assignment inside the same `if` branch. For example, seven consecutive `ARRAY=(...)` lines appear for `prod5` South:

```sh
ARRAY=( "subArray.prod5.South-BL.list" )
ARRAY=( "subArray.prod5.South-Alpha-2LSTs42SSTs.list" )
...
ARRAY=( "subArray.prod5.South-Betab.list" )
```

Source: `Eventdisplay_AnalysisScripts_CTA/CTA.runAnalysis.sh:255-267`.

**Impact:** only the last assignment survives. Earlier options are unreachable dead code and mislead anyone reconstructing which array list was used.

### ENV-3 · Medium — Versioning encoded in script literals

`ANADATE`, `EFFDATE`, `PHYSDATE`, and related IRF/TMVA version tags are string literals in `CTA.runAnalysis.sh`:

Source: `Eventdisplay_AnalysisScripts_CTA/CTA.runAnalysis.sh:66-76`, `:226-233`, `:293-298`.

**Impact:** reproducibility requires preserving the exact script revision; changing a single date literal redirects the whole chain to a different table/TMVA/IRF namespace without any structural safeguard.

---

## 2. Stage: EVNDISP

### EVN-1 · High — `continue` outside a loop terminates the EVNDISP branch abnormally

The EVNDISP dispatch block ends with:

```sh
cd ../
done
continue
```

Source: `Eventdisplay_AnalysisScripts_CTA/CTA.runAnalysis.sh:416-427`.

**Impact:** `continue` outside a loop is a POSIX shell error. The EVNDISP mode does not exit cleanly before worker jobs are submitted.

### EVN-2 · High — Input discovery is fragile when TMPDIR contains zero or multiple simtel files

The worker reconstructs simtel inputs with:

```sh
SIMFIL=`ls $TMPDIR/*.simtel.${EXTE}`
```

Source: `Eventdisplay_AnalysisScripts_CTA/analysis/CTA.EVNDISP.qsub_convert_and_analyse_MC_VDST_ArrayJob.sh:129-162`.

**Impact:** if the glob matches zero or multiple files, the converter call becomes ambiguous. Batches of files are copied into a shared `$TMPDIR` before conversion, making multiple matches plausible.

### EVN-3 · Medium — Extensive unquoted path handling makes file operations shell-fragile

Representative examples:

- `cp -v -f $F $TMPDIR"/"` — `CTA.EVNDISP.qsub_convert_and_analyse_MC_VDST_ArrayJob.sh:77-95`
- `cp $F $DDIR/` — `CTA.MSCW_ENERGY.qsub_analyse_MC.sh:74-77`
- `sed -n "$l,$k p" $TFILE > $IFIL` — `CTA.MSCW_ENERGY.qsub_analyse_MC.sh:45`

**Impact:** paths containing spaces or shell metacharacters break silently. Current DESY-style paths avoid this in practice but the code is brittle.

### EVN-4 · Medium — Missing low-gain calibration multipliers observed in production logs

- [x] Not an issue. Low-gain calibration is not relevant in this context.

From `logFiles/evndisp.log:646-652`:

```text
VEventLoop::analyzeEvent warning: No low gain multipliers available for sumwindow 1 (3 samples); trace integration method 2
```

**Impact:** reconstruction proceeded with incomplete calibration; trace integration method 2 uses a missing conversion coefficient for low-gain channels.

### EVN-5 · Medium — Low-gain average-T0 assignment always overwrites the high-gain field

- [x] Not an issue. Low-gain calibration is not relevant in this context.

`VCalibrationData::setAverageTZero()` is missing an `else` branch:

```cpp
void VCalibrationData::setAverageTZero(double iAverageTzero, bool iLowGain)
{
    if(iLowGain)
    {
        fAverageTZero_lowgain = iAverageTzero;
    }
    fAverageTZero_highgain = iAverageTzero;   // executes unconditionally
}
```

Source: `Eventdisplay/src/VCalibrationData.cpp:380-388`.

**Impact:** every low-gain T0 calibration call corrupts `fAverageTZero_highgain`, invalidating the high-gain timing constants for any subsequent event.

### EVN-6 · Medium — Low-gain average-T0 calculation uses the high-gain charge extraction path

- [x] Not an issue. Low-gain calibration is not relevant in this context.

`VCalibrator::calculateAverageTZero(bool iLowGain)` passes `iLowGainOnly=false` to `calcSums` unconditionally:

```cpp
calcSums(fRunPar->fCalibrationSumFirst,
         fRunPar->fCalibrationSumFirst + fRunPar->fCalibrationSumWindowAverageTime,
         false, false, 2);   // second false = iLowGainOnly
```

Source: `Eventdisplay/src/VCalibrator.cpp:868`.

**Impact:** low-gain T0 constants are derived from charges integrated in high-gain mode; the resulting timing calibration is wrong for low-gain channels.

---

## 3. Stage: Lookup tables and stereo reconstruction

### TAB-1 · High — Table-filling and analysis stages use different stereo-angle cuts

- [x] Note an issue (as long as table filling uses a smaller angle than the analysis.

`MAKETABLES` fixes the stereo-angle cut at 10 degrees:

```sh
MOPT="$MOPT -redo_stereo_reconstruction ... -minangle_stereo_reconstruction=10"
```

Source: `Eventdisplay_AnalysisScripts_CTA/analysis/CTA.MSCW_ENERGY.qsub_make_tables.sh:87-90`.

`ANATABLES` uses 15 degrees for 2-telescope arrays and 5 degrees otherwise:

```sh
if [[ $NTEL == "2" ]]; then
    MOPT="$MOPT -minangle_stereo_reconstruction=15."
else
    MOPT="$MOPT -minangle_stereo_reconstruction=5."
fi
```

Source: `Eventdisplay_AnalysisScripts_CTA/analysis/CTA.MSCW_ENERGY.qsub_analyse_MC.sh:133-138`.

**Impact:** the lookup tables are filled under a different stereo phase space than the analysis stage that consumes them. This biases mean-scaled parameters and energy reconstruction, especially for low-multiplicity configurations where the 10° vs 5° difference is largest.

### TAB-2 · High — `CTA.MSCW_ENERGY.sub_make_tables.sh` uses an undefined variable `MEANDIST` in generated filenames

```sh
FNAM="$SHELLDIR/EMSCW.table-$TAFIL-W$MEANDIST-${ARRAY}${AZ}"
```

`MEANDIST` is never set anywhere in the script.

Source: `Eventdisplay_AnalysisScripts_CTA/analysis/CTA.MSCW_ENERGY.sub_make_tables.sh:99-105`.

**Impact:** job and script names are unstable (expand to an empty-token form). Name collisions are possible if different invocations happen to produce the same effective filename prefix.

### TAB-3 · Medium — `CTA.MSCW_ENERGY.sub_analyse_MC.sh` uses an undefined variable `NC` in output names

```sh
TFIL=$PART$NC"."$SUBAR"_ID${RECID}_${MCAZ}-"$DSET
```

`NC` is never assigned in the script.

Source: `Eventdisplay_AnalysisScripts_CTA/analysis/CTA.MSCW_ENERGY.sub_analyse_MC.sh:121-123`.

**Impact:** the naming convention is broken by a stale variable from an earlier script revision. The empty expansion risks latent name collisions if `NC` becomes set in the environment.

### TAB-4 · Medium — DISP analysis and training cuts are synchronized only by a comment

```sh
# IMPORTANT: this must be the same or lower value as in dispBDT training
if [[ $RECID == "1" ]]; then
    MOPT="$MOPT -maxloss=0.1 -minfui=0."
```

Source: `Eventdisplay_AnalysisScripts_CTA/analysis/CTA.MSCW_ENERGY.qsub_analyse_MC.sh:139-148`.

**Impact:** a training/analysis cut mismatch is not detected automatically and can alter DISP validity or reconstruction quality silently.

### TAB-5 · Medium — `VTableLookupRunParameter` contains contradictory CLI validation

- For `-fill`, the error text documents `2=read lookup tables`, but the code only accepts `0` or `1`. Source: `Eventdisplay/src/VTableLookupRunParameter.cpp:168-183`.
- For `-tmva_nimages_max_stereo_reconstruction`, the code rejects only values `> 40000` while the error message states the maximum is 4. Source: `Eventdisplay/src/VTableLookupRunParameter.cpp:205-214`.

**Impact:** misleading or stale CLI validation on the `mscw_energy` path makes misconfiguration easy to miss.

---

## 4. Stage: DISP training

### DISP-1 · High — `CTA.DISPTRAINING.sub_analyse.sh` argument parsing is broken in two ways

**Argument-position error:** both `TMVAQC` and `QSUBOPT` are gated on `$8` rather than `$7` and `$8` respectively:

```sh
if [ -n $8 ]
then
   TMVAQC="$7"    # should be guarded by [ -n "$7" ]
fi
if [ -n $8 ]
then
   QSUBOPT="$8"
fi
```

Source: `Eventdisplay_AnalysisScripts_CTA/analysis/CTA.DISPTRAINING.sub_analyse.sh:57-64`.

**Unquoted guards always true:** all three positional guards (`$6`, `$8`, `${9}`) are unquoted. In bash, `[ -n $UNSET ]` reduces to `[ -n ]`, which is always true because `-n` is itself non-empty. The guards never detect an absent argument.

Source: lines 52, 57, 61, 91.

**Consequence of empty TMVAQC:** line 72 runs `QCA=\`cat $TMVAQC\``. When `TMVAQC` is the empty string, `cat` reads from stdin and hangs, or returns empty; the inner submission loop `for QC in ${QCA}` then iterates zero times and **no DISP jobs are submitted**.

**Impact:** whenever the quality-cut file argument is absent, the DISP training stage silently submits no jobs. Combined with the unquoted guards, this failure is invisible without explicit log inspection.

### DISP-2 · High — DISP training is submitted for telescope types absent from the selected array (confirmed by logs)

Production logs show repeated training failures:

```text
total number of telescopes: 30 (selected 0)
..nothing to do. Exiting.
```

Source: `logFiles/BDTDisp-10408618.training.log:46-48`, `BDTDisp-201109916.training.log:46-48`, `BDTDisp-205008707.training.log:46-48`.

Successful training occurred only for types `10608418` and `138704810`, which are the only types present in the prod6 North subarray (confirmed by `logFiles/makeTable.log:60-95` and `logFiles/mscwTable.log:181-218`).

