# ARC-AGI Abstraction Pre-Registration

Roadmaps:
[`SUNDOG_V_ARC.md`](../../SUNDOG_V_ARC.md),
[`SUNDOG_V_GRAVITY.md`](../../SUNDOG_V_GRAVITY.md) Candidate 14

Filed: **2026-05-28 (PT)**

Status: **Branch E (deterministic program search) verdict: `branch_e_capability_demonstrated`** (pinned `6886074`). On the 108-task expanded register, a deterministic typed search over registered transformation primitives — selecting programs by train-pair consistency, never by signature geometry — cleared the established non-trivial floor (≥2 distinct held-out tasks solved exactly on both `test_lodo` and `pttest`) that every prior decoder/learner family floored on at zero. The result is **modest** (2 distinct tasks/gated lane, ~3% instance rate; structurally simple tasks via component-extraction, crop+scale, tile/crop+recolor) and is **not** a Blackwell-sufficiency proof, an ARC solve, or any eval/Kaggle claim. **Branch E v2** (deterministic expansion: intricate masks + morphology + depth-3, pinned `a307340`, ~101m, `gitDirty=false`) returned **`branch_e_v2_capability_replicated`**: it retained both v1 gated solves on both lanes but solved **0 new tasks** (material-lift gate ≥4 + ≥2-new not met) and even dropped validation 2→1/lane via top-2 crowding — so the deterministic program-search capability is **bounded near the v1 baseline**; lifting it would need a different solver class. **Branch E3** (a learned MLP ranker over the *frozen* v2 candidate set, targeting that top-2 crowding) is **built + byte-verified but compute-paused** (Amendment A): `enumerate_admitted` reproduces v2's candidate set exactly and the `v2_deterministic_selector` control reproduces the v2 solves, but a timing probe measured ~47.5 s/instance → ~48 h over the frozen ~892-task aux pool, so the sharded binding run was deliberately not launched (tracked in [`docs/TODO.md`](../../TODO.md)). The four Phase 3E certificate verdicts and the seven full-grid-control floors stand; the certificates are cited as a no-collision / no-locality boundary, not a geometric selector.
Phase 0 admitted; Phase 1 synthetic gate strengthened and passed; Phase 2
projection-measurement plus baseline-comparison passed; Phase 3 filed
three deterministic-low-capacity binding receipts (`nn_output_transfer_v1`,
`nn_delta_transfer_v1`, `candidate_combinator_v1`), all verdict
**task hardness / decoder failure**, and one Blackwell sufficiency lane.
The three-receipt convergence is characterised at
[`PHASE3_5_REFLECTION.md`](PHASE3_5_REFLECTION.md) as a finding about
the deterministic-low-capacity-learner family rather than a
sufficiency-failure conclusion about the shadow-projection representation.
The Blackwell sufficiency amendment in
[`PHASE3_SUFFICIENCY_SPEC.md`](PHASE3_SUFFICIENCY_SPEC.md) filed algebraic
conditions, held-out-task splits, Branch A/B/C criteria, a frozen
Python/PyTorch decoder design, a timing probe, and a full clean receipt for
`blackwell_task_decoder_v1` -- **Branch C bounded failure**:
`signature_palette` scored zero exact matches on both held-out-task LODO
and held-out public-training test lanes, and the matched `raw_grid_lowcap`
control also scored zero exact matches. The strengthened follow-up
`blackwell_publictrain_rawgrid_gate_v2` was launched as a raw-grid-only
control gate with all 971 admitted public-training auxiliary tasks (29
excluded by the registered `MAX_DEMOS = 5` token cap), 3 seeds, 120-epoch
budget, on GPU under a shard+merge protocol with byte-equivalence to a
serial run; freeze-marker commit `79C5B060`. The V2 binding receipt is
**`full_grid_control_floor`**: every held-out lane (`pttest`, `test_lodo`,
`validation_lodo`, `validation_pttest`) scored `grid_exact_any_rate = 0.0`
at selected seed `20260529`. The shape-exact-slot1 rate sits at 0.50-0.72
across lanes while palette-exact-slot1 holds at 0.0 -- the same
shape-matches-content-fails character as V1, now reproduced under the
strengthened V2 lane.

A Branch B compact-subset diagnostic lane was then admitted with a
narrower question: does the registered decoder produce any exact-match
signal on the 7 Phase 2 compact-signal tasks? The compact-7 single-seed
binding receipt (
[`PHASE3B_COMPACT_SUBSET.md`](PHASE3B_COMPACT_SUBSET.md), seed `20260529`,
freeze-marker `50EAEBBF`) returned verdict
**`compact_full_grid_control_floor`** with a qualitatively distinct
failure character: every held-out instance gets the output shape exactly
right (`shape_exact_slot1 = 1.000` on all four lanes) and pixel accuracy
jumps to 0.757-0.877, but every prediction collapses to the dominant
background color (palette_exact_slot1 = 0.000 everywhere; 13 of 13
held-out predictions slot-1-use at most 2 colors regardless of target
palette size). This is filed as a named failure mode -- **"dominant-color
mode collapse"** -- distinct from V1/V2's noise-dominated character.
Per the pre-registered §8 stop rule, this verdict closes Branch B in the
deterministic-low-capacity-learner family: `signature_palette` does not
run on this subset, additional seeds do not run, and the path forward
is PHASE3_5_REFLECTION Branch A (stochastic per-task learner) or
Branch D (different framing) rather than further narrowing within the
deterministic family. Kaggle notebook work and public-evaluation grid
inspection remain blocked until Phase 6.

Branch A is filed at
[`PHASE3A_STOCHASTIC_PER_TASK_SPEC.md`](PHASE3A_STOCHASTIC_PER_TASK_SPEC.md).
`per_task_coord_mlp_v1` is a stochastic per-instance coordinate MLP trained
from scratch on only that instance's conditioning demonstrations. The
20-shard binding receipt (4 arms x 5 seeds, ~2.1 h GPU parallel wall on a
GTX 1080, sharded with the spec's (arm, seed) shard+merge protocol and the
`--allow-mixed-commits` operator override after parallel Navier-Stokes
commits landed mid-flight) returned verdict **`branch_a_full_grid_floor`**:
zero exact tasks on both held-out lanes for arm `raw_grid_per_task`, so no
signature_palette_per_task vs. raw_grid_per_task sufficiency comparison is
licensed. Failure character is qualitatively distinct from V1, V2, and
compact-7: **conditioning starvation + shape-generalisation failure**
(61% of held-out instances quarantined as `insufficient_conditioning_pairs`
because registered ARC tasks have k=2-3 train pairs, k-1<=2 conditioning
after LODO). All four arms achieve identical shape_exact_slot1 rates
(0.500 pttest, 0.474 test_lodo) -- the per-task scratch shape MLP memorizes
conditioning shapes regardless of arm. The per-task coordinate-MLP color
head does NOT collapse to background like compact-7 did (max collapse rate
16.7% on pttest, 10.5% on test_lodo).

