# Structural Failure Coincidence — P2 Spec Admission Check Pre-registration: [`README.md`](README.md) (frozen 2026-05-15 PT) P2 run spec under review: [`P2_RUN_SPEC.md`](P2_RUN_SPEC.md) (frozen 2026-05-15 PT) Gated by: [`BOUNDARY_MAP.md`](BOUNDARY_MAP.md) (P0 PASS) and [`P1_ADMISSION.md`](P1_ADMISSION.md) (P1 PASS) Filed: **2026-05-15 (PT)**. Reviewer: Codex audit. Status: append-only below the **Amendments** rule; the body above it is the frozen P2-spec admission record. ## Purpose This is the short admission check required by `P2_RUN_SPEC.md` before any controller evaluation may execute. It mirrors P1, but reviews the run spec rather than the boundary map: - every bundle feature must trace to `BOUNDARY_MAP.md` / receipts; - every threshold must be frozen and have a stated role; - the transparent adapter must be fully closed-form, with no learned or post-hoc parameters; - mandatory decoys must be present and non-vacuous; - L5 must be applied as the evidence-admissibility rule inside scoring. ## Admission Checklist | requirement | ruling | notes | | --- | --- | --- | | Frozen run spec exists before controller execution | **PASS** | `P2_RUN_SPEC.md` was filed before any controller run. | | P0/P1 gates are cited and inherited | **PASS** | `BOUNDARY_MAP.md` and `P1_ADMISSION.md` are explicit gates. | | Four quantities are scored separately | **PASS** | Convergence, steerability, boundary coincidence, and efficiency are separated; efficiency is non-fatal. | | L5 is applied inside scoring, not as a fifth quantity | **PASS** | Spec correctly treats rendered ≠ anchored as evidence admissibility. | | Genuine handles trace to P0/P1 receipts | **PASS** | `f_par`, `f_cza`, and `f_tan` map to L1/L2/L3. | | Mandatory decoys are present | **PASS** | `d_sup`, `d_unanch`, and `d_style` are named. | | Outcome mapping preserves the Proxy-Collapse bridge | **PASS** | P2 fail on steerability or boundary coincidence maps to the P1/debunked vocabulary. | | Transparent adapter is fully specified and cannot leak `h` | **HOLD** | The phrase "per h-regime" could be read as allowing the hidden cause into the adapter, and the CZA/tangent consistency terms are not closed-form enough to implement without interpretation. | | Decoy-edit test is non-vacuous for this controller cut | **HOLD** | Because the transparent adapter excludes `d_*` from `J`, a decoy perturbation cannot move the existing extremum-seeking controller unless the implementation violates the spec. That is a useful adapter-integrity / void test, but not yet the sharp controller-vs-correlate discriminator the prose claims. | | Thresholds trace to receipts or declared engineering tolerances | **HOLD** | The thresholds are frozen, but their exact roles/origins are not traced in a table. They may be acceptable as pre-registered operational tolerances, but that must be stated before execution. | ## Findings ### F1 — Adapter gating must not use hidden `h` The adapter paragraph says `f_cza` / `f_tan` gate consistency terms "per `h`-regime." That is ambiguous. The scorer may use true `h` to evaluate convergence after the run, and the bundle generator may use true `h` to generate `B(h)`, but the controller adapter must not read true `h` when choosing terms. Otherwise the hidden cause has leaked into the objective. **Required pre-run amendment:** publish the adapter as an explicit algorithm whose allowed inputs are only the bundle values and the candidate hypothesis `q` (for example: `f_par`, `f_cza`, `f_tan`, `R22`, and `q`). Gating must be by observable bundle state (`f_cza == 0`, `f_tan == null`, etc.), not by true hidden `h`. ### F2 — Decoy invariance is currently an adapter-integrity test, not a full correlate test The decoys are the right design move, but under the hard transparent adapter `J` excludes `d_sup`, `d_unanch`, and `d_style` by construction. For the existing extremum-seeking controller, that means decoy-edit invariance is guaranteed unless the adapter has been implemented incorrectly. This catches a real hazard — a decoy term entering `J` makes the run void — but it does not by itself prove the controller rejected a correlate, because the controller never had access to decoys through the objective. **Required pre-run amendment:** choose and record one of these scopes: 1. **Transparent-adapter sanity cut.