**Impact:** telescope-type enumeration in the DISP dispatch loop is inconsistent with the array composition. Training jobs are wasted for every absent type; the pipeline provides no error signal for this condition.

### DISP-3 · Medium — Failed telescope-type training is detected only by log-text grep

The qsub worker decides whether training was meaningful by searching for the literal string `Number of telescope types: 0` in the log:

Source: `Eventdisplay_AnalysisScripts_CTA/analysis/CTA.DISPTRAINING.qsub_analyse.sh:79-88`.

**Impact:** failure detection depends on exact log wording. Any upstream message change silently defeats this safeguard.

---

## 5. Stage: TMVA preparation and gamma/hadron training

### TMVA-1 · High — `CTA.prepareTMVA.sub_train.sh` mis-parses optional arguments

The documented interface is `[qsub options] [direction] [job_dir]`, but the implementation assigns:

```sh
MCAZ=${5:-$MCAZ}
QSUBOPT=${5:-$QSUBOPT}
LDIR=${6:-$LDIR}
```

Source: `Eventdisplay_AnalysisScripts_CTA/analysis/CTA.prepareTMVA.sub_train.sh:59-68`.

Argument `$5` is assigned to both `MCAZ` (azimuth direction) and `QSUBOPT` (batch options). The documented direction slot (`$6`) is used as the log directory.

**Impact:** with an empty `QSUBOPT`, the azimuth direction can slide into the batch-options slot and vice versa, silently selecting the wrong input path pattern (`gamma_cone.*ID$RECID$MCAZ*.mscw.root`) and corrupting submission options.

### TMVA-2 · High — Gamma/hadron BDT energy bins 0 and 1 overlap completely (bin 0 ⊂ bin 1)

The energy bin arrays in `CTA.TMVA.sub_train.sh` are:

```sh
EMIN=( -1.90 -1.90 -1.45 -1.20 -0.95 -0.50 -0.10  0.45  0.90 )
EMAX=( -1.40 -1.30 -1.15 -0.80 -0.25  0.25  0.75  1.50  2.50 )
```

Source: `Eventdisplay_AnalysisScripts_CTA/analysis/CTA.TMVA.sub_train.sh:84-85`.

Bin 0 spans log₁₀(E/TeV) = −1.90 to −1.40 (12.6–39.8 GeV). Bin 1 spans −1.90 to −1.30 (12.6–50.1 GeV). Bin 0 is a strict subset of bin 1; both share the same lower energy edge and train on largely identical event populations.

At application time `VTMVAEvaluator::getDataBin()` selects the nearest spectral-weighted mean energy bin. Because bins 0 and 1 share the same lower edge, their spectral-weighted means are nearly identical, and assignment of low-energy events to one or the other is effectively arbitrary.

Source: `Eventdisplay/src/VTMVAEvaluator.cpp:1120-1155`.

**Impact:** two BDTs are trained on near-identical event sets and applied with near-identical selection criteria. No exclusive low-energy classifier exists for the 12.6–40 GeV range. If the intended configuration was two adjacent bins (e.g., −1.90 to −1.40 and −1.40 to −1.30), the lower edge of bin 1 should be −1.40, not −1.90.

### TMVA-3 · High — TMVA training aborts entirely if the first expected training file is missing

`CTA.TMVA.sub_train.sh` exits immediately when `MVA${MCAZ}-${RECID}-${OFFMEA[0]}.training.root` is absent, with no per-file retry or diagnostic:

Source: `Eventdisplay_AnalysisScripts_CTA/analysis/CTA.TMVA.sub_train.sh:123-129`.

**Impact:** one missing training product blocks all remaining offset/energy bins for the affected array and run mode.

### TMVA-4 · Medium — Unquoted positional-parameter checks are always true in `CTA.WPPhysWriter.sub.sh`

```sh
if [ -n $9 ]; then
if [ -n ${10} ]; then
if [ -n ${11} ]; then
```

Source: `Eventdisplay_AnalysisScripts_CTA/analysis/CTA.WPPhysWriter.sub.sh:36-52`.

**Impact:** these guards always evaluate true in bash regardless of whether the arguments are set. Optional late arguments are therefore always consumed.

### TMVA-5 · Medium — Diagnostic artifacts deleted unconditionally after training

`CTA.TMVA.qsub_train.sh` deletes all generated `.C` files and the entire `complete_BDTroot` directory after training:

Source: `Eventdisplay_AnalysisScripts_CTA/analysis/CTA.TMVA.qsub_train.sh:25-31`.

**Impact:** post-mortem analysis of suspicious classifier behavior is not possible without re-running training.

---

## 6. Stage: XGB stereo reconstruction

### XGB-1 · High — XGB apply wrapper can load a different model than training produced

Training writes models under a prefix that includes the requested minimum-image count:

```sh
PREFIX="${ODIR}/dispdir_bdt_mintel${MINTEL}"
```

Source: `Eventdisplay_AnalysisScripts_CTA/analysis/CTA.XGBSTEREO.qsub_train.sh:46`.

The apply wrapper silently caps `MINTEL` at 3 before constructing the same prefix:

```sh
[ "$MINTEL" -ge 3 ] && MINTEL=3
PREFIX="${XGBDIR}/dispdir_bdt_mintel${MINTEL}"
```

Source: `Eventdisplay_AnalysisScripts_CTA/analysis/CTA.XGBSTEREO.qsub_analyse.sh:42-46`.

**Impact:** for any training run with `NIMAGESMIN > 3`, the apply step loads `dispdir_bdt_mintel3_ebin*.joblib.gz` even though training produced `dispdir_bdt_mintel4` (or higher). This is a silent model-path mismatch.

### XGB-2 · High — XGB classification is not wired into the orchestration layer

The script repository contains sub/qsub wrappers for XGB stereo regression training and application. No corresponding wrappers for XGB classification (`train-xgb-classify`, `apply-xgb-classify`) exist in `Eventdisplay_AnalysisScripts_CTA`.

**Impact:** gamma/hadron separation in the shell pipeline relies entirely on the TMVA path. The XGB classification capabilities of `Eventdisplay-ML` are unavailable through the current orchestration.

### XGB-3 · High — `configure_train()` treats `energy_bins_log10_tev` as a dict, but loaders treat it as a list of per-bin dicts

```python
model_configs["pre_cuts"] = pre_cuts_classification(
    e_min=np.power(10.0, model_parameters.get("energy_bins_log10_tev", []).get("E_min")),
    e_max=np.power(10.0, model_parameters.get("energy_bins_log10_tev", []).get("E_max")),
)
```

Source: `Eventdisplay-ML/src/eventdisplay_ml/config.py:162-165`.

The loader and updater logic in `utils.py:109-125` and `models.py:189-202` treat the same field as a list of per-bin dicts.

**Impact:** configuration shape mismatch causes a runtime `AttributeError` (`list` has no `.get()`) or silently applies incorrect energy pre-cuts before classification training.

### XGB-4 · Medium — Training wrapper bypasses the external hyperparameter-config hook

The XGB CLI supports JSON override of hyperparameters via `--hyperparameter_config`. The CTA training wrapper never passes this option:

Source: `Eventdisplay_AnalysisScripts_CTA/analysis/CTA.XGBSTEREO.qsub_train.sh:56-63`.

**Impact:** production XGB training cannot be controlled through `Eventdisplay_AnalysisFiles_CTA/ParameterFiles` or the package's own JSON mechanism without editing the wrapper. Hyperparameters are locked to Python-package defaults.

### XGB-5 · Medium — `CTA.XGBSTEREO.sub_train.sh` advertises a direction argument it never reads

The usage string documents `[direction (e.g. _180deg)]` as argument `$6`, but the implementation only reads `$5` as `QSUBOPT` and `$6` as `LDIR`; no azimuth/direction variable is used anywhere:

Source: `Eventdisplay_AnalysisScripts_CTA/analysis/CTA.XGBSTEREO.sub_train.sh:16-22`, `:48-56`.

**Impact:** callers believe they are training a direction-specific model while all input paths are direction-independent. Azimuth-mixing in XGB stereo models is invisible.

### XGB-6 · Low — `DL2.production.runparameter` is a legacy TMVA example

The shipped file references TMVA products rather than the current XGB production path:

```text
* TMVAPARAMETER noepoch BDT ...
```

Source: `Eventdisplay_AnalysisFiles_CTA/ParameterFiles/DL2.production.runparameter:4-7`.

**Impact:** the file is not an authoritative reference for the current ML production configuration and can mislead users reconstructing the analysis.

---

## 7. Stage: IRF calculation and PHYS writing

### IRF-1 · High — `PHYS` passes a filename with a trailing space

```sh
./CTA.WPPhysWriter.sub.sh \
        "$NFILARRAY "\
        ${EFFFULLDIR}/BDT."$OOTIME"-${EFFVERSION}.$EFFDATE \
```

Source: `Eventdisplay_AnalysisScripts_CTA/CTA.runAnalysis.sh:754-756`.

**Impact:** the subarray-list argument has a spurious trailing space. Depending on the tool's path handling, this can make the wrapper fail to locate the intended file.

### IRF-2 · High — Run-parameter key `THETA2MINENEERGY` appears misspelled

```sh
echo "THETA2MINENEERGY 1." >> "$PARA"
```

Source: `Eventdisplay_AnalysisScripts_CTA/CTA.runAnalysis.sh:590-597`.

The correct key is likely `THETA2MINENERGY`. `VInstrumentResponseFunctionRunParameter` silently ignores unknown keys (see LIB-4 below).

**Impact:** the θ² minimum-energy threshold may be silently dropped, allowing events below the intended energy threshold to enter the IRF calculation.

### IRF-3 · Medium — Effective-area success is inferred from file size rather than content

```sh
minimumsize=300
DS=$(du -k $OFIX.root | cut -f 1)
if [[ ${DS} -le $minimumsize ]]; then
    touch $OLOG.SMALLFILE
    mv -v $OFIX.root ${ODIR}/
    continue
fi
```

Source: `Eventdisplay_AnalysisScripts_CTA/analysis/CTA.EFFAREA.qsub_analyse_list.sh:581-607`.

**Impact:** a semantically broken but large IRF file passes validation; a valid but compact file fails. This is weak validation for a stage that produces final IRF products.

### IRF-4 · Medium — Summary-tree builder is hard-wired to one archived directory convention

`prepareSummaryTrees.sh` unconditionally overrides `PHYSDIR` immediately after setting it:

```sh
PHYSDIR="$DATADIR/Phys-${DDAT}/"
PHYSDIR="$DATADIR/archive.202105/Phys-${DDAT}-${S}/"
```

Source: `Eventdisplay_AnalysisScripts_CTA/analysis/prepareSummaryTrees.sh:31-34`.

