# Phase 0 -- ARC-AGI Task Subset Specification Roadmap: [`../../SUNDOG_V_ARC.md`](../../SUNDOG_V_ARC.md) Filed: **2026-05-28 (PT)** Status: **OPEN -- NO OPERATOR DESIGN ADMITTED**. This is the first ARC-AGI preregistration artifact. Its job is to make the task subset, exclusion criteria, evaluation handling, and baseline slate crisp before any Sundog-specific shadow operator is designed. ## Claim Under Test The Phase 0 claim is not that Sundog can solve ARC. The claim is narrower: > There exists a registered, non-trivial ARC-AGI-2 task subset whose rules are > naturally describable by core-knowledge priors -- objectness, counting, > symmetry, spatial transformation, local completion, and color-role mapping -- > and whose boundaries can test whether a shadow-projected signature is > sufficient without full grid reconstruction. If this subset cannot be registered without leaking evaluation answers or hand-picking post-hoc successes, the ARC coupling is blocked before operator design. ## Data Source and Split Policy Authoritative data source: the public [arcprize/ARC-AGI-2](https://github.com/arcprize/ARC-AGI-2) repository. The local working copy should live outside this repo unless the maintainer explicitly decides to vendor a small fixture. Suggested path: `$env:USERPROFILE\Datasets\ARC-AGI-2`. Split handling: | split | allowed use in Phase 0 | restrictions | | --- | --- | --- | | public training | Inventory, manual inspection, taxonomy, development subset, cheap baselines. | Allowed to inspect; still record every inclusion/exclusion decision. | | public evaluation | Locked adjudication subset only. | No manual grid inspection while designing the operator. Any selection must be by preregistered metadata/hash rule, and scripts must not emit test outputs. | | Kaggle private / semi-private | Phase 6 submission only. | No access in Phase 0. Submission notebook must obey Kaggle no-internet evaluation. | ## Inventory Command The Phase 0 inventory is mechanical and should complete in seconds: ```powershell node scripts/arc-phase0-inventory.mjs --data-dir "$env:USERPROFILE\Datasets\ARC-AGI-2\data" --out results/arc/phase0-inventory ``` The script emits metadata only: task IDs, split counts, pair counts, dimensions, palette sizes, simple shape-change flags, density ranges, connected-component counts, and coarse prior hints. It must not emit public-evaluation test output grids unless explicitly run with the evaluation-output override for a final, post-freeze audit. ## Inclusion Criteria A task may enter the registered development subset if all are true: 1. It is a valid ARC-AGI-2 JSON task with at least two training pairs and at least one test input. 2. Its apparent rule can be assigned to at least one core-knowledge prior: objectness, counting, symmetry, spatial transformation, local completion, or color-role mapping. 3. The expected transformation can be described without task-specific language labels, external world knowledge, or memorized lookup tables. 4. The output grid can be evaluated by exact match, and any auxiliary metric is recorded as diagnostic only. 5. The task has an explicit predicted failure boundary for the future signature audit: non-local information, capacity pressure, gauge-breaking ambiguity, or full-state-only dependency. Target size: **36 public-training tasks**, stratified across the six priors above. Minimum viable size: **18 tasks**, with at least three priors represented by four or more tasks each. If the minimum cannot be met from public training without relaxing the criteria, Phase 0 fails and the ARC lane retreats to synthetic grid warmups. ## Exclusion Criteria Exclude tasks from the Phase 0 development subset if any are true: 1. The rule appears to require long sequential program search rather than a compact abstraction. 2. The answer depends on a global readout that no local signature could plausibly retain. 3. The output information content is so high that exact reconstruction would be the real task. 4. The task is text-like, code-like, table-lookup-like, or otherwise outside ARC core-knowledge priors. 5. The task becomes classifiable only after inspecting multiple candidate solver failures. 6. Public-evaluation task contents were manually inspected before selection. Excluded tasks are not discarded silently. Each excluded public-training task that was manually considered gets a row in the task register with the exclusion reason. ## Task Register Schema The Phase 0 task register lives at: `docs/prereg/arc/P0_TASK_REGISTER.csv`. Required columns: | column | meaning | | --- | --- | | `task_id` | ARC task filename stem. | | `split` | `training` or `evaluation-blind`. | | `status` | `include`, `exclude`, or `hold`. | | `primary_prior` | One of the six registered priors. | | `secondary_priors` | Semicolon-separated optional priors. | | `inclusion_basis` | Short human-readable rule description for included tasks. | | `exclusion_reason` | Required when `status=exclude`. | | `predicted_boundary` | Non-local, capacity, gauge-breaking, full-state-only, or other named boundary. | | `inventory_row_hash` | Hash of the metadata row used during selection. | | `manual_inspection` | `yes` for public-training inspection; `no` for evaluation-blind. | | `notes` | Free text; no solver results. | ## Baseline Slate Phase 0 must record baseline numbers before Phase 1 begins. The first slate is: 1. **Random valid grid baseline.