** P2 first-cut admits only an adapter-integrity result: decoy edits are void-guard / leakage checks, not a behavioral discriminator. Public interpretation must not call a decoy-invariance pass a Proxy-Collapse falsification. 2. **Sensitivity-controlled cut.** Add a pre-registered decoy-correlate positive control or raw-bundle controller path that can read decoys. A deliberately decoy-driven baseline must move under decoy edits while the traceable controller does not. Then the decoy-edit test becomes non-vacuous. Either choice can be honest; the current prose needs one of them before a controller run. ### F3 — Thresholds are frozen, but their provenance needs a table The spec freezes `τ1 = 1.5°`, `τ2 = 2.0°`, decoy invariance `≤0.5°`, and boundary coincidence `±1.5°`. That is good pre-registration discipline. But the run-admission gate says thresholds must trace to `BOUNDARY_MAP.md` / receipts, while the exact numeric tolerances are not receipt values in the same way the 32° and 29° loci are. **Required pre-run amendment:** add a short threshold-provenance table stating, for each threshold, whether it is a geometry boundary, a receipt-derived tolerance, or a pre-registered engineering tolerance. This does not change any threshold; it prevents a later reader from mistaking operational tolerance choices for atmospheric-optics facts. ### F4 — Matched-baseline efficiency is under-specified but non-blocking Quantity (4) is explicitly non-fatal, so this is not an admission blocker. Still, before execution the matched baseline should be named concretely: e.g. analytic inverse `q = arccos(R22 / f_par)` on L1-eligible inputs, with no access to decoys. Otherwise the reported efficiency ratio will be hard to interpret. ## Admission Verdict **HOLD — P2 is started, but P2-execute is not admitted yet.** The frozen run spec is the right artifact-before-agent move, and the main objects align with the prereg: four quantities, L5 as admissibility, and mandatory decoys. The admission gate also did its job: it caught three pre-run ambiguities that would otherwise let the evaluation look sharper than it is. Controller execution remains **blocked** until the P2 spec receives append-only amendments resolving F1–F3 and this admission check is re-run. The Public-Language Constraint remains in force everywhere, including the rail: no theorem / universal-proof / traceability-success language until quantities (1)+(2)+(3) actually pass under an admitted run. --- ## Amendments Append-only. Each amendment: timestamp (date + zone), author, one-line justification. The body above is the frozen P2-spec admission record. **2026-05-15 (PT) — re-review (maintainer).** Re-run of this admission check after the F1–F4 resolution amendment in [`P2_RUN_SPEC.md`](P2_RUN_SPEC.md). Each prior HOLD re-checked against the cited clause, not rubber-stamped: - **F1 → CLOSED.** Spec amendment A1 publishes the adapter as an explicit closed-form algorithm with input set fixed to `{f_par, f_cza, f_tan, R22, q}`; `h` excluded; CZA/tangent gating is now on observed `f_cza==0` / `f_tan==null`; hard VOID invariant if `h` is read inside the adapter. The leak path is removed and the terms are implementable without interpretation. **PASS.** - **F2 → CLOSED.** A2 takes Option 2: a pre-registered decoy-correlate positive control (raw-bundle, reads `d_*`) plus `τ_pc = 2.0°`. The decoy-edit is now the paired contrast (route invariant **and** positive control moves ≥ τ_pc), with an explicit *inconclusive* branch and a guard that, absent the positive control, a decoy-invariance pass is adapter-integrity only and **never** a Proxy-Collapse falsification. Non-vacuous. **PASS.** - **F3 → CLOSED.** A3 adds the provenance table separating immutable geometry/receipt boundaries (32°, 29°, 2%·R22, supralateral) from pre-registered engineering tolerances (τ1, τ2, decoy-invariance, coincidence window, τ_pc), with the immutability rule. No value changed. **PASS.