**Impact:** the nominal output path is never used; only the hardwired archive path is accessible.

---

## 8. C++ library: signal extraction (`VTraceHandler`)

### SIG-1 · High — Sliding-window raw mode returns the wrong quantity

In `calculateTraceSum_slidingWindow(..., fRaw=true)`, the function accumulates raw tail samples into `charge` but returns `FADC[1]` instead:

```cpp
for( unsigned int i = 0; i < ( unsigned int )iIntegrationWindow; i++ )
{
    charge += ( float )fpTrace.at( n - 1 - i );
}
return FADC[1];
```

Source: `Eventdisplay/src/VTraceHandler.cpp:643-654`.

**Impact:** trace integration method 2 in raw mode returns a single FADC sample rather than the integrated charge. Signal estimates are physically wrong for this combination.

### SIG-2 · High — Sliding-window loop can access one sample past the trace end

The search range upper bound is:

```cpp
if( ( n - window ) <= SearchEnd )
    SearchEnd = n - window + 1;
```

The update step is then:

```cpp
xmax = xmax - FADC[i] + FADC[i + window];
```

Source: `Eventdisplay/src/VTraceHandler.cpp:611-615`, `:663-680`.

At the final iteration, `i + window == n`, which is one past the valid `FADC[0..n-1]` range.

**Impact:** undefined memory access at trace edges. This can perturb integrated charge and timing values and may crash on short traces.

### SIG-3 · Medium — Low-gain saturation threshold computed from uninitialized `tmax`

`getTraceMax()` derives the saturation count threshold before resetting `tmax`:

```cpp
unsigned int nMax = ( unsigned int )( fDynamicRange * tmax );
n255 = 0;
tmax = -10000.;
```

Source: `Eventdisplay/src/VTraceHandler.cpp:391-395`.

**Impact:** `nMax` is derived from the previous call's `tmax`, not the current trace maximum. Low-gain saturation diagnostics (`n255`) are unreliable.

### SIG-4 · Low — Muon-ring correction divides by zero at the first grid point

The muon-ring image-size correction loop computes `xi_tmp = (float)i / (float)numSteps` and immediately divides by it before any guard:

```cpp
for(int i = 0; i < numSteps; i++) {
    float xi_tmp = (float)i / (float)numSteps;   // zero when i == 0
    float kTest_tmp = 2.0 * ngExi[i] * kRatio / xi_tmp;  // div-by-zero
```

Source: `Eventdisplay/src/VImageParameterCalculation.cpp:744-745`.

**Impact:** muon calibration produces `inf`/`NaN` at the first grid point and an unstable correction factor. Muon ring analysis is not part of the standard CTA MC-only IRF production, but the defect exists in the shared code and would affect any CTA run that enables muon calibration.

---

## 9. C++ library: stereo reconstruction

### STEREO-1 · Critical — `VStereoReconstruction::removeDataSet` has inverted logic and always reports failure

```cpp
bool VStereoReconstruction::removeDataSet( unsigned int iDataSet )
{
    if( fData.size() < iDataSet )    // removes out-of-range indices, not valid ones
    {
        fData.erase( fData.begin() + iDataSet );
    }
    return false;    // always signals failure
}
```

Source: `Eventdisplay/src/VStereoReconstruction.cpp:289-295`.

**Impact:** valid dataset indices are never removed; out-of-range indices can be erased (undefined behaviour); the function unconditionally reports failure to the caller. Any code path relying on this method to manage reconstruction datasets is broken.

### STEREO-2 · High — `VOnOff` mutates the OFF histogram during on-off subtraction

```cpp
hTemp->Add( hon, hoff, 1., -1.*i_norm_alpha );
hoff->Scale( i_norm_alpha );
```

Source: `Eventdisplay/src/VOnOff.cpp:149-153`.

**Impact:** `hoff` is scaled in-place after being used in the subtraction. Callers expecting the original OFF histogram to remain unchanged reuse a scaled copy, producing wrong background normalization.

### STEREO-3 · High — Method-4 stereo pair-angle rejection evaluates the wrong telescope slopes

In the method-4 all-pair weighted intersection loop, the current pair slopes are loaded into `mm[0]`/`mm[1]` (`m[ii]`/`m[jj]`), but the angle-rejection cut uses `m[0]`/`m[1]` — the slopes of the first two telescopes, not the current pair:

```cpp
mm[0] = m[ii];
mm[1] = m[jj];
rcs_perpendicular_fit(xx, yy, ww, mm, 2, &xs, &ys, &stds);

float i_diff = fabs(atan(m[0]) - atan(m[1]));   // should use mm[0], mm[1]
if(i_diff < fAxesAngles_min / TMath::RadToDeg() ...) continue;
```

Source: `Eventdisplay/src/VArrayAnalyzer.cpp:1339-1354`.

**Impact:** for events with more than two images (routine in CTA multi-telescope arrays), the parallel-image cut is applied to the first telescope pair for every pair in the loop, not to the actual current pair. Valid pairs whose slopes are nearly parallel to telescopes 0 and 1 are accepted; pairs with good angular separation but small `m[0]`–`m[1]` angle are discarded. The weighted direction estimate is biased across the full multi-telescope sample.

### STEREO-4 · Medium — Emission-height pair weights undefined for small image sizes

The pairwise emission-height weight formula `1./((1./log10(size[i])) + (1./log10(size[j])))` is not guarded against `size <= 1`:

```cpp
iEmissionHeightWeightTemp = 1. / ((1. / log10(size[i])) + (1. / log10(size[j])));
```

Source: `Eventdisplay/src/VEmissionHeightCalculator.cpp:61`.

**Impact:** an image with `size == 1` makes `log10(1) = 0`, producing a division by zero; `size < 1` produces a negative weight. Both yield NaN/inf or ill-defined emission-height estimates for low-intensity images that pass the size threshold.

### STEREO-5 · Medium — `VEmissionHeightCalculator` accumulates telescope positions across reinitialisations

Both `setTelescopePositions()` overloads only `push_back()` and never `clear()` the position vectors:

```cpp
void VEmissionHeightCalculator::setTelescopePositions(vector<float> x, ...)
{
    fNTel = x.size();
    for(unsigned int i = 0; i < fNTel; i++)
    {
        fTelX.push_back(x[i]);  // no clear() before the loop
        ...
    }
}
```

Source: `Eventdisplay/src/VEmissionHeightCalculator.cpp:100-128`.

**Impact:** reusing the same calculator instance (e.g., across array configurations or runs) doubles the telescope-position list, making all pairwise baseline distances and emission-height geometry wrong from the second call onward.

---

## 10. C++ library: TMVA training

### ML-1 · Critical — `intersect_mc_error` ignores the Y-component due to a typo

```cpp
intersect_mc_error = sqrt( (Xoff_intersect-MCxoff)*(Xoff_intersect-MCxoff)
                         + (Yoff_intersect-Yoff_intersect)*(Yoff_intersect-Yoff_intersect) );
```

Source: `Eventdisplay/src/trainTMVAforGammaHadronSeparation.cpp:344-346`.

The Y term is `(Yoff_intersect - Yoff_intersect)²`, which is always zero. The intersection angular error is computed from the X offset only.

**Impact:** the angular-reconstruction method selector (`TrainAngularReconstructionMethod`) trains labels that compare DISP against a systematically wrong intersection error whenever the Y component is non-negligible. This biases method-selection in the active angular training stage.

### ML-2 · High — Zenith-bin range check is off by one

```cpp
if( iRun->fZenithCutData.size() < iZenithBin || iRun->fOutputFile[0].size() < iZenithBin )
```

Source: `Eventdisplay/src/trainTMVAforGammaHadronSeparation.cpp:438-441`.

The condition uses `<` instead of `<=`. Index `iZenithBin == size()` passes the guard even though it is out of range.

**Impact:** the subsequent array access at `[iZenithBin]` is out-of-bounds and invokes undefined behaviour.

### ML-3 · High — `get_largest_weight_disp_entries()` sorts ascending, selecting the smallest-weight entries

The function name implies selecting the largest-weight DISP images, but the comparator sorts ascending:

```cpp
sort(indexed_disp_woff.begin(), indexed_disp_woff.end(),
     [](const pair<Float_t, unsigned int>& a, const pair<Float_t, unsigned int>& b) {
         return a.first < b.first;
     });
```

Source: `Eventdisplay/src/trainTMVAforGammaHadronSeparation.cpp:96-101`.

**Impact:** the per-event telescope-image subset selected for BDT features is the lowest-weight group, contradicting the intended selection. This corrupts TMVA feature construction for multi-telescope events.

### ML-4 · High — Per-telescope feature remapping writes into its own source arrays

When `DispNImages < NImages`, the code remaps arrays in place:

```cpp
for( unsigned int i = 0; i < matching_indexes.size(); i++ )
{
    size[i] = size[matching_indexes[i]];
    width[i] = width[matching_indexes[i]];
    length[i] = length[matching_indexes[i]];
    ...
}
```

Source: `Eventdisplay/src/trainTMVAforGammaHadronSeparation.cpp:103-117`.

**Impact:** early writes overwrite values still needed by later iterations. Per-telescope BDT feature vectors are corrupted for events with missing DISP entries.

### ML-5 · Medium — Reduced training tree stores a reconstruction-error feature under a misleading branch name

The helper `get_largest_weight_disp_entries()` fills an energy-error array named `d_erec_mc`:

```cpp
d_erec_mc[i] = (ES[original_index] - ErecS) / ES[original_index];
```

Source: `Eventdisplay/src/trainTMVAforGammaHadronSeparation.cpp:141-143`.

But the reduced tree declares and writes this quantity as `d_erec_es`:

```cpp
Float_t d_erec_es[d_n];
iDataTree_reduced->Branch( "d_erec_es", d_erec_es, htemp );
get_largest_weight_disp_entries(..., d_erec_es, d_R)
```

Source: `Eventdisplay/src/trainTMVAforGammaHadronSeparation.cpp:248`, `:285-286`, `:352`.

**Impact:** the branch name does not match the quantity stored. Any TMVA configuration referencing `d_erec_es` or `d_erec_mc` by name may apply the variable to incorrect physics context.

