# Structural Failure Coincidence — Cut 3 Rendered-Signal Run Spec Pre-registration: [`README.md`](README.md) Parent run spec: [`P2_RUN_SPEC.md`](P2_RUN_SPEC.md) Cut-2 disposition: [`P2_CUT2_WAVE42_DISPOSITION.md`](P2_CUT2_WAVE42_DISPOSITION.md) Admission check: [`P2_CUT3_ADMISSION.md`](P2_CUT3_ADMISSION.md) Phase-15 hazard source: [`SPECULATIVE_HALO_PROOFS.md`](../../calibration/SPECULATIVE_HALO_PROOFS.md) Controller binding: [`P2_CUT2_C1_CONTROLLER_BINDING.md`](P2_CUT2_C1_CONTROLLER_BINDING.md) Filed: **2026-05-16 (PT)**. Status: **CUT-3 SPEC FILED — EXECUTION HELD**. This document opens the Cut-3 documentation front as an artifact-before-agent record. It runs nothing, trains nothing, and admits no controller by itself. ## Purpose Cut 2 closed the closed-form line honestly: C3-A-R removed the Cut-1 vacuity, but C3-A-T/B blocked because the closed-form decoy cannot compete with the promoted parhelion route where that route is eligible. That is the gamma regime-separability finding. Cut 3 is the pre-registered rendered-signal escalation: test whether a learned or image-native correlate can compete in the eligible band where scalar closed-form decoys structurally could not. Cut 3 is not a consolation run. It is a different signal class with a known hard dependency from Phase 15: the pixel-to-degree / centering hazard. This spec resolves that hazard at the protocol level by making angular calibration an admission requirement. If the render corpus cannot satisfy that requirement, Cut 3 is blocked. ## H0 — angular-calibration gate (the Phase-15 hazard) Every rendered frame admitted to Cut 3 must have a per-frame angular calibration record that maps image pixels to sun-centered angular coordinates before the controller or any learned baseline sees the image. The calibration record must include: 1. `sun_px`: the image-space sun center or HaloSim-native sun-origin receipt. 2. `projection`: the projection family used by the frame. 3. `theta_map`: a deterministic function or tabulated ruler that maps pixel location to angular distance from the sun, in degrees. 4. `valid_angular_span_deg`: the angular span over which the map is calibrated. 5. `anchor_check`: a cross-check against at least one known halo locus inside the valid span, preferably the 22 degree halo; if a scored feature lies beyond the span, the frame is inadmissible for that feature. 6. `source_hashes`: hashes for the render and its calibration artifact. Allowed ways to satisfy H0: - **HaloSim-native Scale, full-span.** A scale-stamped render whose ruler covers every scored feature and at least one anchor locus. This is the preferred HaloSim path because the ruler is drawn from the sun and carries HaloSim's own projection. - **Renderer-native angular metadata.** A renderer or export mode that directly emits a sun-centered angular coordinate map. The coordinate map must be saved as an artifact and hashed. - **Independently validated 2D angular fit.** A non-split render with at least two known loci inside the fitted range, such as 22 and 46 degrees, plus residuals below the tolerance below. Disallowed as H0: - A global pixels-per-degree constant copied between renders. - Any scale-lock that changes with HaloSim auto-zoom but is not re-measured per frame. - A ruler or fitted map whose span is shorter than the scored feature field. - A center chosen by whichever downstream detector gives the desired result. - Any frame where the filename, directory, overlay text, or embedded metadata gives the hidden altitude to the agent. H0 tolerance: anchor residuals must be **<= 0.5 degree** for every anchor locus used to admit a frame. A feature outside the valid angular span is not scored; if too many features are outside span, Cut 3 remains blocked for that corpus. Phase-15 precedent: the HaloSim Scale method is accepted as a real instrument, but the pyramidal scale-stamped frames are a negative H0 example for Cut 3 because the ruler span was shorter than the ring field and the 22/46 degree anchors were beyond the calibrated tip. The lesson is not "Scale always passes"; it is "a scale receipt must cover the feature field or the run blocks." ## Signal and corpus Cut 3 input is a rendered image, not the closed-form feature bundle. The hidden cause remains sun altitude `h`. The agent must not receive `h`, altitude-coded filenames, render-order cues, or calibration sidecar values as inputs. Minimum corpus before execution: - `h` grid: at least **0 to 40 degrees** with step no coarser than **2 degrees**, with denser samples within **+/- 3 degrees** of the L1, L2, and L3 boundaries when feasible. - Split: train / validation / held-out test by altitude and nuisance condition, with no identical render seed appearing in both train and test. - Nuisance axes: at least two render/style axes that can be varied independently of `h` for counterfactual tests. Examples: ray seed, crop, color mode, contrast curve, non-scored crystal population, or Monte Carlo noise level. - Calibration manifest: one H0 record per frame. The first admissible corpus should prefer ordinary rendered core handles over pyramidal proof frames. Pyramidal Phase 15 remains a P2/P3 proof chain with its own failure modes; Cut 3 needs an image-domain Proxy-Collapse test, not a pyramidal residual table. ## Agent and baselines The agent-under-test path must be named in Cut-3 admission before any run. Two families are allowed, but must not be mixed silently: - **Existing-controller rendered adapter.