Grounding ledger + cross-domain patent landscape (full-text verified)

Date: 2026-07-22 · Method: two background workflows — (1) full-text fetch + strict verification of every cited paper/patent against the real source (40 agents), (2) cross-domain patent discovery throughdomain lenses →candidates →deep-read (32 agents). This closes the "grounded on search snippets, not the papers" gap flagged earlier.

Part— Citation verification: 33 /hold,partial,wrong,unreachable

Every source was reachable and none was fabricated. Seven citations were overstated in the design docs and are corrected below (fixes applied in the same commit).

Source	Verdict	Correction applied
2407.18418 (abstention survey)	partial	Real error. The paper does not say "DPO drops OOD" — it attributes OOD-generalization failure to abstention-aware instruction-tuning generally (Feng et al.). CPO/ORPO are not in the paper at all. Both sub-claims removed; kept the two that hold (SFT over-abstains; pretraining refusal-data understudied).
2602.06948 (agentic overconfidence)	partial	Real error. The paper does not study or claim "RLHF degrades calibration." Its finding is about assessment framing (adversarial bug-finding reframing improves calibration). RLHF clause dropped; cite instead its real numbers: 73% predicted vs 35% actual confidence; overconfidence gap up to +0.55 (Gemini); AUROC 0.51–0.64.
2511.21437 (model-merging survey)	partial	Real error. This survey does not study DARE. Its six methods: Task Arithmetic, TIES, Model Stock, TSV-Merge, Iso-C, Subspace Boosting. Only Task Arithmetic is reliably net-positive (peak +1.02% at n=12); subspace methods often degrade (down to −5.36%). DARE citation moved to its own paper (2311.03099, which holds).
2601.04170 (agent drift)	partial	Numbers confirmed (−42% success = 36.7pp absolute; 3.2× interventions; >81% is the combined mitigation only). But the "classical ML monitoring insufficient" clause was not in the paper — removed.
2507.06261 (Gemini 2.5)	partial	Thinking-budget range is 128–32768 (Pro) / 0–24576 (Flash), not "1024–32768"; "monotonic" softened to "generally improves."
2511.11500 (Reinforced Hesitation)	partial	Ternary reward +1/0/−λ + Pareto-over-λ + post-training RL all confirmed. But it trains on Knights & Knaves (80k/10k), not GSM8K/MedQA/GPQA (those were frontier-model probes). New citable fact: frontier models essentially never abstain even under heavy penalty (MedQA: zero abstention acrossmodels ×λ) — strong motivation; RH cascading hits 88.1% at 2.2 avg queries.
2606.18206 (Fixed-Point Reasoners)	partial	Convergence is to a single fixed point of the looped shared block, not "per-layer fixed points." Beats HRM/TRM at 7M params (Sudoku-Extreme +19.5pt); ARC-2 the lone loss.

The two load-bearing citations held exactly:

2606.02628 (white-box probe) — confirmed: linear AUROC 0.904–1.000 on 4-bit NF4 7–8B, sampling detectors ≤0.541, mid-late band, ~linear (one nuance: MLP gap is exactly ≤0.010, not strictly <0.01). The spine of v1.10 is solid.
2510.09033 (recall-not-truth) — confirmed: associated-hallucination detection is near-random black-box (0.48–0.49) and only weak white-box (0.57–0.69), vs UH 0.81–0.93. Scope note: their "AH" = spurious Wikidata-triple associations on 8B models; our probe-ladder assoc set = common misconceptions and scored 0.924 at 7B — different constructs, so our result does not overturn theirs; both are cited honestly.

Part— Fresh source (this week): SEA — Self-Evolving Agents with Anytime-Valid Certificates

arXiv:2607.00871 (submitted 2026-07-01). Four-layer stack around a frozen base LLM (no weight fine-tuning): L1 steering adapter (logit-level REINFORCE over a PAC-Bayes posterior, frozen prior), L2 mutable harness (prompts/tools/budgets), L3 loop controllers applying anytime-valid statistical gates. It is an unusually tight independent match for the convergence-engine thesis: persistent improvement without weight modification, gated by verified (anytime-valid) certificates — exactly CLAUDE.md's "learn via memory+retrieval, not retraining" + the Verify gate. Adopt its anytime-valid certificate as the statistical form of the M1/M6 stopping test (a certificate that stays valid under indefinite/optional stopping — precisely what an indefinite-horizon spiral needs).