### ML-6 · Medium — `TrainAngularReconstructionMethod` splits both classes from the same signal tree

Both the "prefer DISP" and "prefer intersection" training classes are extracted from `iSignalTree_reduced`:

```cpp
TTree* sigTree = iSignalTree_reduced->CopyTree(signalCut);
TTree* bkgTree = iSignalTree_reduced->CopyTree(backgrCut);
```

Source: `Eventdisplay/src/trainTMVAforGammaHadronSeparation.cpp:596-602`.

**Impact:** this is the correct design for an angular-method selector (all events are gammas; the label encodes which reconstructor is more accurate). However, combined with the broken `intersect_mc_error` calculation (ML-1), the labels are wrong. The two issues compound.

---

## 11. C++ library: IRF and effective area

### IRF-LIB-1 · High — `makeEffectiveArea` documents the MC header as optional, but the library hard-exits if it is missing

```cpp
// read MC header (might not be there, no problem; ...)
VMonteCarloRunHeader* iMonteCarloHeader = fRunPara->readMCRunHeader();
```

Source: `Eventdisplay/src/makeEffectiveArea.cpp:200-201`.

Inside `readMCRunHeader()`:

```cpp
if( !iMC )
{
    cout << "VInstrumentResponseFunctionRunParameter::readMCRunHeader: no MC run header found ...";
    exit( EXIT_FAILURE );
}
```

Source: `Eventdisplay/src/VInstrumentResponseFunctionRunParameter.cpp:570-577`.

**Impact:** the executable-level comment and the library behaviour disagree. Runs the caller describes as tolerable still terminate with a hard exit.

### IRF-LIB-2 · High — `VInstrumentResponseFunctionReader` disables true-energy-axis handling unconditionally

```cpp
bXaxisIsEtrue = false;
if( bXaxisIsEtrue ) { ... } else { ... }
```

Source: `Eventdisplay/src/VInstrumentResponseFunctionReader.cpp:902-910`.

**Impact:** the code path for migration matrices with true energy on the x-axis is unreachable. Any caller expecting Etrue-axis handling silently receives reconstructed-energy handling.

### IRF-LIB-3 · High — CTA IRF reader compensates for inconsistent histogram naming and units in the files it reads

```cpp
// name and axis units are not consistent in the CTA files!!!
```

The reader then falls back through multiple alternative histogram names (`EffectiveArea`, `EffectiveAreaEtrue`, `harea_gamma`, `ERes`, `EnResol_RMS`, `Ebias`…).

Source: `Eventdisplay/src/VInstrumentResponseFunctionReader.cpp:185-265`.

**Impact:** the reader infers histogram semantics from inconsistent naming conventions. IRF reuse across production versions depends on fallback heuristics that can silently pick the wrong histogram.

### IRF-LIB-4 · High — `VInstrumentResponseFunctionRunParameter` silently ignores unknown runparameter keys

The parser handles a known list of `* KEY ...` directives with no final `else` branch:

```cpp
if( temp == "ENERGYSPECTRUMINDEX" ) { ... }
else if( temp == "MONTECARLOENERGYRANGE" ) { ... }
...
else if( temp == "WOBBLEISOTROPIC" ) { ... }
// unknown keys fall through silently
```

Source: `Eventdisplay/src/VInstrumentResponseFunctionRunParameter.cpp:109-462`.

**Impact:** misspelled or stale script-generated keys (e.g., `THETA2MINENEERGY`, see IRF-2) are silently ignored. Configuration drift between scripts and C++ parser goes undetected.

### IRF-LIB-5 · Medium — IRF binning defaults are duplicated by contract between two classes

`VEffectiveAreaCalculator` documents that changing default energy/offset axes requires synchronized changes in `VEffectiveAreaCalculatorMCHistograms`:

Source: `Eventdisplay/src/VEffectiveAreaCalculator.cpp:53-60`.

**Impact:** this is a maintainability hazard; mismatched changes alter IRF shapes without compile-time protection.

### IRF-LIB-6 · Medium — TMVA XML parsing depends on exact string layout

`VTMVAEvaluator::getTrainingVariables` parses XML weight files using substring offsets around `NVar="`, `Expression="`, and `Label=`:

Source: `Eventdisplay/src/VTMVAEvaluator.cpp:95-151`.

**Impact:** TMVA format changes or slightly different XML serialization silently break variable extraction.

### IRF-LIB-7 · Medium — TMVA weight grid silently passes initialisation with missing interior bins

`VTMVAEvaluator::initializeWeightFiles()` emits a warning and continues when interior energy or zenith bins are missing, allowing an incomplete grid to be used in production:

```cpp
// allow that first files are missing (this happens when there are no training events in the first energy bins)
if(!bGoodRun) {
    if(i == iMinMissingBin || j == jMinMissingBin) {
        cout << "warning: TMVA root file not found ..." << endl;
        continue;   // incomplete grid proceeds
    }
```

Source: `Eventdisplay/src/VTMVAEvaluator.cpp:275-320`.

**Impact:** missing interior bins are filled by the nearest-neighbour fallback in `getDataBin()`; events in those bins are evaluated with an incorrect classifier. There is no hard failure that would flag the incomplete IRF set.

### IRF-LIB-8 · Medium — `VInstrumentResponseFunction` mutates the shared MC azimuth in place

MC azimuth normalisation is applied by modifying the shared event object directly:

```cpp
if(fData->MCaz > 180.) {
    fData->MCaz -= 360.;    // in-place mutation of shared state
}
```

Source: `Eventdisplay/src/VInstrumentResponseFunction.cpp:258-260`.

**Impact:** this mutation occurs inside the azimuth-bin loop. After the first iteration the `MCaz` value is already normalised; subsequent iterations in any outer loop (e.g., spectral index) operate on the already-modified value. If the normalisation is applied more than once, `MCaz` can drift below −180°, causing all azimuth-bin membership tests to fail.

### IRF-LIB-9 · Medium — `makeEffectiveArea` IRF resolution duplication is order-dependent

Resolution products (e.g., angular resolution, core resolution) are filled by looking up the source IRF by vector index via `getDuplicationID()`:

```cpp
else if(f_IRF[i]->getDuplicationID() < f_IRF.size()
        && f_IRF[f_IRF[i]->getDuplicationID()])
{
    f_IRF[i]->fillResolutionGraphs(f_IRF[f_IRF[i]->getDuplicationID()]->getIRFData());
}
```

Source: `Eventdisplay/src/makeEffectiveArea.cpp:297-300`.

**Impact:** this only works correctly if the source IRF has been initialised earlier in `f_IRF`. Any refactor changing the IRF vector ordering or adding new IRF types silently breaks resolution product generation.

### IRF-LIB-10 · Medium — `copyMCHistograms()` leaves the first ROOT input file open

The helper function opens each input file to accumulate MC histograms. For the first file (`z == 0`) it borrows the histogram pointer directly and never closes the file:

```cpp
TFile* ifInput = new TFile(chEl->GetTitle());
if(z == 0) {
    iMC_his = (VEffectiveAreaCalculatorMCHistograms*)ifInput->Get("MChistos");
    // ifInput is never closed
} else {
    iMC_his->add(...);
    ifInput->Close();   // only subsequent files are closed
}
```

Source: `Eventdisplay/src/makeEffectiveArea.cpp:567-579`.

**Impact:** the first input file handle is leaked for the lifetime of the returned histogram object; in large IRF productions with many input files this contributes to file-descriptor exhaustion.

---

## 12. C++ library: `writeCTAWPPhysSensitivityTree`

### PHYS-LIB-1 · High — Hard-coded analysis root path

```cpp
string iDataDirectory = "/lustre/fs22/group/cta/users/maierg/analysis/AnalysisData/" + iSite;
```

Source: `Eventdisplay/src/writeCTAWPPhysSensitivityTree.cpp:248-250`.

**Impact:** the physics summary tree writer is non-portable; it fails on any system where this path does not exist.

### PHYS-LIB-2 · High — SC-MST telescope entries are never stored

The bounds check makes the write path unreachable under normal conditions:

```cpp
if( fTelescopeData.size() > 5 )
{
    fTelescopeData[4].push_back( new VTelescopeData() );
}
```

Source: `Eventdisplay/src/writeCTAWPPhysSensitivityTree.cpp:415-424`.

**Impact:** SC-MST geometry metadata is silently dropped from the summary tree.

### PHYS-LIB-3 · High — Energy bin centres read from ROOT underflow bin

```cpp
for( int i = 0; i < h->GetNbinsX(); i++ )
{
    fEnergy_logTeV[i] = h->GetBinCenter( i );    // ROOT bin 0 = underflow
}
```

Source: `Eventdisplay/src/writeCTAWPPhysSensitivityTree.cpp:515-518`.

**Impact:** the exported energy array starts with the underflow bin and drops the last real bin, shifting all exported energy coordinates by one bin.

---

## 13. C++ library: misc

### MISC-1 · High — `printRunParameter` hides file-open failures and returns success on missing metadata

- If the input file cannot be opened, the program exits with code 0 (success). Source: `Eventdisplay/src/printRunParameter.cpp:563-568`.
- `readObservingDirection()` returns `true` when neither `MC_runheader` nor `runparameterV2` is present. Source: `Eventdisplay/src/printRunParameter.cpp:162-208`.

**Impact:** shell scripts using `printRunParameter` for file naming or control flow continue with empty or misleading metadata instead of failing.

### MISC-2 · Low — `VEventLoop::printRunInfos()` reports wrong telescope types for subset arrays

```cpp
if( i < getDetectorGeometry()->getTelType().size() )
    cout << " (type " << getDetectorGeometry()->getTelType()[i] << ")";
```

Source: `Eventdisplay/src/VEventLoop.cpp:134-142`.

**Impact:** when `fTelToAnalyze` is a subset or reordered telescope list, the reported types do not match the selected IDs. Diagnostic output is misleading.

### MISC-3 · Low — `logFile` falls back to a default log on an unrecognized name

Source: `Eventdisplay/src/logFile.cpp:81-107`.

**Impact:** a misspelled log name succeeds silently, printing to the wrong log. Automated provenance extraction can record incorrect metadata.

---

## 14. FITS / DL2 / DL3 export

### FITS-1 · High — DL2 FITS converter omits reconstructed event-direction coordinates

The ROOT `DL2EventTree` contains reconstructed camera offsets (`xoff`, `yoff`) and pointing, but `generate_DL2_file.py` exports only:
`MC_AZ`, `MC_ALT`, `AZ`, `ALT`, `MC_ENERGY`, `ENERGY`, `MULTIP`, `PNT_AZ`, `PNT_ALT`, `GH_MVA`.

