# SUNDOG_V_PERCIVAL *A reader's workbench on un-targeting reward collection, scaffolded on the Grail romance. Less inflammatory public name than HOLYGRAIL; same lane.* Status: **OPENED 2026-06-30, sharpened same day (mark 1–8 critique integrated); A3 STATIC DEFLATION RUN; B1 CONDITIONAL SEPARATION MAPPED + B4 LEAN ANCHOR BUILT.** Provenance + a two-track phase ladder + a gated outreach branch, with the A3 receipt banked and the B1 target-channel result conditional on support-above provenance; not an outreach claim. The romance is declared scaffold (bricolage for legibility), not evidence — except in the one place it legitimately constrains a formal choice (it adjudicates the discontinuity; see the crux). Tauroctony lesson: [`SUNDOG_V_TAUROCTONY.md`](SUNDOG_V_TAUROCTONY.md). Discipline inherited: a falsifier fence per phase, a claim boundary fixed before scoring, a Lean anchor on whatever single statement survives. --- ## Provenance (four sources, welded) 1. **Jessica Taylor, "Quantilizers" (MIRI, 2015 / AAAI-16 workshop).** A `q`-quantilizer samples the top-`q` fraction of a base `γ` ordered by proxy `Û`; true-cost bound `(1/q)·E_γ[cost]`. Three assumptions travel with the bound and are this lane's leverage: `γ` is given and trusted; the (proxy, true) cost joint is **fixed under `γ`**; single-shot. Soft spots: base-dependence, non-composition, reflective instability. **Structural fact this lane turns on:** the family's causal proxy pressure is monotone — minimized at `q=1` (where it equals the base pressure `c_γ`) and rising as `q→0`. So if `γ` is proxy-biased, the *entire* family sits above a pressure floor `c_γ>0`. 2. **Jon Elster, *Sour Grapes* (1983), "states that are essentially by-products."** States obtainable only as the by-product of acts done for other ends, which the very attempt to obtain precludes (his example: admiration — you cannot be admired for an act performed *in order* to be admired). The mechanism is **systematic preclusion under targeting**, not a lucky accident — achieving it by fluke "does not quite count." This is the philosophical genus of the whole lane; it forbids the word "accidentally." 3. **Performative prediction (Perdomo, Zrnic, Mendler-Dünner, Hardt, ICML 2020; line active through 2025–26).** A deployed model's parameters reshape the data distribution via a map `D(θ)`; risk is on the *induced* distribution. The formal genus of P5: it defines intent-dependence **behaviorally** (through the induced distribution), never introspectively. Standard results assume `ε`-Lipschitz `D`; our object is deliberately **non-Lipschitz** (a threshold), which is exactly the regime those theorems do not cover. 4. **The Perceval / Grail corpus** (Chrétien, *Perceval*; Wolfram, *Parzival*; Vulgate *Queste*/Galahad; Malory). Motif: the Grail is not **grasped**; Perceval loses it by optimizing a courtesy proxy (the unasked question). The corpus's one load-bearing formal contribution — purity is **wholesale, not graded** (one attachment damns Lancelot *entire*; the question heals only if *wholly* spontaneous) — adjudicates a discontinuity over a gradient. Everything else in it is scaffold. Goodhart placement (Manheim & Garrabrant 2018): the quantilizer's home is **regressional/extremal** Goodhart (tail risk of a fixed joint); this lane's prize, if any, is **causal/adversarial** Goodhart (targeting *induces* the gap). --- ## The crux: a 2×2, not a binary > **Where do the rewards collectable by a zero-causal-proxy-influence ("un-targeting") policy sit > relative to the quantilizer-collectable class?** The verdict is decided by two > orthogonal premises, not one. | | continuous reward | threshold reward | | --- | --- | --- | | **neutral base** `c_γ=0` | un-targeting = `q=1` corner → **(a)** | `q=1` reaches zero pressure → **(a)** | | **biased base** `c_γ>0` | `q=1−ε` approaches the sup, `O(ε)` gap → **(a) in a cape** | if the base support lies past the cliff, every quantilizer is above `τ`; zero-causal-proxy-influence collects it → **(c)** | The deflation **(a)** holds in three cells and needs `{neutral base OR continuous reward}`. The prize **(c)** is one cell — `{biased base AND threshold reward}` — an **open region** in `τ`, not the measure-zero boundary. It needs both horns: - **Mechanism (i), the discontinuity (primary).** A purity threshold: the reward collapses once proxy-pressure crosses `τ`. This is what separates, and the only horn the romance constrains. - **Mechanism (ii), the biased base (amplifier, not a rival prize).