Part— Cross-domain patent landscape (20 deep-read; 16 inspiration,adjacent,FTO risk)

Every match is cross-domain (non-AI) and FTO-clear — none reads on an LLM proposer/verifier loop, and the spiral's novelty is precisely the exec/evidence verifier gate they all lack. The value is inspiration: decades-fielded control-theory precedent for the Σ₀ math.

Iterative Learning Control ≡ M4 (contraction / fixed-point halting) — the strongest theme

US7345448B2 (Electro Scientific, laser galvo, 2004) · US8094405B1 (Marvell disk-drive RRO servo, 2007) · US6686716B1 (Exelis/Purdue motor, 2001) · CN111510020B (2020). ILC converges by feeding the measured error back through the plant to refine the input, not by retuning the plant — the servo analogue of "store experience + refine, don't retrain weights." Its convergence guarantee requires the error operator be contractive (spectral radius < 1) — a 40-year-fielded justification for M4's Kreiss/contraction halting: each escalation must strictly reduce residual or halt, else refinement diverges. US6686716B1 goes further — it feeds the error back to re-parameterize the cheap tier itself (= our VTD flywheel).

Statistical-stall halting ≡ M4

US6518892B2 (Broadcom iterative decoder, 2000): halt when a cheap hash of the state repeats (fixed point). US8301987B2 (Western Digital ECC, 2009): halt on N consecutive zero-discrepancy steps. Teaching: don't run a fixed iteration budget — halt on a cheap statistical stall signal (directly implementable as the spiral's turn-cap replacement).

Escalation contract ≡ Act cascade + M5

US6013436A (Visible Genetics hierarchical medical assay, 1994): run the cheapest test on all samples; a positive terminates immediately; only a negative escalates to the costlier assay — a clean, ground-truth escalation contract. US7254641B2 (Intervoice, 2001): bidirectional tier movement — adds a de-escalation direction to M5 (drop back to the cheap tier when load/uncertainty falls). US20120243734A1 (HP, 2009): decouple the gate decision from the confidence estimate — aggregate stage responses into one calibrated probability instead of inferring confidence from where it rejected (a concrete M1 refinement).

Re-grounding / surrogate discipline ≡ M2

US8131656B2 (UIUC evolutionary-computation surrogate fitness, 2006): a cheap surrogate evaluator is trusted only because it is periodically re-fit against real evaluations — the disciplined form of M2's "when may I substitute the cheap check for the expensive one?" (leash the cheap tier to periodic ground-truth).

Calibrated abstain & the learned-gate contrast ≡ M1

US9715723B2 (Applied Materials, 2012): a fielded calibrated-abstain boundary — explicitly refuses to treat uncertain as certain. US12135927B2 (IBM, 2020) is the instructive contrast: it patents a learned-imitation gate (train an ML model to copy an expert's accept/reject) — the spiral deliberately does the opposite (a deterministic exec/evidence verifier with ground-truth outcomes), which is both its novelty and its FTO safety.

Verify-then-refine ≡ Verify → Converge

US20100094676A1 (Bowe Bell + Howell, 2008): measure whether a corrective action actually restored the target metric before closing the loop, and refine the remediation playbook from the outcome — structurally identical to the Fix-Rate ratchet.

Landscape conclusion. The spiral's mechanisms are deeply precedented in control theory, reliability engineering, and operations research (which is reassuring — the Σ₀ math is standing on fielded ground), but no patent gates an LLM proposer with a ground-truth exec/evidence verifier, and none combines the M1–M6 control law. FTO is clean; the owned contribution is the composition applied to a reasoning loop — consistent with "the moat is the system."