Source: `Converters/DL2/generate_DL2_file.py:126-137`; `Eventdisplay/src/VEffectiveAreaCalculator.cpp:408-413`, `:3592-3599`.

**Impact:** the FITS DL2 output contains no reconstructed sky position and no camera offsets from which it could be rebuilt. Downstream physics analysis cannot recover event directions from this file.

### FITS-2 · High — DL3 background FITS header declares the wrong field-of-view coordinate system

The background HDU is filled with synthetic `DETX/DETY` detector-offset coordinates, but the FITS header states:

```cpp
write_fits_keyword( "FOVALIGN", "RADEC", "FOV alignment" );
```

Source: `Converters/DL3-IRFs/src/VDL3IRFs.cpp:187-193`, `:610-699`.

**Impact:** the file claims celestial RA/Dec alignment while the data are detector-offset coordinates. Downstream IRF consumers that honour `FOVALIGN` will misinterpret background geometry.

### FITS-3 · High — DL3 PSF 3-Gauss normalization is dimensionally inconsistent

The sigma is derived in degrees but the normalization uses:

```cpp
scale.push_back( 1./ (2.*TMath::Pi()*data.back() ) );
```

with `SCALE` stored in sr⁻¹ and `SIGMA_*` in deg.

Source: `Converters/DL3-IRFs/src/VDL3IRFs.cpp:465-496`, `:533-546`.

**Impact:** the normalization misses at least one power of sigma and does not convert between degrees and steradians. The written PSF cannot be a correctly normalised 2D angular PDF.

### FITS-4 · Medium — DL2 event-type export is broken for multi-file input

The CLI accepts multiple ROOT files (`nargs=-1`) but in `event_type` mode loads exactly one sidecar file:

```python
event_types = np.loadtxt(filenames[0].replace(".eff-0.root", ".txt"), dtype=float)
data_mask[data_mask] = event_types == event_type
```

Source: `Converters/DL2/generate_DL2_file.py:33`, `:59-63`, `:111-119`.

**Impact:** the event-type mask applies only to events from the first file while `data_mask` spans all files, producing logically inconsistent output.

### FITS-5 · Medium — Legacy `VFITS` has a stack-buffer overrun for non-square sky maps

`createImageFitsFile()` allocates row pointers sized by X bins but initialises them over Y bins:

```cpp
double* array[hSkyMap->GetNbinsX()];
for( int ii = 1; ii < naxes[1]; ii++ )
    array[ii] = array[ii - 1] + naxes[0];
```

Source: `Eventdisplay/src/VFITS.cpp:883-904`.

**Impact:** when Y bins > X bins, the initialisation loop writes past the stack-allocated pointer array — a memory-safety bug.

### FITS-6 · Medium — Legacy `VFITS` table writers read the ROOT underflow bin and miss the last real bin

Multiple writers iterate `i = 0` to `< GetNbinsX()` and call ROOT bin accessors with `i` (not `i+1`):

```cpp
runValues.push_back( h->GetBinCenter( i ) );
runValues.push_back( h->GetBinContent( i ) );
```

Source: `Eventdisplay/src/VFITS.cpp:540-565`, `:618-624`.

**Impact:** FITS tables start with the underflow value and drop the highest real bin, silently shifting histogram exports.

### FITS-7 · Low — Legacy `VFITS` graph writer uses a self-acknowledged wrong Δ*x* formula

```cpp
runValues.push_back( x - x1 ); // Note: this is wrong and should be fixed (delta_x)
```

Source: `Eventdisplay/src/VFITS.cpp:667-670`.

**Impact:** the first `Delta_*` entry is wrong by construction; all later entries represent point-to-point spacing rather than a real bin width.

### FITS-8 · Medium — DL2 and DL3 converters hard-code stale production identifiers

- `generate_DL2_file.py` writes `CREATOR = "Eventdisplay prod5"` regardless of production. Source: line 175-183.
- `convertSensitivityFilesToFITS.cpp` writes `CTA (MC prod5, v0.1.1)` in all primary headers. Source: line 74-76.

**Impact:** provenance metadata in exported FITS files is incorrect for prod6 and later productions.

---

## 15. Analysis parameter files

### PAR-1 · High — prod6 full-moon calibration IPR files are missing for most zenith/site combinations

The EVNDISP worker constructs:

```sh
PEDFIL="$CTA_EVNDISP_AUX_DIR/Calibration/prod6/prod6-${obs}-full-ze${ZE}-IPR.root"
```

Source: `CTA.EVNDISP.sub_convert_and_analyse_MC_VDST_ArrayJob.sh:112-126`.

The directory contains only `prod6-north-full-ze40deg-IPR.root` and `prod6-south-full-ze40deg-IPR.root`. Files for 20°, 52°, and 60° zenith are absent for both sites. Site-generic files (`prod6-full-ze20deg-IPR.root`, etc.) exist but do not match the active naming pattern.

**Impact:** prod6 full-moon EVNDISP jobs at any zenith angle other than 40° fail when resolving the calibration file.

### PAR-2 · Medium — `EFFECTIVEAREA.runparameter` references a non-existent cut file and an obsolete path convention

```text
* CUTFILE ANASUM.GammaHadron.dat
(if no path is given, file is searched in $VERITAS_EVNDISP_ANA_DIR/ParameterFiles/)
```

Source: `Eventdisplay_AnalysisFiles_CTA/ParameterFiles/EFFECTIVEAREA.runparameter:45-46`.

The current CTA wrappers use split `.gamma.dat`/`.CRbck.dat` templates, not `ANASUM.GammaHadron.dat`, and use `$CTA_EVNDISP_ANA_DIR`, not `$VERITAS_EVNDISP_ANA_DIR`.

**Impact:** running `makeEffectiveArea` directly from this file starts from an outdated environment convention and a non-existent cut file.

### PAR-3 · Low — README under-documents active analysis-file inputs

`Eventdisplay_AnalysisFiles_CTA/README.md` does not mention calibration IPR files or `TMVA.BDTDispQualityCuts.runparameter`, both of which are required by the active pipeline.

Source: `Eventdisplay_AnalysisFiles_CTA/README.md:11-31`.

---

## 16. Log file evidence

### LOG-1 · High — DISP training submitted for telescope types absent from the array (log-confirmed)

```text
total number of telescopes: 30 (selected 0)
..nothing to do. Exiting.
```

Observed in `logFiles/BDTDisp-10408618.training.log:46-48`, `BDTDisp-201109916.training.log:46-48`, `BDTDisp-205008707.training.log:46-48`.

Successful training logs exist only for types `10608418` and `138704810` — the only types present in the prod6 North subarray, as confirmed by `logFiles/makeTable.log:60-95` and `logFiles/mscwTable.log:181-218`.

**Impact:** the pipeline submits DISP training jobs for every telescope type regardless of array composition. All jobs for absent types are wasted and generate no usable models.

### LOG-2 · Medium — Production log set spans multiple incompatible production tags

The bundled logs for the same prod6 North array use different tags:

| Stage | Tag |
|-------|-----|
| Lookup tables | `g20260214` |
| TMVA training | `g20250822` |
| Effective areas | `g20260325` |
| PHYS/IRF | `g20260129` |

Source: `logFiles/makeTable.log:16`, `logFiles/tmva.log:5`, `logFiles/effArea_angres.log:7`, `logFiles/irf.log:21`.

**Impact:** these logs do not represent a single coherent end-to-end production run. Cross-stage comparisons within this log set must account for different analysis namespaces.