** Uses observed training output shapes and colors; reports exact-match and pixel-correctness diagnostics. 2. **Identity/copy baseline.** Emits the test input, plus a same-shape color remapping attempt when a one-to-one training color map is obvious. 3. **DSL-lite brute baseline.** Enumerates a frozen small set of transforms: rotate, reflect, translate/crop/pad, recolor, connected-component copy, fill, and compose up to a preregistered depth. 4. **Tiny learned reference.** Optional if it completes under the repo's ten-minute rule; otherwise stage the exact PowerShell command and estimate instead of running it. Subset quality guard: if the DSL-lite brute baseline solves more than **70%** of the development subset, the subset is too easy for the sufficiency question and must be rebalanced before Phase 1. If all cheap baselines solve **0%**, the subset may be too hard or misclassified; file a hold review before operator design. ## Exit Criteria Phase 0 exits only when all artifacts below exist: 1. `results/arc/phase0-inventory/manifest.json` 2. `results/arc/phase0-inventory/tasks.csv` 3. `docs/prereg/arc/P0_TASK_REGISTER.csv` 4. `docs/prereg/arc/P0_BASELINES.md` 5. An admission note appended to this file stating **ADMIT**, **PARTIAL ADMIT**, or **HOLD** for Phase 1. Until then, the following are forbidden: - implementing the ARC shadow-projection operator; - training a signature decoder; - inspecting public-evaluation task grids manually; - tuning on public-evaluation scores; - preparing a Kaggle submission notebook. ## Outcome Branching | outcome | interpretation | next branch | | --- | --- | --- | | Cannot meet minimum subset size | The ARC-AGI-2 public-training subset does not cleanly expose the claimed core-knowledge surface under these criteria. | Retreat to synthetic grid warmups; no Phase 1 ARC operator. | | Subset exists but cheap baselines solve >70% | The selected surface is too easy / too DSL-native. | Rebalance before operator design. | | Subset exists but cheap baselines solve 0% | The selected surface may be too hard or misclassified. | HOLD; audit taxonomy before operator design. | | Subset, blind policy, and baselines are filed | Phase 1 can define the discrete-grid shadow domain. | ADMIT Phase 1. | --- ## Amendments Append-only. Each amendment must carry a timestamp, author, justification, and verdict impact. **2026-05-28 (PT) -- Codex.** Phase 0 inventory/register/baseline receipt filed at [`P0_BASELINES.md`](P0_BASELINES.md). Inventory result: `1120` public tasks (`1000` training, `120` evaluation), with public-evaluation test-output metadata omitted. Register result: `36` included public-training tasks, balanced `6` per registered prior, no duplicate task IDs, all `inventory_row_hash` values resolving into the inventory CSV. Baseline result: `random_valid`, `identity_copy`, and `dsl_lite_v0` all scored `0/36` exact; diagnostic mean pixel accuracy was `0.1530`, `0.5585`, and `0.5585` respectively. Verdict impact: **PARTIAL ADMIT -- subset clean, operator design HOLD**. The subset clears the >70% too-easy guard but fires the zero-floor caveat, so the next admitted work is a Phase 0 hold review / taxonomy rebalance or synthetic-grid warmup only. No ARC shadow-projection operator, signature decoder, or Sundog-specific feature scoring is admitted until a later amendment records **ADMIT**. **2026-05-28 (PT) -- Claude (Opus 4.7).** Phase 0 baseline expansion + verdict upgrade. Following the prior PARTIAL ADMIT amendment's preregistered "Allowed next work" path ("Improve or add non-Sundog cheap baselines only by append-only amendment before looking at any Sundog operator result"), three baselines were added to `scripts/arc-phase0-baselines.mjs`, keeping `random_valid`, `identity_copy`, and `dsl_lite_v0` byte-frozen: - `dsl_lite_v1`: `tile`, `translate`, `palette_permute` (depth 1). - `dsl_lite_v2`: `pad`, `fill_enclosed`, `component_copy_largest`, plus depth-2 composition over the union of v0, v1, and v2 structural transforms with one final color-map fit per composed candidate. - `tiny_learned_v0`: per-task nearest-neighbour over train pairs by padded pixel Hamming distance. Combined primitive coverage now exhausts the spec Baseline Slate item 3 DSL-lite primitive list (rotate, reflect, translate/crop/pad, recolor (palette permute), connected-component copy, fill) at composition depths 1 and 2. Optional tiny learned reference is included per Baseline Slate item 4. Result on the registered 36-task subset: all five non-random baselines score `0/36` exact (mean pixel accuracy `0.5585` for v0/v1/v2/identity fallback; `0.3993` for tiny_learned). Implementation correctness verified by synthetic sanity check at `tests/arc-baselines/out/summary.csv` (v1 solves a 2x2 tile task; v2 solves both a 2x2 tile and a pad-with-zero border task), confirming the registered-subset zero floor is a property of the subset, not a fitter bug. Verdict impact: **ADMIT** (Phase 1 -- ARC grid representation as shadow domain -- admitted). The preregistered "Subset, blind policy, and baselines are filed" outcome row from the Outcome Branching table now governs, with the zero-floor caveat resolved by exhaustive baseline coverage rather than by a nonzero solve. Phase 1 inherits `0/36` exact as the hard preregistered floor that any Sundog-specific result must clear. Public-evaluation discipline (no manual grid inspection, no Kaggle prep) remains in force until Phase 6. **2026-05-28 (PT) -- Claude (Opus 4.7).