** The bound `PhotometricAgent` receives a scalar photometric objective derived from the rendered image. The adapter must be fixed before execution. - **Learned image agent.** A learned policy or regressor receives the rendered image. Architecture, optimizer, training budget, and early-stop rule must be frozen before training. Required baselines: - **Route baseline.** A transparent route extractor that uses only admitted angular handles from the H0-calibrated image, with the same abstain rules as BOUNDARY_MAP. - **Correlate baseline.** A deliberately image-correlate baseline that is allowed to use nuisance/style features but not the angular route. It must be strong enough to move under nuisance edits; otherwise the decoy battery is inconclusive. - **Analytic reference.** The closed-form parhelion route on the calibration sidecar, used only for scoring and sanity checks, never as an agent input. ## Quantities The P2 four-quantity structure carries forward. Quantities 1-3 are traceability; quantity 4 is efficiency/reporting only. 1. **Rendered convergence.** On H0-admitted and BOUNDARY_MAP-eligible frames, the agent's altitude estimate must be within **1.5 degrees** of hidden `h` on at least **90%** of held-out test frames. 2. **Counterfactual steerability.** Two edits are required: - *Route edit:* alter the rendered angular handle while holding nuisance/style cues fixed. Pass iff the agent follows the edited route within **2.0 degrees**. - *Nuisance edit:* alter style/metadata/image nuisance while holding the angular route fixed. Pass iff the agent moves by **<=0.5 degree**, while the correlate baseline moves by at least **2.0 degrees**. If the correlate baseline does not move, the edit is too weak and the result is inconclusive, not pass. 3. **Failure-boundary coincidence.** The agent must degrade, abstain, or switch at the BOUNDARY_MAP loci: - L1 low leverage / invalid parhelion route. - L2 CZA disappearance at the operative 32 degree coded guard. - L3 tangent-to-circumscribed merge at 29 degrees. - L4 supralateral remains a non-handle. - L5 rendered does not equal anchored; unanchored rendered primitives may act as nuisance cues but never as route evidence. 4. **Efficiency and robustness.** Report samples, wall-clock, render count, and failure modes. No pass/fail threshold. ## Outcome mapping The P2 outcome table still governs: - Fails convergence: convergence null. - Converges but follows nuisance edits or ignores route edits: Proxy-Collapse / opaque-correlate confirmation. - Converges and steers but sails through BOUNDARY_MAP boundaries: boundary failure; no traceability claim. - Passes convergence, steerability, and boundary coincidence: narrow apparatus result only, never theorem language. Cut 3 must also report an H0 disposition: - **H0 PASS:** corpus is angularly calibrated and may be used for Cut-3 admission. - **H0 BLOCK:** corpus lacks valid angular calibration; no controller or learned-agent run may start. - **H0 PARTIAL:** some frames/features are admitted and others excluded; execution may start only if coverage remains sufficient after exclusions. ## Publication and language guard Until an admitted Cut-3 run passes quantities 1-3, public surfaces may say only: > Cut 3 is the rendered-signal escalation and is gated on per-frame > angular calibration. They may not say Cut 3 has begun, that a rendered traceability harness passes, or that the controller is traceable. If H0 blocks, the block is the result. ## Execution discipline No Cut-3 script, corpus generator, controller training run, or long HaloSim sweep may run until [`P2_CUT3_ADMISSION.md`](P2_CUT3_ADMISSION.md) records ADMIT or a narrower PARTIAL ADMIT. Any run expected to exceed the repository's ~10-minute rule must be staged as exact PowerShell commands with wall-clock estimates and read-back paths. ## Audit Notes *(reviewer space — append-only below)* **2026-05-16 (PT) — maintainer.** H0 operationalized (append-only; consistent with the frozen §"H0" protocol, no value changed): [`P2_CUT3_H0_CALIBRATION.md`](P2_CUT3_H0_CALIBRATION.md). Turns the frozen H0 protocol into a concrete per-(frame,feature) record schema, a keystone span-measured-before-feature invariant (the Phase-15 failure mechanism made the guarded rule; coverage shortfall ⇒ Cut-3 BLOCKED, not forced), and a two-sided self-test whose immutable known-FAIL fixture is the real Phase-15 pyramidal `pyr_w*_scale.png` set. Runnable checker + per-frame records + residual table + known-PASS fixture remain the maintainer's pre-run fill; execution stays HELD; nothing run.