---

## Summary table

| ID | Severity | Component | Short description |
|----|----------|-----------|-------------------|
| STEREO-1 | **Critical** | `VStereoReconstruction` | `removeDataSet` inverted logic; always returns false |
| ML-1 | **Critical** | TMVA training | `intersect_mc_error` Y-component zeroed by typo |
| ENV-1 | **High** | Orchestration | Hard-coded DESY AFS ROOT path |
| ENV-2 | **High** | Orchestration | Dead array configurations from cascading reassignment |
| EVN-1 | **High** | EVNDISP | `continue` outside loop terminates EVNDISP branch abnormally |
| EVN-2 | **High** | EVNDISP | Fragile simtel input discovery via unguarded glob |
| TAB-1 | **High** | Lookup tables | Stereo-angle cut differs between table filling (10°) and analysis (5°/15°) |
| TAB-2 | **High** | Lookup tables | Undefined `MEANDIST` in table job filenames |
| DISP-1 | **High** | DISP training | Argument parsing broken: wrong positional guards, unquoted tests always true, silent job loss |
| DISP-2 | **High** | DISP training | Training submitted for telescope types absent from array |
| TMVA-1 | **High** | TMVA training | `prepareTMVA` conflates direction and batch-options in positional argument `$5` |
| TMVA-2 | **High** | TMVA training | BDT energy bins 0 and 1 overlap completely; no exclusive low-energy classifier |
| TMVA-3 | **High** | TMVA training | Training aborts on first missing file with no per-bin retry |
| XGB-1 | **High** | XGB stereo | Apply wrapper caps `MINTEL` at 3; loads wrong model when training used `MINTEL > 3` |
| XGB-2 | **High** | XGB stereo | XGB classification not wired into orchestration layer |
| XGB-3 | **High** | XGB / Python | `configure_train()` treats `energy_bins_log10_tev` as dict; loader treats it as list |
| IRF-1 | **High** | PHYS dispatch | Trailing space in subarray filename argument |
| IRF-2 | **High** | IRF runparameter | Key `THETA2MINENEERGY` likely misspelled; silently ignored by parser |
| SIG-1 | **High** | `VTraceHandler` | Sliding-window raw mode returns `FADC[1]` instead of integrated charge |
| SIG-2 | **High** | `VTraceHandler` | Sliding-window loop accesses one sample past trace end |
| IRF-LIB-1 | **High** | `makeEffectiveArea` | MC header documented as optional but library hard-exits if missing |
| IRF-LIB-2 | **High** | IRF reader | True-energy axis handling disabled unconditionally |
| IRF-LIB-3 | **High** | IRF reader | Histogram name/unit inconsistency compensated by heuristic fallbacks |
| IRF-LIB-4 | **High** | IRF runparameter | Unknown runparameter keys silently ignored |
| PHYS-LIB-1 | **High** | Physics tree | Hard-coded LUSTRE path in `writeCTAWPPhysSensitivityTree` |
| PHYS-LIB-2 | **High** | Physics tree | SC-MST telescope entries never stored (unreachable branch) |
| PHYS-LIB-3 | **High** | Physics tree | Energy centres read from ROOT underflow bin |
| ML-2 | **High** | TMVA training | Zenith-bin range check off by one; out-of-bounds access |
| ML-3 | **High** | TMVA training | `get_largest_weight_disp_entries` sorts ascending, selects smallest-weight images |
| ML-4 | **High** | TMVA training | Per-telescope feature remapping overwrites source arrays in place |
| FITS-1 | **High** | DL2 converter | Reconstructed event direction absent from FITS output |
| FITS-2 | **High** | DL3 converter | Background FITS header declares wrong FOV coordinate system |
| FITS-3 | **High** | DL3 converter | PSF 3-Gauss normalization dimensionally inconsistent |
| PAR-1 | **High** | Analysis files | prod6 full-moon IPR calibration files missing for most zenith/site combinations |
| LOG-1 | **High** | Log evidence | DISP training jobs submitted for telescope types absent from selected array |
| STEREO-2 | **High** | `VOnOff` | OFF histogram mutated during on-off subtraction |
| STEREO-3 | **High** | `VArrayAnalyzer` | Method-4 pair-angle rejection tests slopes of first telescope pair, not current pair |
| MISC-1 | **High** | `printRunParameter` | File-open failure exits with code 0; missing metadata returns `true` |
| TAB-3 | **Medium** | Lookup tables | Undefined `NC` in analysis output filenames |
| TAB-4 | **Medium** | Lookup tables | DISP training/analysis cut synchronisation enforced only by comment |
| TAB-5 | **Medium** | `mscw_energy` CLI | Contradictory `-fill` and `-tmva_nimages_max` validation messages |
| DISP-3 | **Medium** | DISP training | Training failure detected only by log-text grep |
| TMVA-4 | **Medium** | PHYS wrapper | Unquoted positional guards always true in `CTA.WPPhysWriter.sub.sh` |
| TMVA-5 | **Medium** | TMVA training | Diagnostic training artifacts deleted unconditionally |
| XGB-4 | **Medium** | XGB stereo | Training wrapper bypasses external hyperparameter-config hook |
| XGB-5 | **Medium** | XGB stereo | Direction argument documented but never read |
| IRF-3 | **Medium** | IRF validation | Effective-area success inferred from file size |
| IRF-4 | **Medium** | IRF writing | Summary-tree builder overrides `PHYSDIR` immediately after setting it |
| SIG-3 | **Medium** | `VTraceHandler` | Low-gain saturation threshold computed from previous call's `tmax` |
| IRF-LIB-5 | **Medium** | IRF binning | Default energy/offset axes duplicated by contract between two classes |
| IRF-LIB-6 | **Medium** | TMVA evaluator | XML variable extraction depends on exact string layout |
| IRF-LIB-7 | **Medium** | TMVA evaluator | Incomplete TMVA weight grids pass initialisation with warnings only |
| IRF-LIB-8 | **Medium** | IRF filling | Shared `MCaz` mutated in-place inside the azimuth-bin loop |
| IRF-LIB-9 | **Medium** | `makeEffectiveArea` | IRF resolution duplication is order-dependent by vector index |
| IRF-LIB-10 | **Medium** | `makeEffectiveArea` | First ROOT input file handle never closed in `copyMCHistograms()` |
| ML-5 | **Medium** | TMVA training | Reconstruction-error feature stored under misleading branch name |
| ML-6 | **Medium** | TMVA training | Angular-method training splits both classes from signal tree (correct design, but compounds ML-1) |
| STEREO-4 | **Medium** | `VEmissionHeightCalculator` | Pair weights divide by zero for `size ≤ 1` |
| STEREO-5 | **Medium** | `VEmissionHeightCalculator` | Telescope positions accumulate without clear across reinitialisations |
| FITS-4 | **Medium** | DL2 converter | Event-type export broken for multi-file input |
| FITS-5 | **Medium** | `VFITS` | Stack-buffer overrun for non-square sky maps |
| FITS-6 | **Medium** | `VFITS` | Table writers read ROOT underflow bin and drop last real bin |
| FITS-8 | **Medium** | Converters | Hard-coded `prod5` creator string in DL2 and DL3 FITS headers |
| PAR-2 | **Medium** | Analysis files | `EFFECTIVEAREA.runparameter` references non-existent cut file and obsolete path |
| ENV-3 | **Medium** | Orchestration | Production version tags encoded as script literals |
| EVN-3 | **Medium** | EVNDISP | Extensive unquoted paths in file operations |
| EVN-4 | **Medium** | EVNDISP | Low-gain calibration multipliers absent for trace integration method 2 |
| EVN-5 | **Medium** | EVNDISP / calibration | Low-gain T0 assignment always overwrites the high-gain field |
| EVN-6 | **Medium** | EVNDISP / calibration | Low-gain T0 calculation uses high-gain charge extraction path |
| LOG-2 | **Medium** | Log evidence | Log set covers multiple incompatible production tags |
| SIG-4 | **Low** | `VImageParameterCalculation` | Muon-ring correction divides by zero at first grid point |
| XGB-6 | **Low** | Analysis files | `DL2.production.runparameter` is legacy TMVA example |
| FITS-7 | **Low** | `VFITS` | Graph writer Δ*x* formula acknowledged wrong in source comment |
| PAR-3 | **Low** | Analysis files | README omits required calibration and quality-cut files |
| MISC-2 | **Low** | `VEventLoop` | Wrong telescope type reported for subset arrays in diagnostic output |
| MISC-3 | **Low** | `logFile` | Misspelled log name silently returns a different log object |

---

## Open items requiring confirmation

- **Signal-efficiency threshold per multiplicity:** `CUTS` applies 80% max signal efficiency for `MST >= 4` and 95% otherwise (`CTA.runAnalysis.sh:722-747`). The physics motivation for this split should be verified.
- **Exposure-dependent re-running:** `CUTS` and `PHYS` are the only stages executed for non-`50h` observation times (`CTA.runAnalysis.sh:548-553`). If upstream lookup tables or BDT models should also vary with exposure, this is a scope inconsistency.
- **`DL2FILLING` mode compatibility:** `FULLTREES` is set for `DL2plus` or `prod6`, `DL2` otherwise (`CTA.runAnalysis.sh:600-608`). Compatibility with downstream DL2 consumers for both modes should be confirmed.
- **Testing coverage gap:** the scripts repository performs presence/count checks on DISP, TMVA, and effective-area outputs (`testProduction/README.md:1-40`), but no end-to-end parameter-consistency validation exists across the `MAKETABLES` → `ANATABLES` → `TRAIN` → `CUTS` chain.

---

## Cross-check against VERITAS pipeline review

The VERITAS (VTS) pipeline uses the same `Eventdisplay` C++ codebase. Issues flagged in `vts_eventdisplay_pipeline_review.md` were cross-checked against the CTA codebase. The following VTS issues were verified as also present in CTA and have been added to this report:

| VTS ID | CTA ID | Notes |
|--------|--------|-------|
| A7 | STEREO-3 | Confirmed at `VArrayAnalyzer.cpp:1350` |
| A10 | STEREO-4 | Confirmed at `VEmissionHeightCalculator.cpp:61` |
| A11 | STEREO-5 | Confirmed in both `setTelescopePositions()` overloads |
| A4 | EVN-5 | Confirmed at `VCalibrationData.cpp:380-388` |
| A5 | EVN-6 | Confirmed at `VCalibrator.cpp:868` |
| A9 | SIG-4 | Confirmed at `VImageParameterCalculation.cpp:744-745` |
| B4 | IRF-LIB-7 | Confirmed in `VTMVAEvaluator::initializeWeightFiles()` |
| B6 | IRF-LIB-8 | Confirmed at `VInstrumentResponseFunction.cpp:258-260` |
| B7 | IRF-LIB-9 | Confirmed at `makeEffectiveArea.cpp:297-300` |
| B8 | IRF-LIB-10 | Confirmed at `makeEffectiveArea.cpp:567-579` |

The following VTS issues were verified as **not applicable** to CTA:

| VTS ID | Reason |
|--------|--------|
| A1 | CTA `VTraceHandler.cpp` has an explicit `maxpos--` guard; access is bounded |
| A2 | Already documented as CTA SIG-1 |
| A3 | `VFitTraceHandler.cpp` is not present in the CTA codebase |
| A6 | CTA `VArrayAnalyzer.cpp:1056` correctly tests both `ximp` and `yimp` |
| A12 | CTA `VInstrumentResponseFunctionData.cpp:485-486` was already fixed (comment: "was xcore both times") |
| A13 | Already documented as CTA ML-2 |
| A14 | CTA `trainTMVAforGammaHadronSeparation.cpp:510` correctly uses `AnalysisType=Regression` |
| B1/B2 | CTA integrates XGB via `E_ReconstructionType::XGBSTEREO` enum, not the VTS `getXGBTree()` pattern |
| B3 | Already documented as CTA IRF-LIB-6 |
| B9 | The `analysisType == 5` dead branch is not present in the CTA `anasum.cpp` |
| B10 | VTS-specific (real-data epoch reevaluation); CTA pipeline is MC-only |
| C1–C22 | VTS-specific shell orchestration (VERITAS DB, laser calibration, epoch handling) |
| D1–D11 | VTS-specific IRF production scripts |
| E6 | Already documented as CTA XGB-3 |
| E13 | CTA TMVA parameter files use `AdaBoost`/`Bagging`; the `NegWeightTreatment+Grad` mismatch does not apply |
| F1–F12 | VTS-specific auxiliary files and epoch metadata |
| G1–G7 | VTS-specific release-test infrastructure |
| H1–H7 | VTS-specific data-analysis performance (real-data `evndisp`/`anasum` throughput) |


Step	Mode key	Entry script	Executable
1	`EVNDISP`	`CTA.EVNDISP.sub_convert_and_analyse_MC_VDST_ArrayJob.sh`	`evndisp`
2	`DISPBDT`	`CTA.DISPTRAINING.sub_analyse.sh`	`trainTMVAforGammaHadronSeparation`
3	`MAKETABLES`	`CTA.MSCW_ENERGY.sub_make_tables.sh`	`mscw_energy -filltables=1`
4	`ANATABLES`	`CTA.MSCW_ENERGY.sub_analyse_MC.sh`	`mscw_energy` + `smoothLookupTables`
5	`XGBSTEREOTRAIN` / `XGBSTEREOANA`	`CTA.XGBSTEREO.sub_train/analyse.sh`	`eventdisplay-ml-train/apply-xgb-stereo`
6	`PREPARETMVA`	`CTA.prepareTMVA.sub_train.sh`	`WriteTrainingEvents`
7	`TRAIN` / `TRAIN_RECO_*`	`CTA.TMVA.sub_train.sh`	`trainTMVAforGammaHadronSeparation`
8	`ANGRES` / `QC` / `CUTS`	`CTA.EFFAREA.sub_analyse_list.sh`	`makeEffectiveArea`
9	`PHYS`	`CTA.WPPhysWriter.sub.sh`	`writeCTAWPPhysSensitivityFiles`

