Pre-registration: self-subgoal conditioning probe (#6 rung (a))

2026-08-07 ~03:5xZ. Immutable once posted. Ideas #6 / #11, from the π0.5 deep read (explicit-HL is their untested-by-us runtime increment) and the Hi-VLA study (2606.10267: explicit language subgoals gain most on long horizon; SELF-generated subgoals untested there — ours is an increment, not a replication). Zero training. The instrument does NOT exist yet: it lands oracle-gated before launch, and if implementation forces any semantic deviation from this post, an amendment posts before launch (the #19 amendment precedent).

Question

Does explicit runtime hierarchy — the AR model decoding its own subgoal for a frame, then conditioning its action decode on that text through the trained [subgoal|…] prompt slot — beat the planner-less deployment baseline (AR-100k greedy panel 5.8026)? We already know semantic co-training shapes the action representation (aux-off costs +0.462, CI [0.387, 0.537] — the Implicit-HL replication). What has never run is the runtime loop: generate the plan, feed it back as an input. AR-100k is the right probe body: it trained the subgoal condition slot (condition_fields = [subgoal, outcome, smoothness]) at subgoal_dropout = 0.5, so the planner-less context is well-trained AND the hint slot saw real text; its aux head generates subgoals (owner steer 08-05 21:43Z: they generalize strikingly OOD).

Instrument (to land, oracle-gated, before launch)

bijou.eval gains a subgoal-conditioning mode for ar_backbone checkpoints (semantics frozen here; flag spelling is implementation’s):

• oracle arm — include SUBGOAL in the collator’s condition fields with no override: each frame renders its TRUE segment label via the existing subgoal_text path ([subgoal|…] trailing bracket, dropout 0); frames without a judge label render nothing and decode identically to baseline. Today this path is reachable only through a single global --condition-override subgoal=X; the instrument makes per-frame truth conditioning a first-class mode.
• self arm — a two-pass policy sharing one model load (the NarratedBijouPolicy pattern): pass 1 greedy-decodes the subgoal value line under a [generate|subgoal actions] request (planner-less prompt; its actions are retained — the narrated-subgoal-only arm comes free); pass 2 re-encodes with item["condition_subgoal"] = <pass-1 text> rendered in the prompt slot and decodes actions on the deployment fast path [generate|actions]. Pass 2 must NOT request subgoal generation — training’s anti-copy coupling (suppress_subgoal) means condition-plus-generate never co-occurred, while condition-plus- fast-path is exactly the trained conditioned context.
• provenance: policy names carry the mode (_selfsubgoal, _oraclesubgoal, _narrsubgoal — a conditioned read must never pass as the deployment read, charter §2); report JSON records the mode; per-frame generated subgoals are retained machine-readable (frame identity triple → text) for the validity table and the results post’s qualitative block.
• oracles (abort-on-red before launch, the usual gate): (i) the self arm with its generated text forced EMPTY reproduces the baseline decode bit-exact (the no-hint limit is the historical path); (ii) the oracle arm on a label-less item ≡ baseline; (iii) conditioned prompt bytes match the training collator’s rendering of the same text (one rendering path, not a re- implementation); (iv) pass 2’s generate list excludes subgoal.

Design — two stages, gated in order

Stage 1 — validity table (eyes before any scalar; the never-generated-subgoal scar is the reason this stage exists and comes first). Generate self-subgoals for a fixed-seed (seed 0) sample of 60 panel frames stratified across episodes/repos, table: frame identity, instruction, TRUE segment label (or —), generated subgoal. Read and commented in the results post, not just attached. Pre-registered go/no-go for stage 2’s self arm: (a) non-empty, non-truncated text on ≥ 90% of rows; (b) no single generated string on > 50% of rows (degenerate-collapse check); (c) qualitatively subgoal-shaped — imperative manipulation clauses, not instruction echoes or judge-artifact fragments (eyes, commented row-by-row where failing). Fail → the self arm does NOT run; the failure is the rung-(a) result for generation quality, recorded with the table. The oracle arm runs regardless — it answers whether the slot is live at all, which stage-1 failure does not touch.

Stage 2 — scalar arms, identical rows via the shared plan file (plans/holdout_curated_v0_k4l2.json), AR-100k (bijou_arb_rcond_100k_ddp4/step_100000), seed 0, k4l2 panel:

Banked context row: the all-fields narrated arm is 5.8565 (+0.054 vs baseline — narrating everything slightly hurts).

Frozen reads

Paired per-row, seeded bootstrap 95% CI (seed 0, 10,000 resamples — the draws-fairness assembly conventions). “Labeled subset” = frames where subgoal_text is non-None; baseline re-pooled onto any subset from its banked npz (the state-probe precedent), never re-run.