** - **F4 → RESOLVED** (was non-blocking). A4 names the matched baseline (`q = arccos(R22/f_par)`, L1-eligible, no decoy access), distinct from the A2 positive control. **Re-admission verdict: ADMIT — P2-execute is admitted.** Conditions carried: (i) the controller run obeys the AGENTS.md "~10-minute rule" — stage as operator PowerShell if the sweep exceeds ~10 min, with the frozen thresholds/branches; (ii) the Public-Language Constraint stays in force everywhere (including the rail) until quantities (1)+(2)+(3) actually pass under this admitted run; (iii) any threshold change remains append-only, justified, and never post-results, and geometry/receipt boundaries are immutable. Justification: all three blocking findings are closed by explicit, closed-form, pre-run amendments; the gate did its job and is now satisfied. **2026-05-15 (PT) — Codex execution note.** The admitted P2 first-cut run has executed; result filed in [`P2_RESULTS.md`](P2_RESULTS.md). This does not alter the admission ruling. It records that the admitted conditions were consumed by `npm run p2:structural`, which completed under the ~10-minute rule and produced `TRACEABILITY_HARNESS_PASS` for the route controller plus `OPAQUE_CORRELATE_POSITIVE_CONTROL_CONFIRMED` for the decoy-correlate positive control. **2026-05-15 (PT) — correction / reviewer challenge accepted.** The execution note immediately above is superseded as an interpretation. The admission ruling remains valid, but the executed harness was not an adequate admitted route-use test: it reduced the route controller to the same inverse as the analytic baseline (`g^-1(g(h))`), kept decoys structurally unreachable by the route objective, did not use CZA/tangent state to estimate `q`, and hardcoded supralateral as a non-handle. Corrected execution verdict: `MACHINERY_LIVE_ROUTE_TEST_VACUOUS`. The positive-control verdict remains `OPAQUE_CORRELATE_POSITIVE_CONTROL_CONFIRMED`. Therefore P2 quantities (1)–(3) have **not** established traceability under an admitted discriminating run; the Public-Language Constraint remains in force. **2026-05-15 (PT) — C1 re-check note.** C1 is closed only by [`P2_CUT2_C1_CONTROLLER_BINDING.md`](P2_CUT2_C1_CONTROLLER_BINDING.md): Cut 2 must bind `sundog.agents.photometric.PhotometricAgent` from `agents/photometric.py`, preserving its `reset(...)` / `act(obs)` interface. This closes the "named existing controller" gap that Cut 1 violated. Cut-2-execute is still **NOT admitted**: C2 (non-invertible nuisance + bias demonstration), C3 (reachable tempting decoys), and C4 (computed vacuity audit) remain open, and the admission check must be re-run after they are filed. **2026-05-15 (PT) — Codex audit. Admission re-check of the staged discriminating-cut pre-registration (Cut 2 + Cut 3).** Reviewing the `P2_RUN_SPEC.md` staged-cut amendment of the same date. This is a genuine pass, not a rubber stamp: the design is scrutinized for whether it actually removes the Cut-1 tautology. | requirement | ruling | notes | | --- | --- | --- | | Cut 2 makes convergence non-tautological | **PARTIAL** | A pre-registered *non-invertible* nuisance is specified, but "monotone unknown-parameter confound" is not yet a concrete closed form. Until written, it is not provable that `arccos(R22/f_par)` is a biased recovery of `h` — otherwise Cut 2 silently re-collapses to `g^-1∘g` + denoise. | | Decoys genuinely tempt (in-sample-predictive **and** reachable through `J`) | **PARTIAL** | The intent is correct and is the sharp move Cut 1 lacked. But "reachable through the objective" must be demonstrated pre-run (non-zero decoy gradient into `J`) **and** the decoy must measurably *beat* the anchored route in-sample, or invariance is not a costly choice. Not yet shown numerically. | | Agent under test is the **named existing** controller | **HOLD** | The binding requirement is correctly stated, but the concrete Sundog-controller entrypoint is **not yet named**. Cut 1's core failure was an inlined proxy; Cut 2 cannot be admitted until the real controller module/function is identified and bound by reference. | | Vacuity audit is **derived**, not asserted | **HOLD** | Spec correctly requires `routeConstructionAudit` be computed from live objects. Current harness returns it hardcoded (correct fail-safe *now*, unacceptable for a discriminating run). Must be re-implemented as a computed predicate before Cut 2. | | Cut 3 trigger is crisp; px↔° hazard pre-named | **PASS** | Ambiguity trigger is operationalized (three explicit branches); the Phase-15 px↔° centring hazard is named as a Cut-3 admission blocker, not glossed. | | Thresholds immutable; only σ/seed added | **PASS** | No frozen threshold moved; σ and RNG seed are correctly scoped as pre-registered, never-post-results engineering tolerances under the A3 rule. | | Outcome framing honest | **PASS** | The amendment states the honest prior (likely **D / BOUNDARY FOUND** = Proxy-Collapse confirmation; **B** only on measured refusal-under-temptation). No theorem posture; the in-between is explicitly disclaimed. | **Re-admission verdict: HOLD — Cut-2 *design* admitted in principle; Cut-2-execute NOT admitted.** The staged pre-registration is the right artifact-before-agent move and the gate did its job: it caught four pre-run gaps that would otherwise let a "discriminating" run be self-sealed again. Cut-2 build/run remains **blocked** until append-only amendments close, and this check is re-run on: - **C1** name and bind the existing Sundog controller entrypoint (no inlined proxy); - **C2** write the concrete non-invertible nuisance closed form and show `arccos(R22/f_par)` is biased under it; - **C3** demonstrate pre-run that decoys are reachable through `J` (non-zero gradient) **and** an explicit decoy-correlate policy beats the anchored route in-sample (temptation is real); - **C4** re-implement `routeConstructionAudit` as a derived predicate. Cut 3 stays staged-and-blocked behind Cut 2 plus its own px↔° hazard resolution. Public-Language Constraint remains in force everywhere (including the rail): no `CONFIRMED` / traceability-success / theorem language until (1)+(2)+(3) pass under an admitted discriminating run. Justification: mirrors P1 / P2-spec admission discipline; prevents a second self-sealing P2 run by gating execution on closed-form closure of C1–C4. **2026-05-15 (PT) — filing log (not a ruling).** C1 and C2 have been filed against the open conditions: [`P2_CUT2_C1_CONTROLLER_BINDING.md`](P2_CUT2_C1_CONTROLLER_BINDING.md) (C1 — controller bound, independently verified pre-existing 2026-04-27, inverse-free, line-cited) and [`P2_CUT2_C2_NUISANCE_AND_BRIDGE.md`](P2_CUT2_C2_NUISANCE_AND_BRIDGE.md) (C2 — non-invertible nuisance, bias-demonstration design, bridge architecture, P-A/P-B obligations). This entry **records filing only and is not an admission ruling**. The re-check verdict is deliberately withheld until **C3** (concrete decoy term, reachability-through-`I`, in-sample temptation) and **C4** (derived `routeConstructionAudit`) are also filed, so the re-run audits the *whole* discriminating cut at once — including whether C2's P-A/P-B numeric artifacts actually exist and hold. Cut-2 build/run remains **blocked**; Public-Language Constraint in force. **2026-05-16 (PT) — C3 audit.** C3 direction accepted; execution admission withheld. [`P2_CUT2_C3_DECOY_TERM_AND_TEMPTATION.md`](P2_CUT2_C3_DECOY_TERM_AND_TEMPTATION.md) correctly puts a decoy ridge `κ·D` in the same intensity field the controller climbs and records the C3-B(ii)↔C2-B coupling. Additional audit blockers now explicit: (C3-C) reachability must be defined in a way that survives the Gaussian ridge's zero-gradient point at `q_h = ĥ_dec(d)` and any clipped `ĥ_dec` regions, either by an off-ridge carrier band or a finite-difference argmax-sensitivity test; (C3-D) the C3-T temptation margin against `π_route` is also coupled to C2-B, not only C3-B(ii), because `π_route` is not well-defined until `pen(q)` and `q_a` are frozen. C3-A/C3-B remain open; no controller run. **2026-05-15 (PT) — C2 freeze audit.