Label	Meaning
Critical	Logic inversion or memory error that corrupts results unconditionally
High	Bug that silently produces wrong physics output, crashes, or causes silent job loss under realistic inputs
Medium	Fragile code or parameter mismatch that can produce wrong results in specific but plausible conditions
Low	Metadata / diagnostic error with no direct physics impact

Stage	Tag
Lookup tables	`g20260214`
TMVA training	`g20250822`
Effective areas	`g20260325`
PHYS/IRF	`g20260129`

ID	Severity	Component	Short description
STEREO-1	Critical	`VStereoReconstruction`	`removeDataSet` inverted logic; always returns false
ML-1	Critical	TMVA training	`intersect_mc_error` Y-component zeroed by typo
ENV-1	High	Orchestration	Hard-coded DESY AFS ROOT path
ENV-2	High	Orchestration	Dead array configurations from cascading reassignment
EVN-1	High	EVNDISP	`continue` outside loop terminates EVNDISP branch abnormally
EVN-2	High	EVNDISP	Fragile simtel input discovery via unguarded glob
TAB-1	High	Lookup tables	Stereo-angle cut differs between table filling (10°) and analysis (5°/15°)
TAB-2	High	Lookup tables	Undefined `MEANDIST` in table job filenames
DISP-1	High	DISP training	Argument parsing broken: wrong positional guards, unquoted tests always true, silent job loss
DISP-2	High	DISP training	Training submitted for telescope types absent from array
TMVA-1	High	TMVA training	`prepareTMVA` conflates direction and batch-options in positional argument `$5`
TMVA-2	High	TMVA training	BDT energy bins 0 and 1 overlap completely; no exclusive low-energy classifier
TMVA-3	High	TMVA training	Training aborts on first missing file with no per-bin retry
XGB-1	High	XGB stereo	Apply wrapper caps `MINTEL` at 3; loads wrong model when training used `MINTEL > 3`
XGB-2	High	XGB stereo	XGB classification not wired into orchestration layer
XGB-3	High	XGB / Python	`configure_train()` treats `energy_bins_log10_tev` as dict; loader treats it as list
IRF-1	High	PHYS dispatch	Trailing space in subarray filename argument
IRF-2	High	IRF runparameter	Key `THETA2MINENEERGY` likely misspelled; silently ignored by parser
SIG-1	High	`VTraceHandler`	Sliding-window raw mode returns `FADC[1]` instead of integrated charge
SIG-2	High	`VTraceHandler`	Sliding-window loop accesses one sample past trace end
IRF-LIB-1	High	`makeEffectiveArea`	MC header documented as optional but library hard-exits if missing
IRF-LIB-2	High	IRF reader	True-energy axis handling disabled unconditionally
IRF-LIB-3	High	IRF reader	Histogram name/unit inconsistency compensated by heuristic fallbacks
IRF-LIB-4	High	IRF runparameter	Unknown runparameter keys silently ignored
PHYS-LIB-1	High	Physics tree	Hard-coded LUSTRE path in `writeCTAWPPhysSensitivityTree`
PHYS-LIB-2	High	Physics tree	SC-MST telescope entries never stored (unreachable branch)
PHYS-LIB-3	High	Physics tree	Energy centres read from ROOT underflow bin
ML-2	High	TMVA training	Zenith-bin range check off by one; out-of-bounds access
ML-3	High	TMVA training	`get_largest_weight_disp_entries` sorts ascending, selects smallest-weight images
ML-4	High	TMVA training	Per-telescope feature remapping overwrites source arrays in place
FITS-1	High	DL2 converter	Reconstructed event direction absent from FITS output
FITS-2	High	DL3 converter	Background FITS header declares wrong FOV coordinate system
FITS-3	High	DL3 converter	PSF 3-Gauss normalization dimensionally inconsistent
PAR-1	High	Analysis files	prod6 full-moon IPR calibration files missing for most zenith/site combinations
LOG-1	High	Log evidence	DISP training jobs submitted for telescope types absent from selected array
STEREO-2	High	`VOnOff`	OFF histogram mutated during on-off subtraction
STEREO-3	High	`VArrayAnalyzer`	Method-4 pair-angle rejection tests slopes of first telescope pair, not current pair
MISC-1	High	`printRunParameter`	File-open failure exits with code 0; missing metadata returns `true`
TAB-3	Medium	Lookup tables	Undefined `NC` in analysis output filenames
TAB-4	Medium	Lookup tables	DISP training/analysis cut synchronisation enforced only by comment
TAB-5	Medium	`mscw_energy` CLI	Contradictory `-fill` and `-tmva_nimages_max` validation messages
DISP-3	Medium	DISP training	Training failure detected only by log-text grep
TMVA-4	Medium	PHYS wrapper	Unquoted positional guards always true in `CTA.WPPhysWriter.sub.sh`
TMVA-5	Medium	TMVA training	Diagnostic training artifacts deleted unconditionally
XGB-4	Medium	XGB stereo	Training wrapper bypasses external hyperparameter-config hook
XGB-5	Medium	XGB stereo	Direction argument documented but never read
IRF-3	Medium	IRF validation	Effective-area success inferred from file size
IRF-4	Medium	IRF writing	Summary-tree builder overrides `PHYSDIR` immediately after setting it
SIG-3	Medium	`VTraceHandler`	Low-gain saturation threshold computed from previous call's `tmax`
IRF-LIB-5	Medium	IRF binning	Default energy/offset axes duplicated by contract between two classes
IRF-LIB-6	Medium	TMVA evaluator	XML variable extraction depends on exact string layout
IRF-LIB-7	Medium	TMVA evaluator	Incomplete TMVA weight grids pass initialisation with warnings only
IRF-LIB-8	Medium	IRF filling	Shared `MCaz` mutated in-place inside the azimuth-bin loop
IRF-LIB-9	Medium	`makeEffectiveArea`	IRF resolution duplication is order-dependent by vector index
IRF-LIB-10	Medium	`makeEffectiveArea`	First ROOT input file handle never closed in `copyMCHistograms()`
ML-5	Medium	TMVA training	Reconstruction-error feature stored under misleading branch name
ML-6	Medium	TMVA training	Angular-method training splits both classes from signal tree (correct design, but compounds ML-1)
STEREO-4	Medium	`VEmissionHeightCalculator`	Pair weights divide by zero for `size ≤ 1`
STEREO-5	Medium	`VEmissionHeightCalculator`	Telescope positions accumulate without clear across reinitialisations
FITS-4	Medium	DL2 converter	Event-type export broken for multi-file input
FITS-5	Medium	`VFITS`	Stack-buffer overrun for non-square sky maps
FITS-6	Medium	`VFITS`	Table writers read ROOT underflow bin and drop last real bin
FITS-8	Medium	Converters	Hard-coded `prod5` creator string in DL2 and DL3 FITS headers
PAR-2	Medium	Analysis files	`EFFECTIVEAREA.runparameter` references non-existent cut file and obsolete path
ENV-3	Medium	Orchestration	Production version tags encoded as script literals
EVN-3	Medium	EVNDISP	Extensive unquoted paths in file operations
EVN-4	Medium	EVNDISP	Low-gain calibration multipliers absent for trace integration method 2
EVN-5	Medium	EVNDISP / calibration	Low-gain T0 assignment always overwrites the high-gain field
EVN-6	Medium	EVNDISP / calibration	Low-gain T0 calculation uses high-gain charge extraction path
LOG-2	Medium	Log evidence	Log set covers multiple incompatible production tags
SIG-4	Low	`VImageParameterCalculation`	Muon-ring correction divides by zero at first grid point
XGB-6	Low	Analysis files	`DL2.production.runparameter` is legacy TMVA example
FITS-7	Low	`VFITS`	Graph writer Δx formula acknowledged wrong in source comment
PAR-3	Low	Analysis files	README omits required calibration and quality-cut files
MISC-2	Low	`VEventLoop`	Wrong telescope type reported for subset arrays in diagnostic output
MISC-3	Low	`logFile`	Misspelled log name silently returns a different log object

VTS ID	CTA ID	Notes
A7	STEREO-3	Confirmed at `VArrayAnalyzer.cpp:1350`
A10	STEREO-4	Confirmed at `VEmissionHeightCalculator.cpp:61`
A11	STEREO-5	Confirmed in both `setTelescopePositions()` overloads
A4	EVN-5	Confirmed at `VCalibrationData.cpp:380-388`
A5	EVN-6	Confirmed at `VCalibrator.cpp:868`
A9	SIG-4	Confirmed at `VImageParameterCalculation.cpp:744-745`
B4	IRF-LIB-7	Confirmed in `VTMVAEvaluator::initializeWeightFiles()`
B6	IRF-LIB-8	Confirmed at `VInstrumentResponseFunction.cpp:258-260`
B7	IRF-LIB-9	Confirmed at `makeEffectiveArea.cpp:297-300`
B8	IRF-LIB-10	Confirmed at `makeEffectiveArea.cpp:567-579`

VTS ID	Reason
A1	CTA `VTraceHandler.cpp` has an explicit `maxpos--` guard; access is bounded
A2	Already documented as CTA SIG-1
A3	`VFitTraceHandler.cpp` is not present in the CTA codebase
A6	CTA `VArrayAnalyzer.cpp:1056` correctly tests both `ximp` and `yimp`
A12	CTA `VInstrumentResponseFunctionData.cpp:485-486` was already fixed (comment: "was xcore both times")
A13	Already documented as CTA ML-2
A14	CTA `trainTMVAforGammaHadronSeparation.cpp:510` correctly uses `AnalysisType=Regression`
B1/B2	CTA integrates XGB via `E_ReconstructionType::XGBSTEREO` enum, not the VTS `getXGBTree()` pattern
B3	Already documented as CTA IRF-LIB-6
B9	The `analysisType == 5` dead branch is not present in the CTA `anasum.cpp`
B10	VTS-specific (real-data epoch reevaluation); CTA pipeline is MC-only
C1–C22	VTS-specific shell orchestration (VERITAS DB, laser calibration, epoch handling)
D1–D11	VTS-specific IRF production scripts
E6	Already documented as CTA XGB-3
E13	CTA TMVA parameter files use `AdaBoost`/`Bagging`; the `NegWeightTreatment+Grad` mismatch does not apply
F1–F12	VTS-specific auxiliary files and epoch metadata
G1–G7	VTS-specific release-test infrastructure
H1–H7	VTS-specific data-analysis performance (real-data `evndisp`/`anasum` throughput)

Uh oh!