Four Phase 3 binding full-grid-control receipts -- V1, V2, compact-7, and
Phase 3A -- now agree on the held-out exact-grid floor across two task
distributions and two learner families (transformer, per-task coordinate
MLP). The remaining admissible Phase 3 reopen path is
PHASE3_5_REFLECTION Branch D (different framing). A future Branch D would
need its own pre-registered spec, arena gate, and learner contract.

Branch D is filed at
[`PHASE3D_DIFFERENT_FRAMING_SPEC.md`](PHASE3D_DIFFERENT_FRAMING_SPEC.md).
`structured_edit_residual_v1` models each output as an input-derived baseline
canvas plus a residual edit mask and edit colors. The 20-shard binding
receipt (4 arms x 5 seeds, ~1h 26m GPU parallel wall via the (arm, seed)
shard+merge protocol with `--allow-mixed-commits` operator override --
3 distinct gitCommits, but `runnerIdenticalAcrossCommits=true` so no WARN
fired) returned verdict **`branch_d_full_grid_edit_floor`**: zero
non-baseline exact tasks on both held-out lanes for arm `raw_grid_edit`,
so no `signature_palette_edit` vs. `raw_grid_edit` sufficiency comparison
is licensed. Failure character is qualitatively distinct from V1, V2,
compact-7, and Phase 3A: **edit-color-rule failure**. The structured-edit
framing recovers shape (1.000 on pttest), partial palette (0.333), pixel
accuracy 0.47-0.57, and edit-mask F1 0.53-0.65 with strong minority recall
(0.61-0.72) -- but the per-instance scratch color learner cannot recover
the exact edit colors. The dominant quarantine label (26% of held-out
instances) is `edit_color_failure` (mask F1 >= 0.50 but color accuracy
< 0.50): the model knows WHERE to edit but not WHAT color.

Five Phase 3 binding full-grid-control receipts -- V1, V2, compact-7,
Phase 3A, and Phase 3D -- now agree on the held-out exact-grid floor
across two task distributions, two learner families (transformer +
per-task MLP), and two output framings (whole-grid + structured edit
residual). The structured-edit framing additionally decomposed the
failure mode into its components (shape picker, canvas picker, edit-mask
learner, edit-color learner) for the first time, isolating the bottleneck
to the edit-color-rule stage. All four PHASE3_5_REFLECTION branches plus
Branch D are now characterised. The first Branch D variant receipt later
shifted the bottleneck to the mask stage, and a second Branch D variant is now
pre-registered to target that mask stage. Any executed reopen still requires
its own arena gate and verdict-amendment discipline.

The first bottleneck-targeted Branch D variant is filed at
[`PHASE3D_EDIT_COLOR_RULE_VARIANT_SPEC.md`](PHASE3D_EDIT_COLOR_RULE_VARIANT_SPEC.md).
`structured_edit_color_rule_v2` replaces only the failed edit-color MLP
with a deterministic conditioning-derived color-rule bank
(`edit_color_rule_bank_v1`) across 10 frozen rule families with LOCO
scoring + spec'd tie-break chain + optional top-3 ensemble. The baseline
picker, edit-mask learner, splits, seed slate, and arena discipline are
inherited unchanged from `structured_edit_residual_v1`. The 20-shard
binding receipt (4 arms x 5 seeds, ~1h 6m GPU parallel wall via the
inherited shard+merge protocol with `--allow-mixed-commits` operator
override -- 3 distinct gitCommits, `runnerIdenticalAcrossCommits=true`,
no WARN) returned verdict **`branch_d_color_rule_full_grid_floor`**:
zero non-baseline exact tasks on both held-out lanes for every arm.

But the variant **achieved its design goal at the diagnostic level**:
palette_exact roughly doubled vs the Phase 3D base (pttest 0.667-0.833
vs 0.333; test_lodo 0.632 vs 0.316), and the `edit_color_failure` mode
named by the Phase 3D base receipt was **eliminated** as a quarantine
label. The dominant failure shifted from edit-color (26% in Phase 3D
base) to **edit_mask_failure (41%)**, with the new diagnostic pair
`color_rule_selection_failure` (16%, oracle >= 0.50 but selected < 0.50)
vs `color_rule_bank_coverage_failure` (9%, oracle < 0.50) firing in a
1.8:1 ratio and `rule_selection_regret_mean` of 0.30-0.35 indicating
~30 percentage points of edit-color accuracy locked up in selector
improvements. The variant thus provides the first **quantitative
bottleneck decomposition** in Phase 3 (41% mask / 16% selection regret
/ 9% bank coverage), telling the next surgical move with unusual
specificity: mask-extension variant (biggest leverage), selection-
refinement variant (large locked accuracy), or rule-bank extension
(smallest expected payoff).

The mask-targeted Branch D variant (`structured_edit_mask_target_v3` /
`edit_mask_candidate_bank_v1`) then replaced the trained edit-mask MLP
with a 13-family deterministic mask candidate bank, keeping the
deterministic color bank from the prior variant. The 20-shard binding
receipt (4 arms x 5 seeds, ~1h 58m GPU parallel wall via shard+merge
with `--allow-mixed-commits` -- 8 distinct gitCommits,
`runnerIdenticalAcrossCommits=true`, no WARN; a runner O(C^2) LOCO-rescore
blowup was found and fixed in the probe phase) returned verdict
**`branch_d_mask_target_full_grid_floor`** -- the **seventh** Phase 3
full-grid control floor. The central finding is a clean negative result:
**the mask repair did not shift the bottleneck off the mask.** Mask-stage
quarantine labels still dominate (137 of ~196 selected-seed labels:
`mask_selection_failure` 29%, `mask_overedit_failure` 26%,
`mask_candidate_coverage_failure` 12%, `mask_underdetection_failure` 4%),
the deterministic mask bank does not dominate the learned mask MLP (the
inherited `legacy_mlp_threshold_mask` family still won 13.8% of
selections, and `test_lodo` mask F1 dropped to 0.49-0.51 vs the
color-rule variant's 0.53-0.58), and the bank traded higher minority-edit
recall (0.70-0.79) for over-editing.

Seven Phase 3 binding full-grid-control receipts -- V1, V2, compact-7,
Phase 3A, Phase 3D base, Phase 3D color-rule variant, and Phase 3D
mask-target variant -- now agree on the held-out exact-grid floor across
two task distributions, three learner families, two output framings, and
-- within the structured-edit framing -- both the learned and the
deterministic-bank variants of *each* edit component (mask and color).
With both named bottlenecks of the structured-edit framing probed by
deterministic banks and the floor holding at both, **Branch E (a
different framing entirely -- e.g., generative program / DSL search over
conditioning pairs, or test-time prompting of a frozen large LM) is the
live frontier.** A selection-targeted within-framing variant (a joint
mask+color selector, since both banks fail more by selection than by
coverage) remains the smaller alternative. Any reopen needs its own
pre-registered spec, arena gate, and verdict-amendment discipline.

