AI Novelty Verification — Protocol and Literature Grounding (2026-07-23)

1. Why — the traps, each grounded

AI novelty verification fails in known, measured ways. Each trap below is tied to external literature (all IDs verified against arxiv.org and/or F:/arxiv-corpus on 2026-07-23) and to the live failure we caught in-repo the same day.

2. The protocol

1. Decompose the idea into functional atoms (claim-style), per the patent-examiner workflow

(arXiv:2502.06316). A verdict on the undecomposed idea is not admissible.

1. Per-atom prior-art table — each atom gets its own citations with an explicit overlap

statement. Every atom published ⇒ the only possible novelty is the composition.

1. Structural search beyond the home literature — deliberately query other vocabularies:

classical algorithms (iterative deepening, anytime/contract algorithms, restart schedules, cascades/reject rules), adjacent fields, and patents. This is the RND lesson (arXiv:2503.01508): compute the density of the whole neighborhood, not the distance to the nearest modern paper.

1. Citation authenticity audit — verify every load-bearing ID resolves (local corpus, then

arxiv.org) and that the paper says what the clearance claims it says. Title-match is not claim-match (our RLTT case). This is the arXiv-ban standard applied to ourselves.

1. Adversarial refutation — an independent pass whose brief is to kill the novelty claim

and the supporting math. Diversity of lens, not repetition, is what raises ensemble accuracy (arXiv:2604.15145).

1. Verdict vocabulary (capped): breakthrough — forbidden from AI passes entirely;

novel-synthesis / incremental / known-repackaged — AI-assignable; plus empirical-instantiation: the algorithm is known but the measurement on our substrate is absent — a legitimate, publishable residue that must never be dressed as an algorithmic contribution.

1. Human ratification — the operator ratifies the final grade; docs may not carry a novelty

claim above candidate-novel (seam_open) before that, and never above the ratified grade after.

3. Grounding table

All six items resolved on 2026-07-23; none of the pasted search-snippet claims had to be taken on faith. (The "Relative Neighbor Density" and "Automated Novelty Evaluation" items arrived without IDs and were resolved to 2503.01508 and 2507.11330 by search.)

4. Worked example — the depth-escalation clearance (same day)

Idea assessed: use an external correctness verifier as the inference-time deferral rule that buys more recurrence depth (R4→R8→R16) on one shared weight-tied looped model — a cascade whose tiers are depths, not models.

• Pass(modern-LLM clearance, 14-entry prior-art table): graded incremental; sliver = "no

single paper reports this exact inference-time wiring." Correctly resisted the self-speculative decoding surface match (their "verify" is lossless distributional equality for speed, not external correctness).

• Pass(adversarial refutation): upheld incremental; separately collapsed the economic

headline (mediant non-dominance; regime sandwich; measured depth plateau at ~2 loops).

• Pass(trap audits): citations — all real, two claim-drift nits. Classical-CS structural

audit — sliver killed: the composition is Korfdepth-first iterative deepening (same procedure, deeper cutoff, external goal test, geometric amortization), Luby–Sinclair–Zuckerman(provably optimal escalating-budget schedules for generate-then-externally-verify), Hansen–Zilberstein monitored anytime stopping (the deferral rule, DP-optimal), and in the LLM era Reflexion/Self-Debugging and AlphaCode/CodeT (same weights, exec-verifier buys more compute). Patent sweep (manual, free surfaces): corroborates the kill — Intel US11869232B2 (2024) claims "an exit determination as to whether an output of the first subset of layers satisfies one or more exit criteria," selectively bypassing (equivalently: continuing) deeper layers, with speculative deep execution while pending; with US12217172B2 (adaptive off-ramps) and US20220358358 (dynamic early exits), criterion-gated depth continuation is anticipated at patent-claim generality. Coverage limits: free-surface search only (Google Patents snippets/USPTO); no examiner-grade claim charts, no EPO OPS family search — sufficient here only because the classical-CS audit had already killed the sliver independently.

• Final grade: known-repackaged (algorithm) + empirical-instantiation (the depth-rescue

profile of a looped LM under an external verifier is unmeasured — a measurement claim only).

• What survived untouched: the deployment argument — depth escalation is the only ΔRAM=0

local escalation axis on an 8GB box — which never depended on novelty, and which now importsyears of optimality theory (Luby schedules, Hansen–Zilberstein monitoring) instead of deriving its own.

The example is the protocol's own justification: a disciplined single AI pass still overstated by exactly one trap (structural echo under foreign vocabulary), and the ensemble caught it.

5. Limits

• Novelty is falsifiable, never provable. Every verdict is "no anticipation found within the

searched coverage"; coverage is always partial (non-English literature, unindexed industry systems, unpublished internal work).

• Patent coverage is shallow without professional search tooling; free surfaces (Google

Patents, EPO OPS) miss claim-family breadth. A seam_open grade must state what was not searched.

• The verifier can drift even on real citations — title-match ≠ claim-match; stepexists

because our own pass attached a wrong mechanism to a real paper.

• Ensemble ≠ oracle. 90.1% (arXiv:2604.15145) is the measured ceiling of the best current

ensemble on their benchmark — one in ten verdicts is still wrong. Hence the human gate.