Eventdisplay CTAO Pipeline — Issue Report #198

Description

Eventdisplay CTAO Pipeline — Issue Report

Pipeline overview

Severity definitions

1. Environment and orchestration

ENV-1 · High — Hard-coded DESY AFS path makes the pipeline non-portable

ENV-2 · High — Dataset-to-array mapping uses cascading ARRAY=() reassignments, leaving dead configurations

ENV-3 · Medium — Versioning encoded in script literals

2. Stage: EVNDISP

EVN-1 · High — continue outside a loop terminates the EVNDISP branch abnormally

EVN-2 · High — Input discovery is fragile when TMPDIR contains zero or multiple simtel files

EVN-3 · Medium — Extensive unquoted path handling makes file operations shell-fragile

EVN-4 · Medium — Missing low-gain calibration multipliers observed in production logs

EVN-5 · Medium — Low-gain average-T0 assignment always overwrites the high-gain field

EVN-6 · Medium — Low-gain average-T0 calculation uses the high-gain charge extraction path

3. Stage: Lookup tables and stereo reconstruction

TAB-1 · High — Table-filling and analysis stages use different stereo-angle cuts

TAB-2 · High — CTA.MSCW_ENERGY.sub_make_tables.sh uses an undefined variable MEANDIST in generated filenames

TAB-3 · Medium — CTA.MSCW_ENERGY.sub_analyse_MC.sh uses an undefined variable NC in output names

TAB-4 · Medium — DISP analysis and training cuts are synchronized only by a comment

TAB-5 · Medium — VTableLookupRunParameter contains contradictory CLI validation

4. Stage: DISP training

DISP-1 · High — CTA.DISPTRAINING.sub_analyse.sh argument parsing is broken in two ways

DISP-2 · High — DISP training is submitted for telescope types absent from the selected array (confirmed by logs)

DISP-3 · Medium — Failed telescope-type training is detected only by log-text grep

5. Stage: TMVA preparation and gamma/hadron training

TMVA-1 · High — CTA.prepareTMVA.sub_train.sh mis-parses optional arguments

TMVA-2 · High — Gamma/hadron BDT energy bins 0 and 1 overlap completely (bin 0 ⊂ bin 1)

TMVA-3 · High — TMVA training aborts entirely if the first expected training file is missing

TMVA-4 · Medium — Unquoted positional-parameter checks are always true in CTA.WPPhysWriter.sub.sh

TMVA-5 · Medium — Diagnostic artifacts deleted unconditionally after training

6. Stage: XGB stereo reconstruction

XGB-1 · High — XGB apply wrapper can load a different model than training produced

XGB-2 · High — XGB classification is not wired into the orchestration layer

XGB-3 · High — configure_train() treats energy_bins_log10_tev as a dict, but loaders treat it as a list of per-bin dicts

XGB-4 · Medium — Training wrapper bypasses the external hyperparameter-config hook

XGB-5 · Medium — CTA.XGBSTEREO.sub_train.sh advertises a direction argument it never reads

XGB-6 · Low — DL2.production.runparameter is a legacy TMVA example

7. Stage: IRF calculation and PHYS writing

IRF-1 · High — PHYS passes a filename with a trailing space

IRF-2 · High — Run-parameter key THETA2MINENEERGY appears misspelled

IRF-3 · Medium — Effective-area success is inferred from file size rather than content

IRF-4 · Medium — Summary-tree builder is hard-wired to one archived directory convention

8. C++ library: signal extraction (VTraceHandler)

SIG-1 · High — Sliding-window raw mode returns the wrong quantity

SIG-2 · High — Sliding-window loop can access one sample past the trace end

SIG-3 · Medium — Low-gain saturation threshold computed from uninitialized tmax

SIG-4 · Low — Muon-ring correction divides by zero at the first grid point

9. C++ library: stereo reconstruction

STEREO-1 · Critical — VStereoReconstruction::removeDataSet has inverted logic and always reports failure

STEREO-2 · High — VOnOff mutates the OFF histogram during on-off subtraction

STEREO-3 · High — Method-4 stereo pair-angle rejection evaluates the wrong telescope slopes

STEREO-4 · Medium — Emission-height pair weights undefined for small image sizes

STEREO-5 · Medium — VEmissionHeightCalculator accumulates telescope positions across reinitialisations

10. C++ library: TMVA training

ML-1 · Critical — intersect_mc_error ignores the Y-component due to a typo

ML-2 · High — Zenith-bin range check is off by one

ML-3 · High — get_largest_weight_disp_entries() sorts ascending, selecting the smallest-weight entries

ML-4 · High — Per-telescope feature remapping writes into its own source arrays

ML-5 · Medium — Reduced training tree stores a reconstruction-error feature under a misleading branch name

ML-6 · Medium — TrainAngularReconstructionMethod splits both classes from the same signal tree

11. C++ library: IRF and effective area

IRF-LIB-1 · High — makeEffectiveArea documents the MC header as optional, but the library hard-exits if it is missing

IRF-LIB-2 · High — VInstrumentResponseFunctionReader disables true-energy-axis handling unconditionally

IRF-LIB-3 · High — CTA IRF reader compensates for inconsistent histogram naming and units in the files it reads

IRF-LIB-4 · High — VInstrumentResponseFunctionRunParameter silently ignores unknown runparameter keys

IRF-LIB-5 · Medium — IRF binning defaults are duplicated by contract between two classes

IRF-LIB-6 · Medium — TMVA XML parsing depends on exact string layout

IRF-LIB-7 · Medium — TMVA weight grid silently passes initialisation with missing interior bins

IRF-LIB-8 · Medium — VInstrumentResponseFunction mutates the shared MC azimuth in place

IRF-LIB-9 · Medium — makeEffectiveArea IRF resolution duplication is order-dependent

IRF-LIB-10 · Medium — copyMCHistograms() leaves the first ROOT input file open

12. C++ library: writeCTAWPPhysSensitivityTree

PHYS-LIB-1 · High — Hard-coded analysis root path

PHYS-LIB-2 · High — SC-MST telescope entries are never stored

PHYS-LIB-3 · High — Energy bin centres read from ROOT underflow bin

13. C++ library: misc

MISC-1 · High — printRunParameter hides file-open failures and returns success on missing metadata

ENV-2 · High — Dataset-to-array mapping uses cascading `ARRAY=()` reassignments, leaving dead configurations

EVN-1 · High — `continue` outside a loop terminates the EVNDISP branch abnormally

TAB-2 · High — `CTA.MSCW_ENERGY.sub_make_tables.sh` uses an undefined variable `MEANDIST` in generated filenames

TAB-3 · Medium — `CTA.MSCW_ENERGY.sub_analyse_MC.sh` uses an undefined variable `NC` in output names

TAB-5 · Medium — `VTableLookupRunParameter` contains contradictory CLI validation

DISP-1 · High — `CTA.DISPTRAINING.sub_analyse.sh` argument parsing is broken in two ways

TMVA-1 · High — `CTA.prepareTMVA.sub_train.sh` mis-parses optional arguments

TMVA-4 · Medium — Unquoted positional-parameter checks are always true in `CTA.WPPhysWriter.sub.sh`

XGB-3 · High — `configure_train()` treats `energy_bins_log10_tev` as a dict, but loaders treat it as a list of per-bin dicts

XGB-5 · Medium — `CTA.XGBSTEREO.sub_train.sh` advertises a direction argument it never reads

XGB-6 · Low — `DL2.production.runparameter` is a legacy TMVA example

IRF-1 · High — `PHYS` passes a filename with a trailing space

IRF-2 · High — Run-parameter key `THETA2MINENEERGY` appears misspelled

8. C++ library: signal extraction (`VTraceHandler`)

SIG-3 · Medium — Low-gain saturation threshold computed from uninitialized `tmax`

STEREO-1 · Critical — `VStereoReconstruction::removeDataSet` has inverted logic and always reports failure

STEREO-2 · High — `VOnOff` mutates the OFF histogram during on-off subtraction

STEREO-5 · Medium — `VEmissionHeightCalculator` accumulates telescope positions across reinitialisations

ML-1 · Critical — `intersect_mc_error` ignores the Y-component due to a typo

ML-3 · High — `get_largest_weight_disp_entries()` sorts ascending, selecting the smallest-weight entries

ML-6 · Medium — `TrainAngularReconstructionMethod` splits both classes from the same signal tree

IRF-LIB-1 · High — `makeEffectiveArea` documents the MC header as optional, but the library hard-exits if it is missing

IRF-LIB-2 · High — `VInstrumentResponseFunctionReader` disables true-energy-axis handling unconditionally

IRF-LIB-4 · High — `VInstrumentResponseFunctionRunParameter` silently ignores unknown runparameter keys

IRF-LIB-8 · Medium — `VInstrumentResponseFunction` mutates the shared MC azimuth in place

IRF-LIB-9 · Medium — `makeEffectiveArea` IRF resolution duplication is order-dependent

IRF-LIB-10 · Medium — `copyMCHistograms()` leaves the first ROOT input file open

12. C++ library: `writeCTAWPPhysSensitivityTree`

MISC-1 · High — `printRunParameter` hides file-open failures and returns success on missing metadata

MISC-2 · Low — `VEventLoop::printRunInfos()` reports wrong telescope types for subset arrays

MISC-3 · Low — `logFile` falls back to a default log on an unrecognized name

FITS-5 · Medium — Legacy `VFITS` has a stack-buffer overrun for non-square sky maps

FITS-6 · Medium — Legacy `VFITS` table writers read the ROOT underflow bin and miss the last real bin

FITS-7 · Low — Legacy `VFITS` graph writer uses a self-acknowledged wrong Δx formula

PAR-2 · Medium — `EFFECTIVEAREA.runparameter` references a non-existent cut file and an obsolete path convention