1. Primary: Δ_self = chunk_mae(self-subgoal) − 5.8026 on all core frames (deployment-honest: a planner-less rig can always self-generate), with the labeled-subset value quoted beside it.
2. Bound: Δ_oracle = chunk_mae(oracle) − baseline on the labeled subset. Interpretive frame, fixed now: Δ_oracle bounds what the slot can transmit. Δ_oracle ≈ 0 ⇒ the slot is inert (0.5-dropout training made the hint ignorable) and rung (a) closes regardless of generation quality; Δ_oracle < 0 with Δ_self ≥ 0 ⇒ the gap is generation quality, not the slot.
3. Channel read: Δ_narr (suffix voice) vs Δ_self (prompt slot) — nearly the same text, different entry point; separates “where the text enters” from “whether text helps”.
4. Horizon decomposition: per-step-in-horizon MAE curves for every arm from --dump-predictions npz, the fontaine/scripts/flow_vs_ar_paired.py conventions. The Hi-VLA anchor predicts the gain concentrates LATE-horizon (their flat 25.30% → 67.08% gap is long-horizon-only), i.e. chunk-tail moves, first_mae barely.
5. first_mae mirrors of 1–3.
6. Execution oracles (abort): state-copy / state-copy-norm rows byte-match banked panel values; names/JSON carry the modes.

Numbered expectations (banked before data)

1. Δ_oracle < 0 on the labeled subset (a true hint helps a condition-trained model) — confidence medium.
2. Δ_self < 0 but |Δ_self| < |Δ_oracle| (self text is a noisy version of truth) — confidence medium-low; this is the probe’s genuinely open number.
3. The gain (if any) concentrates late-horizon; first_mae moves little — confidence medium (external anchor, our #1 precedent shows such predictions can half-fail).
4. Δ_narr ≈ +0.05-ish (suffix narration keeps slightly hurting, as the all-fields arm did) and does NOT beat Δ_self — confidence low (record-only comparison).
5. Falsified if Δ_self ≥ 0: explicit self-hierarchy gives nothing at panel granularity on this body. Recorded with the 2-vs-oracle diagnostic split; any escalation (rollout-granularity refresh policies, planner-side work, subgoal-quality training) needs a NEW pre-reg citing this result. No prompt fishing, no post-hoc subgoal re-phrasing.

Cost & scheduling

All arms are greedy panel evals (banked rate 0.081 s/frame → ~35 min each; the self arm ~2× for its two encode+decode passes). Stage 1 ≈ minutes. Pre-registered ceiling ≤ 8 GPU-h total; if a first-200- frame rate measurement projects past it, all arms drop to the frozen q4 subset (4,301 rows; the #19 clause verbatim) and the switch is recorded. Venue: local GPU, first quiet window at or after the draws10_t1 boundary AND its frozen reads (that pre-reg’s obligations come first); never co-located with a training run’s eval chain. First-poll util+rate check per standing rule.

Amendment 1 — oracle-(i)/(ii) comparator (2026-08-08 ~00:1xZ, posted BEFORE the stage-2 launch)

The pre-launch adjudication run falsified an assumption inside the oracle spec, not the oracle’s semantics. Recorded here per this post’s own amendment clause; no read, arm, gate or expectation changes.

The live oracle-(i) run (q4 subset, --selfsubgoal-force-empty) was NOT bit-exact against the banked full-panel baseline npz: 1207/4,301 rows differed. Diagnosis, in order run:

• state-copy / state-copy-norm rows byte-match the banked panel (plan/data alignment exact); differing rows hit subgoal-labeled and label-less rows at the same rate (no hint-leak signature);
• the pooled effect of the differing rows is −0.0008 chunk MAE (per-frame CI [−0.016, +0.015] — mean-zero decode noise, not a systematic shift);
• a plain baseline eval on the same q4 plan with zero instrument code involved reproduces the same class of row flips against the banked npz (count quoted in the adjudication log). Mechanism: greedy AR decode is batch-composition-sensitive at the kernel level — padding/shape-dependent reduction order perturbs logits at ulp scale, near-tie argmaxes flip, and one flipped token cascades through the row. Action quantiles are per-item (verified at interface.py::_stats), so bins are composition-independent — kernel numerics are the only channel.

Amended comparator (semantics unchanged: “the no-hint limit is the plain path”): oracle (i) is adjudicated bit-exact against a plain baseline decode of the SAME plan at the SAME batch composition (the stateprobe_q4_diagbaseline run), not against the banked full-panel npz. Oracle (ii)’s label-less half is decode-checkable only up to the same composition noise (label-bearing batchmates change padding), so its abort-grade live form is the wiring check (≥ 1 labeled row moves the decode) and the label-less byte-equality stands on the pinned CPU prompt-byte oracle; the label-less decode count is recorded descriptively. All composition-independent execution oracles keep their banked comparator and abort grade: identity columns, state-copy rows, provenance fields.

Consequence for the frozen reads: none. The stage-2 arms run the full panel — the SAME plan, order and batch size as the banked baseline — and the paired reads keep the banked npz as baseline, exactly as frozen. The measured mean-zero composition noise (−0.0008 pooled, CI ±0.016 per-frame) is quoted in the results post beside Δ_self as the decode-noise floor context; it does not modify the read definitions or the E5 falsifier.