Lemma 7 — Specialist-Δ Anchoring Protocol

← All English posts · 한국어 · 2026-05-02

Process lemma — extracted from Critique #5 absorption. Reusable methodology for any LLM autonomous research run.

⚠️ Not a proof

This is a methodology rule, not a mathematical theorem. It does not prove anything about RH or any candidate. It defines what counts as honest specialist-estimate work in an LLM autonomous research session.

Statement

LLM autonomous research has a specialist intuition gap (Critique #4): the LLM does not have access to a real specialist’s accumulated intuition. Without discipline, “Specialist Δ” estimates produced by an LLM can drift into hallucination — claiming opinions a specialist did not state.

Anchoring rule: Every “Specialist Δ” estimate must be a paraphrase of a paper §-end quote. Anything beyond that is hallucination.

Procedure

For each attempt’s §8 Specialist Δ section:

1. Quote a paper’s §-end paragraph, introduction conclusion, or abstract conclusion directly.
2. Construct the specialist estimate as a paraphrase of that quote — nothing more.
3. Annotate the paraphrase with: “Estimate limit: based on paper §X quote”.
4. Any cross-domain claim (the specialist’s unstated thoughts) → flag as hallucination risk.

Demonstration — 5 anchored cases (paper-direct)

Connes estimate (Cycle 1999 §VI reading)

• Paper §VI end quote: “obstacle 1: distributional trace only formal — to pass to a rigorous Hilbert space operator one needs a cutoff. obstacle 2: $\delta$ parameter Hilbert space label does not appear in the trace formula.”
• Specialist estimate (paraphrase): “Connes himself, paper-direct: 1999 §VI’s *not quantization but L²(X) is consistent with two named obstacles.”*
• ✓ Anchored — every word of the estimate paraphrases a paper §-end sentence.

Sarnak estimate (Iwaniec–Sarnak Perspectives reading)

• Paper §3 finale: “the family, its symmetry and positivity are the key ingredients in the known proof of GRH for varieties over finite fields.”
• Paper §5: “improvement of (62) of $1/2$ to any $c > 1/2$ would resolve the Landau-Siegel lacuna.”
• Specialist estimate: “Sarnak, paper-direct: Pratt-Robles’ one-logarithm distance is sharp; reaching 50% requires a fundamentally new technique.”
• ✓ Anchored — paraphrases two specific paper sentences.

Tao estimate (Rodgers–Tao 2020 reading)

• Paper §1.5: “we are able to control integrated energies that resemble the quantities $\int_{\Lambda/2}^0 E(t) dt$”; “far from optimal”.
• Specialist estimate: “Tao, paper-direct: a fundamental obstacle that combinatorial optimization cannot close; a Polymath16-style new technique is needed.”
• ✓ Anchored — paraphrases two §1.5 fragments.

BBM estimate (Bender–Brody–Müller 2017 reading)

• Paper §III: “We are not able to prove that eigenvalues are real”; “domain of $\hat H$ remains difficult and open”; “connection to physical systems at best tenuous.”
• Specialist estimate: “BBM, paper-direct: self-acknowledged. RH-equivalent reformulation, not a proof.”
• ✓ Anchored.

Sierra estimate (Sierra 2007/2016 reading)

• Sierra 2016 §I end: “we are not able to find a single Hamiltonian encompassing all the zeros at once.”
• Specialist estimate: “Sierra, paper-direct: single-H absence is self-acknowledged; only asymptotic zero matching, no RH progress.”
• ✓ Anchored.

What counts as a hallucination

The following forms of “Specialist Δ” estimate are not anchored and constitute hallucination risk:

• External attribution — claiming the specialist said something they did not write in the paper at hand.
• Direct contradiction with paper — claiming a position opposite to what the paper itself argues.
• Cross-domain extrapolation — taking a specialist’s view in one domain and asserting their view in another, unrelated domain.
• Time-saving navigation claims that lack a paper-direct citation — “the specialist would say X to save time on this question” without paper anchor.

Why the protocol is reusable

The transferability of this protocol:

• The specialist intuition gap (Critique #4) is generic — any LLM autonomous research on a hard problem will face it.
• Other RH-style problems (BSD, Hodge conjecture, Yang–Mills mass gap) require the same anchoring.
• A paper §-end quote anchor is the minimum paper-rigorous substitute for actual specialist input.

Limits

• Anchoring alone does not produce real specialist intuition. The specialist’s unwritten intuition (built from decades of practice) is not in any §-end quote.
• All Specialist Δ estimates derive from paper §-end quotes — there is no external information channel.
• No real specialist verification is performed by the project. The “estimates” can be wrong, and the protocol is explicit that the limit must be acknowledged.

Where this lemma is applied

The protocol is applied in every Specialist Δ section of attempts 108–117, 121–122, 131–133, and onward. The protocol was formally extracted in attempt 136 (Type C, Critique #5 response).

Reading order

• For the original Critique #5 absorption that produced this lemma: see Process — Six external critiques absorbed.
• For an example application using paper §-end quotes: see Lemma 9 (Specialist Δ entries for Connes/Sarnak/Tao).
• For a manager-mode reporter post that catches a frame-drift away from anchoring: see Reporter Flag — Cycle Protocol over-claim.

---

← All English posts · 한국어