## Official Anchors

Checked **2026-05-28** against:

- [ARC Prize 2026 overview](https://arcprize.org/competitions/2026): competition
  start March 25, 2026; submissions due November 2, 2026; papers due
  November 8, 2026; results announced December 4, 2026.
- [Paper Prize](https://arcprize.org/competitions/2026/paper): paper
  submissions must link to a Kaggle code submission for ARC-AGI-2 or ARC-AGI-3;
  a high score is not required for eligibility, but the score feeds the
  Accuracy rubric.
- [ARC-AGI-2 track](https://arcprize.org/competitions/2026/arc-agi-2): static
  grid reasoning track; Kaggle notebook submission; no internet during
  evaluation; two predicted outputs per test input; exact-match task score.
- [Official ARC-AGI-2 repo](https://github.com/arcprize/ARC-AGI-2): 1,000
  public training tasks and 120 public evaluation tasks, with private sets held
  out for the competition.

## Current Phase Artifacts

- [`PHASE0_TASK_SUBSET_SPEC.md`](PHASE0_TASK_SUBSET_SPEC.md) -- frozen Phase 0
  work order for task inventory, subset registration, baselines, and
  evaluation leak control. Two amendments on file (PARTIAL ADMIT then ADMIT).
- [`P0_BASELINES.md`](P0_BASELINES.md) -- Phase 0 inventory/register/baseline
  receipt with both verdicts (frozen 3-baseline PARTIAL ADMIT, then 6-baseline
  ADMIT after baseline expansion).
- [`P0_TASK_REGISTER.csv`](P0_TASK_REGISTER.csv) -- 36 public-training tasks,
  6 per registered prior, all manually inspected.
- [`PHASE0_CONTEXT_EXPANSION_FOR_FIBERS_SPEC.md`](PHASE0_CONTEXT_EXPANSION_FOR_FIBERS_SPEC.md)
  -- Phase 0 register-amendment spec for the Phase 3E v2 sparsity blocker. It
  keeps `program_sketch_v2`, `epsilon_primary=0.05`, k=3, and all context
  geometry frozen, and pre-registers a balanced 108-task public-training
  expansion target (18 per prior). **Expansion ADMITTED**
  (`phase0_fiber_expansion_admit`, Amendment B): a candidate-ordering script
  emitted a frozen queue (sha256 `43ea4550`), the maintainer's agent inspected
  it in queue order (0 hard-excludes, 0 discipline tripwires, coordinator
  spot-verified), and the binding **108-task expanded register**
  (`P0_TASK_REGISTER_EXPANDED_FOR_FIBERS.csv`, sha256 `7c56141c`; 36 original +
  72 new, 18/prior) is filed. The runner gained `--split-mode sha256_expansion`
  (geometry untouched; default `frozen_v2` byte-reproduces the v2 receipt).
  **Expanded certificate verdict (Amendment C, pinned `a0f1a4b`,
  `runnerSha256 9D456143`, ~5m47s, deterministic):
  `phase3e_v2_expanded_oracle_regression`.** The expansion densified the context
  universe (U_primary 25 -> 336, U_all 49 -> 491) but produced two findings:
  (1) the v2 oracle's **anti-solver-leakage gate fails** on the larger register
  (`core_sketch_exact_lookup_fraction 0.04 -> 0.351`, threshold 0.20; vacuity and
  laundering still pass) -- the gate calibration does not transfer from 36 to 108
  tasks; (2) the fibers stay **sparse** -- tripling the register contracted the
  min cross-task distance only `0.207 -> 0.196` (still 0 near pairs, 0 fidelity,
  0 collisions), far from the ~4x needed for `epsilon_primary=0.05`. The verdict
  neither promotes nor blocks signature sufficiency or a Branch E solver; the
  36-task v2 receipt and the seven full-grid floors stand. Next moves (each needs
  its own pre-registered spec): a register-size-robust oracle v3, accepting the
  sparsity result as durable, or Branch E on capability grounds.
- [`PHASE3E_RELATIVE_LOCALITY_CERTIFICATE_SPEC.md`](PHASE3E_RELATIVE_LOCALITY_CERTIFICATE_SPEC.md)
  -- Phase 3E rank-based locality spec. It keeps the 108-task expanded register,
  `program_sketch_v2`, and all absolute-epsilon findings frozen, but changes the
  certificate question from "are there pairs within `epsilon_primary=0.05`?" to
  "are cross-task nearest neighbors under `signature_palette_context` more
  program-sketch-coherent than controls?" Primary k is 5, with k=1/3/10
  diagnostics, metadata/raw/random/permutation controls, and explicit
  leakage-caveated public language. **Binding verdict (Amendment B, pinned
  `f7850f0`, runnerSha256 `3DD2C97A`, U_primary=336, 5 seeds x 1000 perms, ~56m):
  `phase3e_relative_locality_negative`.** At k=5, palette mean neighbor sketch_sim
  0.420 vs metadata 0.414 (Δ +0.006), signature_only 0.406, raw_grid **0.456**
  (palette is *below* raw_grid by 0.036). Palette is significantly above
  prior-stratified random (p=0.0002) but the effect is small (Δ vs uniform +0.072
  < 0.15 bar), metadata-tied, raw-grid-dominated, sign-unstable across k, and the
  palette hard-incompatibility rate **0.285 > 0.25** fires the negative gate.
  `oracle_invalid=False` (vacuity/laundering pass; the 108-task leakage regression
  `exact_lookup 0.351` is carried as a diagnostic, not retuned). The rank test
  **strengthens** the absolute-sparsity result: no usable fiber or rank-local
  locality for `signature_palette` on the expanded register (and raw_grid edges it
  on both rank-local metrics).
- [`PHASE3_BRANCH_E_PROGRAM_SEARCH_SPEC.md`](PHASE3_BRANCH_E_PROGRAM_SEARCH_SPEC.md)
  -- Branch E, the first **solver** lane (capability grounds). A deterministic
  typed search composes registered transformation primitives (structural
  transforms + the imported Branch D color-rule bank + simple deterministic mask
  families + candidate-combinator candidates), admitting only programs consistent
  with **all** of a task's conditioning train pairs (selection is train-pair
  consistency, never signature geometry); the four Phase 3E certificates are cited
  as a no-collision / no-locality **boundary**, not a geometric-selector
  motivation. **Binding verdict (Amendment B, pinned `6886074`, runnerSha256
  `036EA6A9`, gitDirty=false, U_primary=336, ~2m38s, deterministic):
  `branch_e_capability_demonstrated`.** It cleared the established non-trivial
  floor — ≥2 distinct held-out tasks solved exactly on both `test_lodo` and
  `pttest` (`be94b721` via `extract_largest_component`, `f25fbde4` via
  `crop_bbox>>scale`; validation diagnostic also 2) — that every prior
  decoder/learner family floored on at zero. The result is **modest** (2 distinct
  tasks/gated lane, ~3% instance rate; 3 of 4 solves are depth-2 compositions) and
  is explicitly **not** a Blackwell-sufficiency proof, an ARC solve, or any
  eval/Kaggle claim; the certificate verdicts and seven floors stand. A v2 may
  admit the deferred intricate mask families + deeper composition to test whether
  the rate rises above this floor-clearing baseline.
- [`PHASE3_BRANCH_E_V2_PROGRAM_SEARCH_SPEC.md`](PHASE3_BRANCH_E_V2_PROGRAM_SEARCH_SPEC.md)
  -- Branch E v2, a deterministic expansion of the v1 solver lane. It admits
  the exact v1-deferred deterministic mask families (`row_col_periodic_mask`,
  `source_color_pair_mask`, `object_role_mask`,
  `nearest_residual_patch_mask`), the morphology cross-product, and depth-3
  structural composition. **Binding verdict (Amendment B, pinned `a307340`,
  runnerSha256 `1D486AE1`, gitDirty=false, U_primary=336, ~101m, deterministic):
  `branch_e_v2_capability_replicated`.** v2 retained both v1 gated solves
  (`be94b721`, `f25fbde4`) on both lanes but solved **0 new tasks** (material-lift
  gate ≥4-distinct + ≥2-new not met), and the larger search space actually
