Every baseline model has been unrunnable since 2026-08-02 — catalog.py reads the retired config['targets'] key (27 models) #84

Summary

views_baseline/model/catalog.py reads self.config["targets"] at 7 call sites. That key was
deliberately retired by views-pipeline-core on 2026-08-02 (507ae11, "fix!: retire the
synthesized targets config key", refs #380 / C-269). pipeline-core migrated its own 13 read sites
to a new helper; views-baseline was never swept.

Every baseline model has therefore been unrunnable since 2026-08-02 — that's 27 models in
views-models whose algorithm this catalog serves.

Found from views-hydranet, where it blocks a pre-registered acceptance criterion (details at the
bottom). Filing here rather than fixing it, since this is your repo.

The failure

File "views_baseline/model/catalog.py", line 75, in _get_conflictology_model
    targets=self.config["targets"],
            ~~~~~~~~~~~^^^^^^^^^^^
KeyError: 'targets'

Reproduced 2026-09-07 running light_strider on the validation partition.

Why the key is gone

pipeline-core's ConfigurationManager.get_combined_config used to synthesise
config["targets"] from regression_targets + classification_targets for every model. It was a
backward-compatibility shim. views-evaluation retired targets in its 0.4.0 API and now raises
on a config carrying it — and pipeline-core hands its whole merged config to NativeEvaluator — so
the shim broke every evaluation run.

507ae11 removed the synthesis rather than hiding it, and its message is explicit that this was
the deliberate choice:

An earlier idea was to strip it at the evaluator boundary, which would have kept a dead key alive
everywhere else and concealed it from the one component honest enough to complain.

The 7 sites

views_baseline/model/catalog.py — lines 57, 62, 67, 75, 85, 98, 112, one per builder:

The replacement pipeline-core provides

views_pipeline_core.managers.configuration.configuration.combined_targets(config) — verified
importable from the currently pinned pipeline-core. It returns regression targets then
classification targets, and raises if a retired key reappears (the old shim let a stale
targets take precedence, which is how it could quietly outrank the split keys).

So each site is plausibly:

targets=combined_targets(self.config),

We have not tested this — it is what the migration did in pipeline-core, offered as a starting
point, not a verified patch. Two things worth checking that we could not check from outside:

1. Ordering. combined_targets returns regression-then-classification. If any baseline model
   depends on a different order, concatenation order matters and should be pinned by a test.
2. Models with no classification targets. light_strider declares only regression_targets;
   the helper returns [] for the missing key where the old shim reportedly crashed on None. Fine
   for us, but worth confirming against the other 26.

Blast radius

27 models in views-models carry an algorithm served by this catalog:

MixtureBaseline 8 · ParametricHurdleConflictology 6 · LocfModel 5
ConflictologyModel 4 · ZeroModel 2 · AverageModel 2

Why we care downstream

light_strider is the datafactory climatology comparator for a pre-registered kill criterion
(K3: "worse than climatology on crps_events/mcr_events for more than half the months") in
views-hydranet's ensemble-roster dossier. Because it cannot run, K3 is recorded VOID, not PASS —
the 8-member roster has never been compared to a baseline. Tracked at views-hydranet#249.

No urgency implied for your planning; we have recorded it as untested rather than assuming a pass.

Suggested check once fixed

A test that constructs each of the 7 builders from a realistic merged config would have caught this
and would catch the next retired key. Right now the first signal is a KeyError in a training run.

Cross-refs: pipeline-core 507ae11 (#380, C-269) · views-models#445 (filed 2026-09-07, same root
cause, no comments yet) · views-hydranet#249.