← all documents · raw markdown · 13 KB

SRHT Literature Positioning — Sweep of 2026-08-24

Purpose: place every load-bearing element of the SRHT master paper against its nearest

published prior art, so the paper's "based on things people had, worked out with operators

that worked" claim becomes precise: which parts are inherited, which parts appear to be

original contributions, and what would confirm or overturn each placement.

Method: targeted web sweep (constrained beam decoding; constraint handling in

evolutionary/heuristic optimization; penalty-method theory; rank normalization), August 24,

2026. Caveat that applies to every "no analogue found" below: absence of a search hit is

evidence, not proof — the definitive check is a reviewer in the heuristic-search community.

---

1. The framework family — INHERITED (and the paper already says so)

Constraint-aware beam search over sequences is an established line:

constraints organize the beam into a grid by constraints-satisfied count.

divided across constraint-satisfaction levels; the efficiency fix for GBS.

generation, satisfaction folded into the beam objective with penalties + pruning +

grouping.

("A\*esque") estimates of *future* constraint satisfaction to the beam score.

Placement: SRHT §1's self-description ("constraint-aware best-first beam framework") is the

same genus as these. The decoding-side novelty cannot be the *existence* of such a score.

Differences of form: the above operate additively in log-likelihood space with

count/step-shaped penalties; SRHT's score is a bounded multiplicative product

g(R)·(1−χ)² over a rank-normalized cost. A multiplicative two-variable score is not, in

general, order-equivalent to any additive one (their level sets differ), so the *form* is

distinct — but distinct form alone is weak novelty, since infinitely many forms produce

similar behavior. What the decoding line does NOT contain (as far as this sweep found): any

theorem about when its scores collapse to plain likelihood ranking, or a closed-form

violation crossover. See §3, §4.

2. Hard gate vs graded gate — the COMPONENTS are inherited, the CHARACTERIZATION looks sharper than what is published

The evolutionary/heuristic constraint-handling literature has debated exactly this for ~30

years under other names:

*Theoretical and numerical constraint-handling techniques used with EAs*, CMAME 2002;

Smith & Coit's penalty-functions chapter; Kramer's 2010 review. Folklore verdict:

death penalty (discard infeasibles outright) works when feasible regions are large;

graded penalties usually better when infeasible space must be traversed.

lower violation wins — a *lexicographic* hard rule, the standard baseline.

objective-sort and violation-sort; the classic evidence that neither pure regime wins

everywhere.

handling techniques in EAs*, 2009: the literature home of "repairable constraints" as a

category.

Placement: SRHT's empirical finding — **irreversible constraints ⇒ hard gate optimal;

repairable constraints ⇒ tuned graded gate strictly dominant** (paired sign tests, z=2.18

and z=6.98) — is a crisper organizing hypothesis than the folklore ("it depends on the

problem"), and this sweep found no published statement of that specific

irreversibility-based dichotomy. But every ingredient (both gate shapes, the knowledge that

neither wins universally, repair as a distinct regime) is inherited. The honest claim:

*a sharpened, statistically disciplined restatement of a known tension, with a proposed

mechanism (repairability) for when each side wins.* What would confirm it as a

contribution: the dichotomy holding on the standard constrained-optimization test suites

(CEC benchmark sets) against Deb's rules and stochastic ranking as baselines — not just on

the two in-house task families. What would deflate it: Deb's rules or stochastic ranking

matching the tuned graded gate on the repairable family, which would reduce the result to

"known baselines already achieve this."

3. The Redundancy Theorem (Theorem 1) — underlying fact is FOLKLORE; the explicit statement for constraint-gated beam scores appears UNCLAIMED

The mathematical core — a strictly monotone transform of a scalar preserves order, so any

score that is a monotone function of the optimized cost is sort-equivalent to cost — is

folklore (it is the ordinal-utility invariance argument from decision theory, and the

reason rank-based fitness in GAs is invariant to monotone fitness rescaling). The OR

literature on "redundant constraints" (identifying constraints that never bind) is a

different notion and not prior art for this.

Placement: this sweep found **no published theorem stating the collapse condition for

constraint-aware beam scores** — i.e., "if violation is monotone in the optimized cost,

the constraint-gated score is order-equivalent to cost sort; the apparatus adds nothing."

The nearest neighbors (GBS/DBA/NeuroLogic) publish no such self-limiting result; the

penalty literature knows penalties can be redundant but states it empirically. Verdict:

shallow but apparently original as an explicit, proved statement in this setting — and

its main value is hygienic: it is the theorem that forced the SVP/crypto retraction.

Mathematically it is an easy consequence of monotonicity; no one should claim depth for

it, and a referee might call it a remark rather than a theorem. It is still the right

remark, correctly proved.

4. The crossover χ\* = 1 − √(g(R_A)/g(R_B)) — ADJACENT concept exists at a different level; this per-pair closed form appears UNCLAIMED

Exact-penalty theory (Di Pillo–Grippo lineage; SIAM J. Control & Optim.) proves thresholds

on the penalty *coefficient* above which the penalized optimum coincides with the

constrained optimum — a global, coefficient-level crossover. SRHT's χ\* is a different

object: a per-sibling-pair, closed-form violation threshold at which a

valid-but-costlier branch overtakes an invalid-but-cheaper one inside the beam, derived

directly from the score's algebra.