  **dropped validation 2→1/lane** via top-2 crowding (more train-consistent-but-
  test-wrong programs displace the correct one). So the deterministic
  program-search capability is **bounded near the v1 ~3% baseline**: more
  deterministic families/depth crack no new held-out task here. Same caveats —
  capability on held-out public-training, not sufficiency/solve/eval/Kaggle;
  certificates + seven floors + the v1 verdict stand. Lifting the rate would need a
  different solver class (learned ranker, richer object-centric primitives, or an
  LM proposer) under its own spec.
- [`EVAL_BLIND_SELECTION.md`](EVAL_BLIND_SELECTION.md) -- stub pattern for
  future Phase 1+ evaluation-blind register rows (no manual grid inspection,
  selection by preregistered metadata/hash rule).
- [`PHASE1_SHADOW_DOMAIN_SPEC.md`](PHASE1_SHADOW_DOMAIN_SPEC.md) -- Phase 1
  discrete grid shadow-domain spec. Two amendments: original 5-fixture
  synthetic gate (later audited as construction-only) and a strengthened
  9-fixture + 50-grid discrimination gate that adds falsifiable cases
  (1-cell flip, color collision, stencil-bag invariance, signature
  discrimination). Phase 2 projection scaffold admitted with falsifiable
  support.
- [`PHASE2_PROJECTION_SPEC.md`](PHASE2_PROJECTION_SPEC.md) -- Phase 2
  registered-subset projection measurement receipt with baseline-comparison
  addendum. `36` tasks / `266` grids projected; shadow operator collapses 6
  byte-distinct grids correctly as gauge equivalents, but its train-pair
  residual (`0.594`) is the highest of all four representations tested
  (raw-pixel Hamming `0.279`, shape/palette/density `0.103`, cell-count
  `0.071`). Phase 3 sufficiency-spec writing admitted with the harder
  framing: signature space does not shorten the input-output gap, so
  Phase 3 must explicitly model the transformation. No decoder or
  public-evaluation work admitted by this receipt.
- [`PHASE3_SUFFICIENCY_SPEC.md`](PHASE3_SUFFICIENCY_SPEC.md) -- Phase 3
  sufficiency audit spec. It freezes the harder post-Phase-2 posture:
  `input_sig -> output_sig` is not a small-delta problem, and the primary
  Sundog representation is `signature_palette`, not signature alone.
  Three binding receipts filed (`nn_output_transfer_v1`,
  `nn_delta_transfer_v1`, `candidate_combinator_v1`), all verdict
  task hardness / decoder failure. A Blackwell sufficiency lane is also
  filed: the full clean `blackwell_task_decoder_v1` receipt is Branch C
  bounded failure with the raw-grid exact-floor caveat. The strengthened
  follow-up `blackwell_publictrain_rawgrid_gate_v2` is filed as a
  raw-grid-only control gate with all 971 admitted public-training aux
  tasks, 3 seeds (`20260528/20260529/20260530`), 120 epochs, GPU
  shard+merge protocol with shard-equivalence verification, freeze-marker
  commit `79C5B060`. Selected seed `20260529` (val_loss 1.6249,
  validation_metric ties at 0.0). V2 binding-receipt verdict:
  **`full_grid_control_floor`** -- `grid_exact_any_rate = 0.0` on every
  held-out lane (`pttest`, `test_lodo`, `validation_lodo`,
  `validation_pttest`); `shape_exact_slot1` 0.50-0.72 while
  `palette_exact_slot1 = 0.0`, the same shape-matches-content-fails
  character as V1. A Branch B compact-subset diagnostic lane is then
  admitted and the compact-7 single-seed receipt is filed (see next
  bullet); Phase 3 then pauses with Branch B closed.
- [`PHASE3B_COMPACT_SUBSET.md`](PHASE3B_COMPACT_SUBSET.md) +
  [`PHASE3B_COMPACT_SPLIT.csv`](PHASE3B_COMPACT_SPLIT.csv) -- Phase 3
  Branch B compact-subset diagnostic lane. 7 Phase 2 compact-signal
  tasks, pre-registered 4/1/2 internal split, minimal floor (≥1 exact
  task on `pttest` AND `test_lodo`), decoder/hyperparams/seed slate/aux
  pool frozen relative to V2. Compact-7 single-seed binding receipt
  filed under freeze-marker `50EAEBBF` with seed `20260529` (best_epoch
  55, val_loss 0.7628). Verdict: **`compact_full_grid_control_floor`**.
  Failure character is qualitatively distinct from V1/V2: every
  held-out instance has `shape_exact_slot1 = 1.000` and pixel accuracy
  0.757-0.877, but every prediction collapses to the dominant
  background color (13 of 13 held-out predictions slot-1-use at most 2
  colors regardless of target palette size). Named failure mode filed:
  **"dominant-color mode collapse"**. Per the pre-registered §8 stop
  rule, Branch B closes in the deterministic-low-capacity-learner
  family; no additional seeds, no `signature_palette` arm on the
  compact subset, no further raw-grid V-bumps within the family.
- [`PHASE3A_STOCHASTIC_PER_TASK_SPEC.md`](PHASE3A_STOCHASTIC_PER_TASK_SPEC.md)
  -- Branch A stochastic per-task learner spec. It reserves
  `per_task_coord_mlp_v1`, freezes the `raw_grid_per_task`,
  `signature_palette_per_task`, `signature_only_per_task`, and
  `metadata_only_per_task` arms, requires the raw-grid arm to open the arena
  before any signature sufficiency comparison, and adds dominant-color collapse
  audits. **Binding 20-shard receipt filed**: `branch_a_full_grid_floor` at
  selected seeds (raw=`20260530`, sig_palette=`20260528`, sig_only=`20260530`,
  metadata=`20260601`). Zero exact tasks on both held-out lanes for any arm;
  all four arms achieve identical shape_exact_slot1 rates (0.500/0.474).
  Failure mode: **conditioning starvation + shape-generalisation failure**
  (61% of held-out instances quarantined as `insufficient_conditioning_pairs`).
  Run executed via the (arm, seed) shard+merge protocol with the operator
  `--allow-mixed-commits` override (parallel Navier-Stokes commits landed
  mid-launch; runner-content audit verified shard-time computational contract
  unchanged across 6 distinct gitCommits / 2 distinct runner SHAs;
  `mixedCommitsAudit` recorded in the merged manifest). No
  signature_palette vs. raw_grid sufficiency comparison licensed.
- [`PHASE3D_DIFFERENT_FRAMING_SPEC.md`](PHASE3D_DIFFERENT_FRAMING_SPEC.md)
  -- Branch D structured edit/residual framing spec. It freezes
  `structured_edit_residual_v1`, with `raw_grid_edit`,
  `signature_palette_edit`, `signature_only_edit`, and `metadata_only_edit`
  arms, a conditioning-selected baseline family, edit-mask/edit-color residual
  learners, and a non-baseline arena gate. **20-shard binding receipt
  filed** under `mergeGitCommit 9F3193D7` (3 distinct gitCommits via
  `--allow-mixed-commits` override; `runnerIdenticalAcrossCommits=true` so
  no WARN; 14 dirty + 6 clean shards). Selected seeds: raw=`20260529`,
  sig_palette=`20260531`, sig_only=`20260529`, metadata=`20260528`.
  Verdict: **`branch_d_full_grid_edit_floor`**. Zero non-baseline
  exact tasks on both held-out lanes for every arm; baseline_exact also
  zero (the baselines never produce the exact target). Shape exact 1.000
  on pttest, 0.316 on test_lodo; palette exact 0.333; pixel mean
  0.47-0.57; edit-mask F1 0.53-0.65; minority edit recall 0.61-0.72.
  6 of 9 quarantine labels fire; dominant label is **`edit_color_failure`**
  (51/196 = 26%), naming the new failure mode: the per-instance scratch
  color learner knows WHERE to edit but not WHAT color. Wall: ~1h 26m
  parallel (2.9x sharded GPU speedup vs ~5.3 h serial projection).
- [`PHASE3D_EDIT_COLOR_RULE_VARIANT_SPEC.md`](PHASE3D_EDIT_COLOR_RULE_VARIANT_SPEC.md)
  -- first Branch D bottleneck variant. It freezes