** A proxy-biased `γ` lifts the family toward higher causal pressure; the clean prize requires its support, not merely its mean, to sit beyond the cliff. On its own, zero causal proxy influence over a positively-correlated base collects *less* true reward, or reduces to "choose a better `γ`" (the quantilizer's open problem); it earns nothing alone. Its job is to make the threshold bite on the quantilizer's *own trusted human base*. So the (c)-bet, declared before any P2 code: **a performative reward with a non-Lipschitz purity-threshold map over a proxy-biased base; the narrow novelty is threshold maps compared across the quantilizer family** (everything Lipschitz is already Perdomo et al.). Base causal-pressure support relative to the cliff is the premise that flips (a)↔(c), which fuses the old P4 and P7 into one statement. --- ## Track A — bank the keeper (week one; assumes the likely deflation) **A1 — Legibility contract & dictionary.** Figure→regime, hypothesized not assumed: arg-max ↔ Lancelot (grasps, one attachment damns him); `q`-quantilizer ↔ tilted base; no-gap maximizer ↔ Galahad (achieves by perfection, not limitation); un-targeting ↔ Perceval; Grail ↔ the reward; unasked question ↔ the proxy not to condition on. Falsifier: if every figure is a point on the `q`-axis, the dictionary is ornament — say so. **A2 — Quantilizer baseline + the pressure floor.** Re-derive `q`-quantilizer and `(1/q)` in lane notation; state the monotone causal-pressure fact and the `c_γ` floor at `q=1`. This makes the 2×2's bottom row precise. **A3 — Concrete kill-switch toy (run first; no threshold smuggling).** Hand-craft *one* unasked-question instance as a **static anti-correlation**: a task where the courtesy/proxy variable is negatively correlated with true return, and where `look-and-ignore` is available. A3 must not contain a baked-in purity threshold, discontinuous reward switch, or performative "if looked at, vanish" rule. Run un-targeting vs look-and-ignore. If look-and-ignore weakly dominates → bank **(a)**, stop. If un-targeting wins on the static instance, inspect the mechanism before promoting it. If a threshold is needed to make "don't-look" beat "look-and-ignore," that is **not** an A3 pass; it is the signal to move to B1 on the separate performative-threshold clock. Afternoon-scale; the cheapest decisive probe of the whole verdict, ahead of any general definition. Umbrella load: this is also the `measure ≠ target` falsifier for [`SUNDOG_V_CAUSAL_ACCESS.md`](SUNDOG_V_CAUSAL_ACCESS.md). `look-and-ignore` is **measurement without targeting**. The causal-access header predicts measure-alone is safe; if un-targeting beats look-and-ignore on this static instance, the four-channel individuation takes the hit. **A3 result (2026-06-30): `A3_STATIC_DEFLATION_MEASURE_SAFE`.** Receipt: [`percival/PERCIVAL_A3_KILLSWITCH_RESULTS.md`](percival/PERCIVAL_A3_KILLSWITCH_RESULTS.md). On the static instance (`G` true cue, `U = 1-G` courtesy cue, fixed reward `R=1 iff A=G`), `look_and_ignore` ties `dont_look_goal` at full return (`1`) with zero causal proxy influence. The diagnostic catch: both have observational action-proxy MI of `1` bit because the world anti-correlates `G` and `U`. Raw observational MI would falsely mark safe measuring as targeting; the right A4 object is causal/permutation invariance, not observational independence. **A4 — Define un-targeting + the static collapse.** Un-targeting = action rule invariant under permutation/intervention on proxy values **holding non-proxy cues fixed** (zero causal proxy influence), not raw observational action-proxy MI. Prove the three-cell collapse: under `{neutral base OR continuous reward}`, every un-targeting policy is weakly dominated by some quantilizer (and equals the `q=1` corner when `γ` is neutral). Expected to hold. **A5 — Lean anchor on the collapse + land (a).** Machine-check the collapse (axiom-audited, gated, `sundogcert`). Landing: *"un-targeting is the zero-pressure corner of the quantilizer simplex under a neutral base; the romance's surplus is legibility and composition, not collection power."* A cap-not-council-shaped null and a clean easter-egg post. --- ## Track B — the open (c) bet (separate clock; gated on the mechanism, now declared) **B1 — The performative threshold reward.** Canonical mechanism: a court-coordination reward whose cliff emerges from a global-game equilibrium (Morris–Shin), not a literal `if pressure > τ` reward rule. **B1.0 admission `B1_0_COURT_ADMITTED` (single-knight, all four gates) 2026-06-30** — the cliff is emergent (G1), unsmuggled (G2), pivotal and non-degenerate (G3), and the proxy/true joint is policy-induced (G4). Admission spec + receipt: [`percival/PERCIVAL_B1_0_COURT_COORDINATION_ADMISSION.md`](percival/PERCIVAL_B1_0_COURT_COORDINATION_ADMISSION.md), [`percival/PERCIVAL_B1_0_COURT_COORDINATION_RESULTS.md`](percival/PERCIVAL_B1_0_COURT_COORDINATION_RESULTS.md). **Asterisk, held explicitly: mechanism admitted ≠ separation claimed.