Placement: no hit for a per-pair closed-form crossover in beam/heuristic search. Verdict:

apparently original in form, modest in depth — it is two lines of algebra once the

score is fixed, and any smooth multiplicative gate yields *some* crossover; the specific

√ form is an artifact of the (1−χ)² choice. Its real value is diagnostic: it predicts,

quantitatively, when the score departs from cost-sort, and that prediction was verified

(2,000 pairs, zero violations). A companion observation with content: for siblings of

comparable cost, g(R_A)/g(R_B)→1 so χ\*→0 — even mild violations flip the order. That

sensitivity claim is testable against additive-penalty baselines, where the flip threshold

scales differently.

5. Everything else — INHERITED, competently chosen

(rank selection in GAs; score normalization in IR hybrid search). Not original; the

emergent floor ε/(s+ε) replacing a hand-set constant is good craftsmanship, not a result.

engineering with unusually honest bookkeeping (the superset-not-lossless counterexample

is the kind of thing most write-ups omit). Not a literature event.

---

Summary verdict (ranked, with confidence)

1. The proofs are correct. Near certain — machine-verified independently

(properties A.1–A.4, Theorem 1 over 2,000 sets, χ\* over 2,000 pairs, subsumption both

directions).

2. The framework genus is inherited from constrained beam decoding and penalty-based

constraint handling. Certain — and the paper's §1 already concedes this.

3. Three elements appear unclaimed in the literature as explicit statements: the

redundancy/guidance dichotomy as proved theorems for constraint-gated beam scores; the

per-pair closed-form crossover; the irreversibility-based hard-vs-graded

characterization with paired statistics. Moderate confidence (~70%) — a targeted sweep

found no analogues, but only expert review or a deeper citation crawl settles novelty.

4. None of the three is deep, and a hostile referee could downgrade Theorem 1 to a

remark. Their joint value is coherence: one bounded score, its exact limits, its exact

crossover, and an empirical regime map, all self-consistent and all falsifiable.

5. The publishable unit, if pursued: a short paper / workshop paper in heuristic

search or constrained optimization, contingent on one more experiment — the dichotomy

of §2 run against Deb's feasibility rules and stochastic ranking on standard (CEC-style)

constrained benchmarks. If it survives that, it is placed; if the baselines match the

tuned graded gate, the contribution reduces to the two theorems and the crossover.

Key sources

---

ADDENDUM (2026-08-24, same day): the deciding experiment was RUN

benchmarks/srht_baselines.py (pre-registered verdicts, fresh out-of-sample

seeds). Outcome: REDUCED. Stochastic ranking (Runarsson & Yao 2000) matches

the tuned graded gate's validity and beats its cost (z=3.87); Deb's rules beat

its validity (z=2.45). The hard-vs-graded characterization is therefore NOT a

contribution beyond known baselines. Elements 1 and 2 of the "apparently

unclaimed" list (the theorems and the crossover) stand; element 3 falls. The

earlier in-sample dominance claim (v9.10.775, z=6.98) is corrected in

benchmarks/srht_baselines_results.md - out-of-sample it is a trade, not

dominance. Section 2's placement above should now be read as: sharpened

restatement only, empirically matched by the 2000-era baselines.

ADDENDUM 2 (2026-08-24, later same day): the repaired gate WINS its confirmatory study

benchmarks/srht_capx10.py: CAPX — C-hat with chi = dip / (exact reachable

repair from the remaining move menus), deterministic, parameter-free — vs

stochastic ranking, 10 fresh seeds, 3,000 paired instances, pre-registered.

Verdict: SUPERIOR (pooled validity z=2.75, pooled cost z=3.33, never

significantly worse on any seed). Standing of element 3 is therefore revised:

the ORIGINAL fixed-vmax gate remains matched by the 2000-era baselines

(REDUCED, Addendum 1 stands), but the repaired gate — which computes the

hard-vs-graded dichotomy PER BRANCH from reachable repair — beats the champion

baseline on this family. Scope: requires a computable repair bound; one task

family; breadth untested.

ADDENDUM 3 (2026-08-24, evening): breadth study — CAPX does not generalize

benchmarks/srht_breadth.py: five families, pre-registered. F1 anchor: TIE at

n=900 (consistent with the small z=3.33 effect at n=3000). F2/F4: CAPX maxes

validity but Deb's rules dominate it outright (equal validity, cheaper).

F3 (deliberately optimistic repair bound): INFERIOR — CAPX's documented

failure mode is an unsound bound (chi<1 on truly dead branches). F5: trade.

Overall: DOES NOT GENERALIZE. Final practitioner map: sound bound + expensive

feasibility -> CAPX; cheap feasibility -> Deb; no bound -> stochastic ranking.

The 30-year folklore "no constraint-handling technique dominates" is

reproduced independently. SRHT's durable contributions after the full program:

the analytical layer, and per-branch irreversibility detection valid where a

sound repair bound exists.