  `structured_edit_color_rule_v2` / `edit_color_rule_bank_v1`, keeps the
  baseline picker and edit-mask learner inherited from
  `structured_edit_residual_v1`, and replaces only the edit-color MLP with a
  deterministic color-rule bank (10 frozen families; LOCO scoring; spec'd
  tie-break chain; optional top-3 ensemble when 3+ candidates tie within
  0.05 conditioning accuracy). **20-shard binding receipt filed** under
  `mergeGitCommit 8177A578` (3 distinct gitCommits via
  `--allow-mixed-commits` operator override; `runnerIdenticalAcrossCommits=
  true` so no WARN; 9 dirty + 11 clean shards; ~1h 6m parallel GPU wall =
  2.89x sharded speedup). Selected seeds: raw=`20260528`, sig_palette=
  `20260528`, sig_only=`20260528`, metadata=`20260531`. Verdict:
  **`branch_d_color_rule_full_grid_floor`** (zero non-baseline exact tasks
  on any held-out lane). But: **the variant achieved its diagnostic design
  goal** -- palette_exact roughly doubled vs the base
  (pttest 0.667-0.833 vs 0.333; test_lodo 0.632 vs 0.316) and the
  `edit_color_failure` quarantine label was eliminated. The failure shifted
  to a quantitative 3-component bottleneck: **41% `edit_mask_failure`**
  (mask MLP now dominant blocker), **16% `color_rule_selection_failure`**
  (oracle >= 0.50 but selector picked wrong;
  `rule_selection_regret_mean = 0.30-0.35`), **9%
  `color_rule_bank_coverage_failure`** (no rule in bank covers). Selected
  rule family distribution across 980 rows: `ensemble_top3` 40.8%,
  `nearest_edited_neighbor_color` 36.7%, `relative_palette_rank_map` 20.4%,
  `row_col_periodic_color` 2.0%. Per-arm: signature_palette slightly higher
  on pixel + color_rule_accuracy than raw_grid, but identical
  shape/palette_exact (baseline-picker-dominated, arm-invariant).
- [`PHASE3D_MASK_TARGET_VARIANT_SPEC.md`](PHASE3D_MASK_TARGET_VARIANT_SPEC.md)
  -- second Branch D bottleneck variant. It freezes
  `structured_edit_mask_target_v3` / `edit_mask_candidate_bank_v1`, keeps the
  baseline picker and deterministic edit-color rule bank inherited from
  `structured_edit_color_rule_v2`, and replaces only the edit-mask predictor
  with a deterministic conditioning-derived 13-family mask-candidate bank. It
  targeted the largest remaining bottleneck from the edit-color-rule receipt
  (41% `edit_mask_failure`). **20-shard binding receipt filed** under
  `mergeGitCommit 07F29513` (8 distinct gitCommits via `--allow-mixed-commits`;
  `runnerIdenticalAcrossCommits=true`, no WARN; ~1h 58m parallel GPU wall =
  2.88x sharded; an O(C^2) LOCO-rescore blowup was found and fixed in the
  probe phase by generating each fold's candidate bank once). Selected seeds:
  raw=`20260530`, sig_palette=`20260529`, sig_only=`20260529`,
  metadata=`20260531`. Verdict: **`branch_d_mask_target_full_grid_floor`**
  (the **7th** Phase 3 floor; zero non-baseline exact tasks on any held-out
  lane). Central finding -- **the mask repair did NOT shift the bottleneck off
  the mask**: mask-stage quarantine labels still dominate (137/196:
  `mask_selection_failure` 56, `mask_overedit_failure` 50,
  `mask_candidate_coverage_failure` 24, `mask_underdetection_failure` 7), the
  deterministic mask bank does not dominate the learned mask MLP (legacy MLP
  family still won 135/980 = 13.8% of selections; `test_lodo` mask F1 fell to
  0.49-0.51 vs the color-rule variant's 0.53-0.58), and it traded higher
  minority-edit recall (0.70-0.79) for over-editing. `mask_selection_failure :
  mask_candidate_coverage_failure` = 2.3:1 -- same "candidate exists but
  selector misses it" shape as the color stage. Selected mask family
  distribution: `conditioning_mask_union` 60.8%, `legacy_mlp_threshold_mask`
  13.8%, `conditioning_bbox_fill` 9.8%, then 5 others. With both the mask and
  color stages now deterministic banks and the floor holding at both, Branch E
  is the live frontier.
- [`PHASE3E_SIGNATURE_FIBER_CERTIFICATE_SPEC.md`](PHASE3E_SIGNATURE_FIBER_CERTIFICATE_SPEC.md)
  -- Phase 3E certificate spec. It freezes the first Branch E question:
  whether registered ARC contexts contain exact or near
  `signature_palette_context` fiber collisions with incompatible required
  behavior. This is not a solver lane and does not train a decoder. It uses the
  primary context universe `test_lodo union pttest`, reports the full
  validation+test universe diagnostically, and branches into exact-fiber
  collision, near-fiber incompatibility, fiber-locality positive, sparse-fiber
  deferral, or label-vacuity deferral. Its binding receipt returned
  **`phase3e_deferred_label_vacuity`**: no registered exact or near
  signature-fiber collision was found, but the registered fibers are sparse at
  the frozen radius and `program_sketch_v1` is prior-blind on 68% of primary
  contexts.
- [`PHASE3E_PROGRAM_SKETCH_ORACLE_V2_SPEC.md`](PHASE3E_PROGRAM_SKETCH_ORACLE_V2_SPEC.md)
  -- Phase 3E oracle-repair spec. It freezes a framing-agnostic
  `program_sketch_v2` labeler with nine facets, anti-vacuity,
  anti-prior-laundering, and anti-solver-leakage gates, and the same frozen
  Phase 3E signature geometry thresholds. **Binding receipt filed** (pinned to
  freeze-marker `3B1B76D2`, runnerSha256 `6163C208`, ~8 s deterministic,
  U_primary=25): verdict **`phase3e_v2_deferred_sparse_fibers`**. The v1 oracle
  defect is **repaired and certified** -- all three gates pass: anti-vacuity
  vacuous fraction **0.00** (v1 was 0.68; every prior 0-vacuous across both
  lanes), anti-prior-laundering 0% violations (min extra facets 5),
  anti-solver-leakage clean (`unique_core_sketch_fraction 0.36` <= 0.60,
  `core_sketch_exact_lookup_fraction 0.04` <= 0.20). Facet richness 6-7 of 9
  per primary context; the four priors that were 100% vacuous under v1
  (`counting`, `local_completion`, `spatial_transform`, `symmetry`) are now
  0/4-0/5 vacuous. With gates cleared, the **unchanged frozen geometry** is
  byte-identical to v1 (0 collisions, 0 near pairs, min cross-task distance
  **0.207**, 0% fidelity-passing), so the certificate now defers on **geometric
  sparsity alone** rather than label vacuity. The two v1 deferral causes are
  thus separated and one closed; the unambiguous next certificate move is to
  **expand the registered context universe** (more Phase 0 tasks, requiring a
  Phase 0 register amendment), not another oracle iteration. The alternative is
  to proceed to a Branch E solver on capability grounds (the certificate found
  no collision to block it and cannot certify locality to motivate a smooth
  selector). Either path needs its own pre-registered spec + arena gate +
  verdict discipline.
- [`PHASE3_5_REFLECTION.md`](PHASE3_5_REFLECTION.md) -- reflection doc
  naming the three-receipt convergence under the
  deterministic-low-capacity-learner family as a methodological finding,
  scoping what kind of learner would produce a non-trivial sufficiency
  verdict, and surveying four Phase 4 branch options. Admits no
  execution and changes no verdict. The
  Pass A representation/decoder contract, Pass B split/floor/discrimination
  contract, Pass C learner/metric contract, and Pass D receipt/command contract
  are filed. The freeze-marker runner wiring is implemented; the first receipt
  is admitted only under the frozen Pass A-D contracts.