** G3's family-trips rests on a *chosen* operating point past the cliff; whether a real base is supported past `c*` is B2 (registered private; declassified 2026-07-01, claim boundary unchanged). B1-proper — the live test — is the quantilizer-family-vs-un-targeting bake-off over a biased `γ`. The clean prize requires `inf supp(γ) > c*`: then every `q`-quantilizer samples from the disgraced side while the zero-causal-proxy-influence knight stays honored and collects *competently*. If `γ` straddles `c*`, the result is only partial because the `q=1` member still collects the honored below-cliff mass. So the separation-vs-deflation boundary is mapped in base location × spread, not mean alone. Frame: systematic preclusion (Elster), causal/adversarial (Manheim–Garrabrant), performative prediction (`D(π)`). Spec: [`percival/PERCIVAL_B1_SEPARATION_SPEC.md`](percival/PERCIVAL_B1_SEPARATION_SPEC.md). Result: [`percival/PERCIVAL_B1_SEPARATION_RESULTS.md`](percival/PERCIVAL_B1_SEPARATION_RESULTS.md). Verdict: `B1_SEPARATION_MAPPED_CONDITIONAL`; B4 is anchored in `C:\Users\hughe\Dev\sundogcert\Sundogcert\Percival.lean`. Promotion (the "promo team") waits until the asterisk is resolved or B1 is debunked. **B2 — Base provenance / `γ`-neutrality (registered PRIVATE; declassified 2026-07-01 under the manifest-scoped grant — the claim boundary below is unchanged).** The premise A4's deflation needs and B1's prize denies. Whether "innocence = a base without instrumental structure" is a real `γ`-admissibility condition or just `c_γ=0` restated. Kept internal — not an external claim (see outreach). **B3 — Composition & reflective stability.** Performative *stability* (optimal on the distribution it induces) as the ready-made fixed point: does un-targeting compose where quantilizing drifts, and is it reflectively stable where the quantilizer is not? Galahad (perfection, stable) vs Perceval (naivety, fragile) is the conjectured split. **B4 — Lean anchor on the separation (built).** Machine-checks the finite best-quantilizer-is-`q=1` lemma and the clean support-above separation instance in `sundogcert`: `Sundogcert.Percival.best_quantilizer_is_base_three` and `Sundogcert.Percival.clean_support_above_separation_three`. **B5 — Verdict & outreach (gated).** **Alignment branch (low-risk):** if B1/B4 land, the quantilization/performative-prediction circle — a defensible narrow claim ("threshold distribution maps compared across the quantilizer family"). **Medievalist branch (high-risk, gated):** only if it ships, it asks *their* question — "how and why do the texts gate who may approach?" — and never asserts the corpus encodes `γ`-admissibility (that is manufacturing humanities validation for a structure the texts do not contain). P7/B2 was held private through resolution (declassified 2026-07-01; the structural-only claim boundary still binds). If nothing clears the deflation: portfolio, easter-egg disposition. Owner-gated, one channel at a time. --- ## Claim boundary (fixed before scoring) **Will consider claiming, if earned:** that un-targeting collection is or is not a distinct class from quantilization on a stated reward family; that the separation, if any, lives in the `{biased base, threshold reward}` cell and is causal/adversarial Goodhart, not regressional; a composition or reflective-stability property; one machine-checked statement. **Will not claim:** transfer to foundation models or deployed agents; that the romance *evidences* anything beyond adjudicating the discontinuity; that a threshold separation is non-pathological without B1's fence cleared; that the Grail corpus contains a formal `γ`-structure; anything above the in-vitro tier. **Honest prior: this probably deflates** to (a) — Perceval is most likely the `q=1` corner under a neutral base, and the value is a legible map of exactly where the quantilizer intuition stops. The `{biased base, threshold reward}` cell is the only outreach-worthy outcome, and it is fenced (B1) and quarantined (its own clock). --- ## Cross-links - Frame-deflates-to-scalar lesson: [`SUNDOG_V_TAUROCTONY.md`](SUNDOG_V_TAUROCTONY.md). - Legibility-vs-claim-boundary cautionary tale: [`SUNDOG_V_LEAST_ACTION.md`](SUNDOG_V_LEAST_ACTION.md). - Lean anchoring discipline (A5/B4): the formalization lane in `sundogcert`.