** Discipline hardening (no verdict change). Following the ADMIT amendment above, the operational tooling that enforces the public-evaluation and Kaggle restrictions was hardened so that the discipline rules are checked by code, not only documented: - `scripts/arc-phase0-inventory.mjs` now requires *two* flags to emit evaluation test outputs: `--include-evaluation-test-output` AND `--authorize-evaluation-leak`. The `--out` path must end in `_PRIVILEGED_AUDIT`; the manifest emits `privilegedAudit: true` and a banner `evaluationPolicy`. This formalises the "evaluation-output override for a final post-freeze audit" provision in the Data Source and Split Policy section above into a runtime gate. - `scripts/arc-phase0-leak-check.mjs` (new) audits five invariants on every invocation: default inventory manifest is non-privileged; register has no `split=evaluation` rows and every `split=evaluation-blind` row carries `manual_inspection=no`; baseline predictions are a subset of registered training task IDs; no Kaggle scaffolding (kaggle.json, non-empty .ipynb) exists under `scripts/`, `docs/`, `tests/`, `notebooks/`; non-allowlisted ARC scripts (anything other than the inventory and leak-check themselves) contain no `evaluation` literal. - `.githooks/pre-commit` runs the leak check before every commit, installed by `scripts/setup-githooks.mjs` via the npm `prepare` script. - `.github/workflows/arc-discipline.yml` runs the leak check on push and PR for changes touching ARC paths. - [`EVAL_BLIND_SELECTION.md`](EVAL_BLIND_SELECTION.md) is filed as a stub pattern for the first Phase 1+ amendment that adds evaluation-blind register rows, so the rule shape is preregistered before any concrete selection. Verdict impact: none. The ADMIT verdict above continues to govern. This amendment is a tooling tightening, not a scope or floor change. Future amendments may add or modify these tools, but the verdict logic and preregistered constraints in the body above remain unchanged. **2026-05-29 (PT) -- Codex.** Context-expansion-for-fibers spec filed: [`PHASE0_CONTEXT_EXPANSION_FOR_FIBERS_SPEC.md`](PHASE0_CONTEXT_EXPANSION_FOR_FIBERS_SPEC.md). This is an append-only Phase 0 register-amendment path motivated by the Phase 3E v2 receipt `phase3e_v2_deferred_sparse_fibers`: the v2 oracle defect is repaired, but the 36-task registered context universe has zero cross-task neighbors within `epsilon_primary = 0.05` and minimum cross-task distance `0.207`. The new spec freezes a balanced expansion target of 108 included public-training tasks (18 per prior), mechanical pre-inspection candidate ordering, no use of Phase 3E distances/sketches/solver outputs during selection, and unchanged Phase 3E v2 oracle/geometry thresholds. Verdict impact: **no register change and no execution admission**. The original 36-task Phase 0 register remains binding. A future expansion receipt must file `P0_CONTEXT_EXPANSION_REGISTER.csv`, `P0_TASK_REGISTER_EXPANDED_FOR_FIBERS.csv`, audit artifacts, leak-check coverage, and a freeze-marker amendment before any expanded Phase 3E v2 certificate rerun is admitted. **2026-05-29 (PT) -- Jeffery Hughes Jr.** Phase 0 fiber context expansion admitted and expanded certificate filed. The expansion was executed under the spec above: - commit `f9a1dd9`: freeze marker, candidate-ordering script, `--split-mode`, inspection renderer, Amendment A; - commit `a0f1a4b`: inspection + register assembly, **`phase0_fiber_expansion_admit`**; - commit `fa079b9`: expanded Phase 3E v2 certificate, **`phase3e_v2_expanded_oracle_regression`**. The binding expanded register contains 108 public-training tasks, exactly 18 per prior (36 original + 72 new), selected from the frozen queue with 0 hard-exclusions, 0 discipline tripwires, and no cross-prior rebalancing. The expanded certificate did not densify the fibers enough to adjudicate locality: `U_primary` increased 25 -> 336, but min cross-task distance moved only 0.207 -> 0.196, with 0 near pairs and 0% fidelity at `epsilon_primary = 0.05`. The certificate branch is nevertheless `phase3e_v2_expanded_oracle_regression` because the anti-solver-leakage gate fails on the expanded register: `core_sketch_exact_lookup_fraction = 0.351 > 0.20`. Anti-vacuity and anti-prior-laundering still pass. Verdict impact: the expanded register is filed, but it does not promote or block signature sufficiency or Branch E. The 36-task Phase 3E v2 receipt and the seven full-grid floors stand unchanged.