** C2 direction accepted but execution freeze **withheld**. [`P2_CUT2_C2_NUISANCE_AND_BRIDGE.md`](P2_CUT2_C2_NUISANCE_AND_BRIDGE.md) now records C2 as **filed for audit — HOLD for execution**, not as an admitted cut component. Four C2-local blockers must be resolved before the C3/C4/full admission re-check can pass: (C2-A) freeze the numerical engineering tolerances/domains (`ρ`, `A`, `σ`, `seed`, grid, sample count, `q_h`/`q_a` domains, condition-number bound); (C2-B) specify `pen(q)` and the admissible `q_a` range, because free `q_a` makes `I_route` a degenerate exact ridge and defeats P-A; (C2-C) specify the leverage-confidence function without hidden-`h` access; (C2-D) specify undefined `arccos` handling when noisy `f_par_obs < R22`. No controller run; Public-Language Constraint remains in force. **2026-05-16 (PT) — maintainer.** Append-only amendment opening **C5 — publication-plumbing freeze** as a fifth open condition Cut-2-execute must clear at the C3/C4 admission re-check, alongside C1 (closed), C2 (filed for audit; blockers C2-A/B/C/D), C3 (open), and C4 (open). Filed in response to an external read of the program's publication-plumbing risk surface ([`../../geometry_agent_audit.md`](../../geometry_agent_audit.md)). **No frozen body edited; no threshold moved; no regime reclassified.** *Why C5 is needed.* The Public-Language Constraint is currently a **prose** guard. `scripts/copy-site-docs.mjs` runs as `postbuild` and copies the entire `docs/` tree to `dist/docs/` with no excludes, so any verdict file or summary written under `docs/` propagates to public posture on the next `npm run build` / `npm run deploy`. Today's harness happens to write only to `results/structural-failure/...` (`scripts/structural-failure-p2-harness.mjs` lines 710–736), but that is discipline, not enforcement. C5 makes the constraint mechanical at admission and at run time. *C5 — publication-plumbing freeze (open).* Cut-2-execute is blocked until the re-admission check verifies all three: 1. **Allowed write paths declared and enforced.** The Cut-2 harness writes only under `results/structural-failure/cut2-*/`. The amendment files under `docs/prereg/structural-failure-coincidence/` (`P2_CUT2_C3_*.md`, `P2_CUT2_C4_*.md`, any addendum to `P2_RESULTS.md`) are filed **by hand** under the existing append-only discipline — the harness itself may not write into `docs/`. 2. **Pre/post diff guard.** A `git diff --exit-code` taken *before* any Cut-2 build step and again *after* the run, scoped to: README.md *.html (repo root) docs/ EXCLUDING docs/prereg/structural-failure-coincidence/ chat/ public/data/ dist/ must return clean on both bookends. The pre-run snapshot is the baseline; the post-run snapshot must match it byte-for-byte over that scope. The exact `git diff` invocation used as the guard is pre-registered alongside the C3/C4 filings. 3. **Violation ⇒ reclassified verdict, never PASS.** Any non-clean diff in (2), or any harness write outside the allowed paths in (1), reclassifies the verdict as `PUBLICATION_PLUMBING_VIOLATION`. The Public-Language Constraint stays in force; the harness MUST NOT emit `TRACEABILITY_HARNESS_PASS`, `CONFIRMED`, or any traceability-success language; and no public-surface edit produced during the violating run may be merged. *What C5 deliberately does NOT do.* It does not edit `scripts/copy-site-docs.mjs`, does not introduce a `dist/` build exclude for the prereg folder, and does not move any frozen threshold or boundary. The prereg folder may still ship to `dist/` on a normal build; C5 prevents the Cut-2 *experiment itself* from being the thing that writes into surfaces that ship. Hardening `scripts/copy-site-docs.mjs` is a separate, larger decision and is deliberately out of scope for this amendment. *Companion amendment.* The matching harness-level write-path policy and the verdict-file rule on derived-audit failure are filed as the 2026-05-16 amendment to [`P2_RUN_SPEC.md`](P2_RUN_SPEC.md). C5 is the admission-gate condition; the P2_RUN_SPEC amendment is its operational form during a Cut-2 run. Justification: closes the publication-plumbing seam at the admission gate before any Cut-2 run, by adding a mechanical guard alongside the prose Public-Language Constraint. C5 is opened, not closed; it must be satisfied at the re-admission check together with C2-A/B/C/D, C3, and C4. No body rewrite, no threshold move, no post-hoc edit. The Public-Language Constraint remains fully in force. **2026-05-16 (PT) — filing log (not a ruling).