## Discipline Tooling

The following enforces public-evaluation and Kaggle discipline rather than
just documenting it:

- `npm run arc:phase0:leak-check` -- audits inventory manifest, register
  splits, predictions ⊆ register, no Kaggle scaffolding, no non-inventory ARC
  script with `evaluation` literals. Exits nonzero on any FAIL.
- `.githooks/pre-commit` -- runs the leak check before every commit. Installed
  automatically by the `prepare` npm script (sets `core.hooksPath=.githooks`).
- `.github/workflows/arc-discipline.yml` -- runs the leak check on every push
  or PR that touches `docs/prereg/arc/**`, `scripts/arc-*`,
  `tests/arc-baselines/**`, or related infrastructure.
- `scripts/arc-phase0-inventory.mjs` double-flag override -- emitting
  evaluation test outputs requires `--include-evaluation-test-output` *and*
  `--authorize-evaluation-leak`, and `--out` must end in `_PRIVILEGED_AUDIT`.
  Single-flag-by-mistake leaks are blocked at runtime.

## Append-Only Rule

Once a phase spec is used to admit or block a run, the body above its
amendments line is frozen. Corrections or refinements must be appended with:

- date and timezone;
- author;
- one-line justification;
- explicit statement of whether the prior verdict changes.

## Public-Language Constraint

Phase 3 has **seven** full-grid-control binding receipts on file (V1, V2,
compact-7, Phase 3A `per_task_coord_mlp_v1`, Phase 3D
`structured_edit_residual_v1`, Phase 3D color-rule variant
`structured_edit_color_rule_v2`, and Phase 3D mask-target variant
`structured_edit_mask_target_v3`), all floored at zero exact matches; no
sufficiency support adjudication is on file. Public copy may say:

- ARC-AGI abstraction coupling roadmap;
- registered task-subset audit;
- shadow-projection hypothesis for static grid reasoning;
- falsifiable sufficiency test, filed; the deterministic-low-capacity
  learner branch produced three task-hardness verdicts, the
  `blackwell_task_decoder_v1` lane produced a Branch C bounded-failure
  receipt with the raw-grid exact-floor caveat, the strengthened
  `blackwell_publictrain_rawgrid_gate_v2` lane produced a
  `full_grid_control_floor` receipt on all four held-out lanes, the
  Branch B compact-subset diagnostic lane (compact-7) produced a
  `compact_full_grid_control_floor` receipt with a qualitatively
  distinct failure -- dominant-color mode collapse -- on every
  held-out instance, the Branch A stochastic per-task lane
  (`per_task_coord_mlp_v1`) produced a `branch_a_full_grid_floor`
  receipt with a third qualitatively distinct failure --
  conditioning starvation + shape-generalisation failure, the Branch D
  structured-edit-residual lane (`structured_edit_residual_v1`)
  produced a `branch_d_full_grid_edit_floor` receipt with a fourth
  qualitatively distinct failure -- edit-color-rule failure, and the
  bottleneck-targeted Branch D variant
  (`structured_edit_color_rule_v2`) replaced the failed color MLP
  with a deterministic conditioning-derived color-rule bank, returning
  `branch_d_color_rule_full_grid_floor`. The variant achieved its
  design goal at the diagnostic level -- palette recovery roughly
  doubled vs the base (pttest 0.67-0.83 vs 0.33, test_lodo 0.63 vs
  0.32), the `edit_color_failure` quarantine label was eliminated,
  and the failure decomposition shifted to a quantitative
  3-component bottleneck (41% `edit_mask_failure`, 16%
  `color_rule_selection_failure`, 9% `color_rule_bank_coverage_failure`),
  and the second Branch D variant (`structured_edit_mask_target_v3`)
  then replaced the edit-mask predictor with a 13-family deterministic
  mask-candidate bank, returning `branch_d_mask_target_full_grid_floor`
  -- the seventh floor -- and a clean negative result: the mask repair
  did NOT shift the bottleneck off the mask (mask-stage labels still
  137/196; the learned mask MLP still won 13.8% of selections;
  `test_lodo` mask F1 fell). All seven full-grid controls now agree on
  the floor across two task distributions, three learner families
  (transformer + per-task MLP + per-task structured-edit), two output
  framings (whole-grid + structured edit residual), and -- within the
  structured-edit framing -- both the learned and deterministic-bank
  variants of each edit component (mask and color). With both named
  bottlenecks of the structured-edit framing probed by deterministic
  banks and the floor holding at both, **Branch E (a different framing
  entirely -- e.g., generative program / DSL search, or test-time
  prompting of a frozen large LM) is the live frontier**; a
  selection-targeted within-framing variant (a joint mask+color
  selector, since both banks fail more by selection than coverage)
  remains the smaller alternative.

Phase 3E reframed the question as a decoder-independent certificate:
does the frozen `signature_palette` context representation itself have
exact or near fiber collisions with incompatible required behavior on
the registered contexts? The binding certificate (pinned to
freeze-marker `C133FF70`, 25 primary contexts / 49 total, ~4 min,
deterministic) returned **`phase3e_deferred_label_vacuity`** with a
clean dual finding: (1) **zero** exact or representation-level
collisions, and more strongly **zero** cross-task context pairs within
any tested radius -- the minimum cross-task
`d_context_signature_palette` is **0.207**, 4.1x the primary threshold
(`epsilon_primary=0.05`) and 2.1x the loose threshold -- so
`signature_palette` *maximally separates* distinct registered tasks
and there is no near-fiber locality to certify either way; and (2) the
`program_sketch_v1` oracle is **prior-blind** (68% of primary contexts
vacuous), cleanly structured: the Branch D edit-composition banks
characterize `color_role` and `objectness` (0/4 vacuous each) but are
blind to `counting`, `local_completion`, `spatial_transform` (4/4
each) and `symmetry` (5/5) -- exactly the priors whose output is not a
local edit of an input-derived baseline. The certificate therefore
**neither blocks nor licenses Branch E**: it found no registered
collision (so the seven floors are not explained by a signature
collision, though this does not prove sufficiency), but it cannot
issue `fiber_locality_positive` (0% fidelity-passing neighborhoods,
vacuous oracle). To get an adjudicable certificate, a future amendment
needs a framing-agnostic program-sketch oracle (covering the four
edit-blind priors) or a larger registered context universe; otherwise
Branch E must be justified by capability rather than by certified
fiber geometry.

The v2 oracle repair
([`PHASE3E_PROGRAM_SKETCH_ORACLE_V2_SPEC.md`](PHASE3E_PROGRAM_SKETCH_ORACLE_V2_SPEC.md))
is a certificate labeler, not a solver: it cannot emit candidate grids,
choose outputs, train a selector, or retune the frozen Phase 3E geometry after
seeing labels. Its binding receipt (pinned to freeze-marker `3B1B76D2`,
runnerSha256 `6163C208`, ~8 s deterministic, 25 primary contexts) returned
**`phase3e_v2_deferred_sparse_fibers`** and **repairs + certifies** the v1
oracle defect: the framing-agnostic nine-facet `program_sketch_v2` labeler is
**0.00 vacuous** (v1 was 0.68), and all four priors that were 100% vacuous under
v1 (`counting`, `local_completion`, `spatial_transform`, `symmetry`) are now
0/4-0/5 vacuous, with 6-7 of 9 facets populated per primary context. All three
spec gates pass: anti-vacuity (0.00 <= 0.20), anti-prior-laundering (0%
violations, min extra facets 5 >= 2), and anti-solver-leakage
(`unique_core_sketch_fraction 0.36` <= 0.60, `core_sketch_exact_lookup_fraction
0.04` <= 0.20 -- the sketch is neither a per-context fingerprint nor an
input->output lookup table). With the labeler defect closed, the **frozen
geometry is byte-identical to v1** (0 collisions, 0 near pairs, min cross-task
`d_context_signature_palette` **0.207**, 0% fidelity-passing), so the two v1
deferral causes are now **separated and one closed**: label vacuity is repaired,
and **geometric sparsity is the sole remaining deferral cause**. The certificate
still **neither blocks nor licenses Branch E**, but the next certificate move is
now unambiguous -- **expand the registered context universe** (a Phase 0 register
amendment), not another oracle iteration.

Avoid:

- "Sundog solves ARC";
- "human-level abstraction";
- "the 5D subspace is universal";
- any claim that a Kaggle entry validates the theory without a
  non-trivial Phase 3 sufficiency receipt;
- any Branch A, Branch B, or Branch D support claim from any of
  the seven filed binding receipts (V1, V2, compact-7, Phase 3A,
  Phase 3D base, Phase 3D color-rule variant, Phase 3D mask-target
  variant);
- any claim that Phase 3E found a signature-fiber collision (it did
  not -- 0 collisions), ruled signature_palette insufficient, proved
  it sufficient, or licensed a Branch E solver -- the v1 binding
  certificate is `phase3e_deferred_label_vacuity` and the v2 binding
  certificate is `phase3e_v2_deferred_sparse_fibers`; both adjudicate
  neither sufficiency nor a locality-positive Branch E license;
- any claim that the maximal cross-task separation (min distance
  0.207) proves signature sufficiency -- it only shows the
  representation does not conflate distinct registered tasks at the
  frozen radius, which the spec explicitly forbids reading as
  sufficiency;
- any claim that the v2 oracle repair (`program_sketch_v2`,
  `phase3e_v2_deferred_sparse_fibers`) repaired Phase 3E itself, licensed
  Branch E, ruled signature_palette (in)sufficient, or found a collision
  -- it repaired only the *labeler* (vacuity 0.68 -> 0.00, all four
  edit-blind priors now 0-vacuous, all three gates pass); the frozen
  geometry is byte-identical to v1 (still 0 collisions, still min distance
  0.207, still 0% fidelity-passing), so the certificate is still
  **deferred**, now on geometric sparsity alone, and still adjudicates
  neither sufficiency nor a Branch E license;
- any claim that the v2 gates passing (anti-vacuity, anti-prior-laundering,
  anti-solver-leakage) is evidence about ARC sufficiency -- the gates only
  certify that the oracle is a faithful, non-laundering, non-leaking
  *labeler*, which is a precondition for an adjudicable certificate, not a
  result about the representation;
- any claim that the 108-task context expansion
  (`phase0_fiber_expansion_admit` / `phase3e_v2_expanded_oracle_regression`)
  proved or disproved signature sufficiency, licensed or blocked Branch E,
  certified fiber locality, or found a collision -- the expanded certificate
  could not adjudicate fibers because the v2 oracle's anti-solver-leakage gate
  **fails** on the larger register (`core_sketch_exact_lookup_fraction 0.35 >
  0.20`); separately the fibers stayed sparse (min cross-task distance only
  `0.207 -> 0.196`). It is an oracle-regression + still-sparse finding, nothing
  more;
- any claim that the oracle regression means the v2 oracle was "wrong" or that
  the gate should be loosened -- the gate fired exactly as pre-registered;
  loosening `leakage_exact_lookup_max` after seeing the expanded value would be
  a post-hoc p-hack. A register-size-robust metric requires a new pre-registered
  oracle spec, not a retune;
- any claim that the relative-locality NEGATIVE verdict
  (`phase3e_relative_locality_negative`) changes the absolute-epsilon verdicts,
  proves or disproves sufficiency, licenses or blocks Branch E, or proves
  `raw_grid` is sufficient -- it found no USABLE rank-local sketch geometry for
  `signature_palette` (weak signal, metadata-tied, below raw_grid, 28.5%
  hard-incompatible > the 25% gate); raw_grid merely edges palette on this one
  metric (anti-privilege signal, not a sufficiency claim);
- any claim that palette being statistically above prior-stratified random
  (p=0.0002) is a positive result -- the positive branch additionally requires
  Δ-vs-metadata >= 0.10, Δ-vs-uniform >= 0.15, and hard-incompat <= 0.10, none of
  which hold; a small significant-but-unusable signal is the negative reading;
- any claim that the Branch E `branch_e_capability_demonstrated` verdict means
  "Sundog solves ARC", proves Blackwell `signature_palette` sufficiency, validates
  the 5D subspace, or says anything about the public-evaluation split or a Kaggle
  entry -- it is a **modest capability** result on held-out **public-training**
  lanes (2 distinct tasks per gated lane, ~3% instance rate, structurally simple
  tasks) from a deterministic train-consistency program search; it does not change
  the four Phase 3E certificate verdicts or the seven full-grid-control floors,
  and Phase 6 still gates all eval/Kaggle work;
- any claim that Branch E's positive result reverses the seven floors or the
  certificate program -- those characterized *learning/representation* families
  and signature fiber geometry; Branch E is an independent *program-search
  capability* probe that selects by train-pair consistency (not signature
  geometry) and rests on the certificates' no-collision / no-locality boundary;
  searching a fixed library clearing a floor that learning-from-scratch floored on
  is the expected come-apart, not a contradiction;
- any claim that Branch E v2 (`branch_e_v2_capability_replicated`) lifted the
  capability, or that its validation drop (2->1/lane) means a regression toward the
  floor -- v2 retained both v1 gated solves on both lanes (it did not floor) but
  solved 0 new tasks, so it *replicated, did not lift*; the validation drop is a
  top-2-crowding artifact of the larger search space, not a capability collapse.
  The honest reading is that deterministic program-search capability is bounded
  near the v1 ~3% baseline; do not read v2 as either a lift or a regression;
- any "color-rule variant flipped the bottleneck -> mask is the new
  limit" claim that omits the still-floor verdict -- the color-rule
  variant achieved a diagnostic improvement (palette doubled,
  decomposition gained) but the arena did not open;
- "the mask bank beats the learned mask" -- it does not; the legacy
  mask MLP still won 13.8% of selections and `test_lodo` mask F1 fell;
- "the mask-target variant shifted the bottleneck off the mask" -- the
  opposite; mask-stage labels still dominate (137/196) and the floor
  held;
- describing any Phase 3 binding receipt as a signature-specific
  falsification independent of decoder capacity;
- describing the seven filed Phase 3 receipts as a sufficiency-failure
  conclusion -- per the reflection, they characterise the
  deterministic-low-capacity, per-task-scratch, and structured-edit
  framing families on the registered task class, not the
  shadow-projection representation;
- claiming "V2 failed -> signature representation is favoured" -- V2
  floored the same way V1 did; no signature-vs-full-grid comparison
  is licensed by either receipt;
- claiming "compact-7 failed -> compact tasks are unsolvable" -- the
  failure is of this learner family on this slice; a different
  framing (Branch D) might or might not pass on the same slice, and
  that is the next question, not a settled one;
- claiming "Phase 3A floored -> signature representation is favoured"
  -- the Phase 3A raw-grid arm did not open the arena, so no per-arm
  comparison is licensed; the all-arms-equal pattern in the receipt
  is a consequence of the floor, not evidence of per-arm equivalence;
- claiming "Phase 3D floored -> signature representation is favoured"
  -- same caveat: the raw_grid_edit arm did not open the
  non-baseline arena; `signature_palette_edit`'s slightly higher
  pixel / minority recall in the documentation-only per-arm table
  is unlicensed-for-adjudication data, not a support signal;
- claiming any single Phase 3 receipt closes the sufficiency
  question -- the seven floors together close PHASE3_5_REFLECTION
  Branches A, B, C, and D in the filed learner families and
  framings, but the sufficiency question remains open under the
  filed Phase 3E signature-fiber certificate, a later Branch E solver,
  or a smaller selection-targeted Branch D variant;
- any spatial_transform or local_completion claim from the compact-7
  receipt -- those priors are not represented in the compact-signal
  slice;
- claiming "Phase 3D's baseline-exact = 0 means the input copy
  baseline doesn't work" -- the baseline picker correctly identifies
  same_as_input + identity_top_left as the dominant pick (28/49 in
  smoke) but the target outputs of registered tasks genuinely differ
  from the baselines after the editing operation; `baseline_exact = 0`
  means "no task in the registered subset is solved by baseline
  alone", which is the expected design property of ARC tasks.