alberta: Delight/DG stack is development-only; KondoGate not integrated, no run artifacts

[lalalune]

The Delight / Delightful Policy Gradient (DG) stack in alberta/ implements the bounded discrete-action experiment described in WP5 (specifically §5.4) of alberta/CONTINUAL_AGENT_IMPLEMENTATION_PLAN.md: a detached delight-gated policy-gradient loss (alberta_framework/core/delight.py), a standalone Kondo forward-screening gate (alberta_framework/core/kondo_gate.py), a typed learning-value router (alberta_framework/core/learning_value_router.py), a fail-closed stateful actor-critic boundary (alberta_framework/core/delightful_actor_critic.py), and a matched development benchmark (alberta_framework/benchmarks/delightful_policy_gradient_development.py). The mechanism layer is built and unit-tested, but the stack is development-only end to end: nothing in the repo integrates the Kondo gate into a training path, nothing runs the benchmark outside pytest, and no run artifacts are checked in.

What is missing / unfinished

• alberta_framework/core/delight.py — DelightfulPolicyGradientConfig.kondo_enabled is a reserved fail-closed flag: it defaults to False and __post_init__ raises ValueError on True ("Kondo compute gating is unavailable in this full-batch helper"), because the full-batch loss helper cannot skip compiled backward work. The docstring directs callers to apply KondoGate before invoking autodiff instead, but no caller does.
• alberta_framework/core/kondo_gate.py — KondoGate (screen / gather_sparse, top_k_rate and bernoulli_price modes) is implemented and unit-tested (tests/test_kondo_gate.py) but has no runtime consumer: neither delightful_actor_critic.py nor the development benchmark imports it, and the module describes itself as "L0 mechanism code, not evidence of wall-clock savings".
• alberta_framework/core/learning_value_router.py — hard-codes MECHANISM_STATUS = "development_mechanism_only" and SCIENTIFIC_PROMOTION_ALLOWED = False, embeds both into its config schema and checkpoint schema, and its strict from_config loaders reject any payload claiming otherwise ("learning-value router must remain mechanism-only" / "cannot claim promotion").
• alberta_framework/core/delightful_actor_critic.py — the stateful boundary for the DG experiment (§5.4 of CONTINUAL_AGENT_IMPLEMENTATION_PLAN.md); its docstring states that passing tests establish mechanism contracts only, not policy quality or evidence for a research claim.
• alberta_framework/benchmarks/delightful_policy_gradient_development.py — hard-codes DEVELOPMENT_ONLY = True, SCIENTIFIC_PROMOTION_ALLOWED = False, and KONDO_IMPLEMENTED = False, and states it "implements neither Kondo selection nor a promotion path". Its runner run_delightful_policy_gradient_development is referenced only from tests/test_delightful_policy_gradient_development.py; there is no console script for it in alberta/pyproject.toml and no wiring in alberta_framework/cli.py.
• alberta/outputs/ — contains no delight/DG directory or report; the benchmark's versioned development records exist as a schema and validator but no artifacts are checked in.

Suggested direction

• Integrate KondoGate.screen / gather_sparse ahead of autodiff in the actor-critic or benchmark path, as the kondo_enabled docstring in delight.py already directs, updating KONDO_IMPLEMENTED only once that is true.
• Add a script/CLI entry point for run_delightful_policy_gradient_development so development reports can be produced outside pytest, matching the track convention that benchmark executions run through scripts/CLIs.
• Any scientific claim would additionally require a separate frozen preregistered protocol under the track's evidence-promotion rules; the current modules deliberately cannot promote and that fail-closed posture should be preserved.

Verification notes

Verified against the working tree at issue-filing time: existence and docstrings of all five modules; the fail-closed kondo_enabled validation in DelightfulPolicyGradientConfig.__post_init__; the MECHANISM_STATUS / SCIENTIFIC_PROMOTION_ALLOWED constants and their embedding plus strict from_config rejection in the router's config and checkpoint schemas; the DEVELOPMENT_ONLY / SCIENTIFIC_PROMOTION_ALLOWED / KONDO_IMPLEMENTED constants in the benchmark module; a repo-wide reference search showing the benchmark module is referenced only by its own test file, that KondoGate has no runtime importer, and that the [project.scripts] table has no delight entry; and the absence of any delight/DG artifact under alberta/outputs/ (the only delight-named files there are frozen source snapshots, not run reports).

---

Filed from a full-track audit of alberta/; each claim re-verified against the working tree immediately before filing.

[lalalune]

Kondo selection accounting landed (commit 48e9c57):

Kondo selection accounting is now implemented in the development benchmark; actual Kondo compute gating deliberately remains unimplemented and KONDO_IMPLEMENTED stays False.

Determination: honest implementation in the benchmark alone was possible for SELECTION + ACCOUNTING only. Per core/kondo_gate.py's own bar, "implemented Kondo" requires demonstrably skipped compiled backward work (screen -> gather a smaller PyTree -> autodiff on the gathered tree); masking or post-hoc bookkeeping does not qualify. The benchmark's jax.lax.scan lives execute every actor/critic/average-reward update in full, so the addition is named "Kondo selection accounting" and is explicitly a measured post-hoc counterfactual.

What was added (schema bumped v1 -> v2 since the committed v1 field set changed; validator remains fail-closed on exact field sets):

• Selection policy: the deterministic maximum-probability decision of the paper's Bernoulli(sigmoid((delight - price)/temperature)) gate with frozen price=0.0 — a step is selected iff delight > price. Delight (advantage x action_surprisal) is a forward-pass quantity already bound into the trace, so each decision is causally valid at the point a real gate would skip the backward pass.
• Per-run report section "kondo_selection_accounting": per-channel compute accounting (actor gateable; critic and average_reward declared non-gateable with rationale, per delight.py's boundary that non-policy-gradient losses stay outside the gate), executed vs counterfactual-selected vs counterfactual-skipped scalar updates, selection rate (overall, per A/B/A phase, and expected Bernoulli rate), an actual_compute_gating=False flag, and frozen accounting-semantics text stating no backward work was skipped and no saving is claimed.
• Fail-closed validation: the sect

Stays open for: actual compute gating (skipping compiled backward work), which requires changes in core/delight.py under active development.