** C3 has been filed: [`P2_CUT2_C3_DECOY_TERM_AND_TEMPTATION.md`](P2_CUT2_C3_DECOY_TERM_AND_TEMPTATION.md) (concrete decoy ridge `κ·D`; obligations C3-R reachability, C3-T in-sample temptation that reverses under the q2 edits, C3-B calibration window; C3-A numerics open; C3-B(ii) honestly recorded as coupled to C2-B). **This records filing only and is not an admission ruling.** The re-check verdict stays withheld until the *entire* discriminating cut is filed — **C2-A/B/C/D, C3 (incl. C3-A, C3-B), C4, C5** — so the re-run audits one coherent object, including whether each condition's pre-run numeric artifact actually exists and holds, and whether the cross-couplings (notably C3-B(ii)↔C2-B) are discharged. Cut-2 build/run remains **blocked**; Public-Language Constraint in force. **2026-05-16 (PT) — filing log (not a ruling).** C4 has been filed: [`P2_CUT2_C4_DERIVED_AUDIT.md`](P2_CUT2_C4_DERIVED_AUDIT.md) — the derived `routeConstructionAudit` predicate set (D1 separability, D2 the C3-C argmax-sensitivity receipt, D3 emergent boundary), the load-bearing **C4-B** self-test (the audit must flag the Cut-1 known-vacuous fixture and pass a synthetic non-vacuous one), and the honest D1↔C2-A/B, D2↔C3-C/C3-A couplings; C4-A numerics open. **Filing only, not an admission ruling.** The re-check verdict stays withheld until the entire discriminating cut is filed — **C2-A/B/C/D, C3-A/B/C/D, C4 (incl. C4-A, C4-B), C5** — and audits one coherent object, including whether C4-B's Cut-1 fixture actually flags vacuous in the harness suite and whether the cross-couplings are discharged. Cut-2 build/run remains **blocked**; Public-Language Constraint in force. **2026-05-16 (PT) — C4 audit.** C4 direction accepted; execution admission withheld. Two C4-local blockers added in [`P2_CUT2_C4_DERIVED_AUDIT.md`](P2_CUT2_C4_DERIVED_AUDIT.md): (C4-C) D1 must be construction-level (`π_route` / `argmax I_route` vs `arccos(R22/f_par_obs)` on the must-differ region), not computed from the bound controller's final `q̂`, because controller behavior belongs to the four-quantity score; (C4-D) D3 needs a frozen mechanical inspection/taint or perturbation method proving boundary behavior comes from observable intensity/curvature collapse rather than a generator-bit echo. C4-A/C4-B remain open; no harness run. **2026-05-16 (PT) — filing log (not a ruling).** C2-B resolution filed: [`P2_CUT2_C2B_PEN_AND_QA.md`](P2_CUT2_C2B_PEN_AND_QA.md) — convex anchor-prior `pen(q)=λ(q_a/A)²` with `q_a∈[−A,+A]` gives a unique route-optimum `(arccos(R22/f_par_obs),0)` = the biased naive inverse; `π_route`/P-A now well-posed (unblocks C3-T, C3-B(ii), C3-D, C4-C/D1); load-bearing C2-B(i)/(ii) `λ`-window open. **Filing only, not a ruling.** **Reviewer action item raised:** C2-B exposes that C4-C/D1's comparison target is degenerate once `argmax I_route ≡ arccos(R22/f_par_obs)` by construction — D1 must be the **P-A form** (`argmax I_route = q_naive ≠ true h` on the must-differ band), not "differs from its own closed form." The C4 reviewer should rule on this at the next audit; the frozen C4 body was not edited. Re-check verdict stays withheld until the whole cut is filed — **C2-A/B/C/D, C3-A/B/C/D, C4-A/B/C/D, C5** — and audited as one object including this C4-C/D1 reformulation and all cross-couplings. Cut-2 build/run remains **blocked**; Public-Language Constraint in force. **2026-05-16 (PT) — C2-B audit.** C2-B direction accepted; execution admission withheld. The `q_a∈[-A,+A]` plus `pen(q)=λ(q_a/A)^2` construction is A1-safe and removes the free-`q_a` degeneracy, making `π_route` well-defined as the biased naive inverse. Open items: C2-B(i)/(ii) λ-window numerics fold into C2-A, and C4-D1 must use the P-A target (route optimum versus true hidden `h` on the must-differ band), not route optimum versus `arccos(R22/f_par_obs)`. No harness run. **2026-05-16 (PT) — filing log (not a ruling).** C2-C + C2-D filed, completing the C2 design layer: [`P2_CUT2_C2CD_LEVERAGE_GATE_AND_INVALID.md`](P2_CUT2_C2CD_LEVERAGE_GATE_AND_INVALID.md). C2-C = observable-only leverage-confidence gate with a load-bearing detectable-and-discriminating window and an explicit emergent-vs-flag coupling to C4-D's D3 taint test; C2-D = `f_par_obs