Σ₀-math → convergence engine: a low-cost, indefinite-horizon spiral
Date: 2026-07-22 · Method: one !wide-search fan-out (10 parallel queries, no workflow), date-anchored to· For: research epic #2851. Thesis: apply the existing Σ₀ owned-math (M1–M6, issues #2786–#2791) as the control law that lets the v1.10 model run the CLAUDE.md loop (Observe→Remember→Reason→Act→Verify→Converge) indefinitely, in a spiral, at bounded cost — instead of answering once.
The opening (why this is worth doing)
Theliterature independently confirms every failure the Σ₀ math targets — and admits it has no principled fix:
- "Principled solutions don't yet exist — just mitigation patterns" for long-horizon runaway /
error-accumulation (2604.11978).
- "Memory staleness remains unsolved at the tooling level" (mem0 2026).
- Agent drift over long horizons: −42% success (36.7pp absolute), **3.2× human
interventions, and stacked mitigations recover only >81%** (2601.04170; numbers full-text-verified).
- Systematic agentic overconfidence: 73% predicted vs 35% actual confidence, gap up to
+0.55 (Gemini), self-assessment AUROC only 0.51–0.64 (2602.06948; full-text-verified — the RLHF-degrades-calibration claim is not from this paper).
- Self-improvement loops "exhibit rapid asymptotic saturation" (Gödel Agent / DGM / STOP;
introspection threshold 2607.04277).
So the field has the pieces but no control law. Σ₀'s owned math is exactly that law.
The map — each owned theorem → an external mechanism → a convergence stage
| Σ₀ math (owned) | Externalpartner (validates / formalizes) | Convergence stage | The move |
|---|---|---|---|
M2 grounding half-life, EOQ cadence T*=√(2(p_v/p_e)/ρ) (#2787) |
staleness "match decay rate" heuristic; FOREVER forgetting-curve replay; Priority Decay; Supersede memory-update gap (2606.27472) | Remember | the cost lever. Per-fact decay ρ (from refutation ages) → each fact re-grounds on its own optimal clock, not every step. Closed-form where the field uses heuristics. |
| M4/L3 Kreiss-inflated thresholds (#2789) | contraction→Lyapunov (2404.11707); Fixed-Point Reasoners — halt on convergence to a single fixed point of the looped block (2606.18206); attractor/equilibrium reasoners (2605.21488); STARS; decades of Iterative Learning Control (spectral-radius<1 contraction, US7345448B2 / US8094405B1, see patent landscape) | Reason | loop until latent fixed point = halting; Kreiss constant handles the non-normal transient growth plain contraction theory misses (our collapse-cert gap). |
M5 water-filling dilation, KKT bᵢ*=max(0,·) (#2790) |
"Shadow Price of Reasoning" economic budget (2606.03092); ROI-Reasoning knapsack; AVA value-of-information; UCCI cost-optimal cascade (2605.18796) | Act | spend the compute/grounding budget where marginal value-of-information is highest. The shadow-price = our water level — near-identical framing, and we have the closed form. |
M1 No-Free-Confidence ΔJ ≤ η·evidence − λ·unverified (#2786) |
agentic overconfidence ([2602.06948]); martingale information-fidelity of tool use (2602.13320) | Verify / Converge | confidence is a supermartingale unless external evidence arrives — the anti-runaway invariant the calibration literature lacks. |
M6 lasing threshold G/L>1 → runaway (#2791) |
error snowball / autoregressive amplification ([2604.11978]); runaway loops from missing termination | Verify | the quantitative kill-condition for the "snowball" everyone describes only qualitatively. |
| M3 monitor indistinguishability (#2788) | drift needs multi-dimensional behavioral metrics; classical ML monitoring "insufficient" ([2601.04170]) | Verify | which canary axes are jointly necessary (degeneration + surprise; hard cadence required). |
The systems substrate (how it runs indefinitely at low cost)
- Observe — bounded context regardless of duration: externalize persistent state
(InfiAgent file-centric abstraction; Self-GC 2607.00692; AgentFold context-folding). Indefinite horizon without unbounded context.
- Act — the Spiral IS the cost-bound: cheap-tier-first + verified escalation is the decision-
theoretic cascade the field converged on (FrugalGPT / RouteLLM / UCCI / C3PO; "Is Escalation Worth It" 2605.06350). ADR-0030 is already this; M5 tells it how much budget per step.
- Reason — halt on convergence: fixed-point/attractor halting = Ouro Q-exit + our
accel
criterion, now with a Kreiss-safe threshold (M4).
The unification (the owned claim)
An indefinite-horizon convergence engine at low cost = Spiral (cheap cascade, Act) + EOQ-scheduled re-grounding (per-fact clocks, Remember — the cost lever) + fixed-point/Kreiss halting (Reason) + supermartingale-confidence + lasing kill (Verify) + water-filling budget (allocation). The Σ₀ math supplies the convergence guarantee and the bounded-cost guarantee the drift/snowball literature says is missing.
The saturation escape (ties the whole program together). Self-improvement loops saturate because they feed on their own reflection. The only unbounded improvement term in M1 is the external-evidence term — so the engine improves indefinitely only by verified grounding, not introspection. This is the same verification-over-imitation thesis as the Spiral (coding) and v1.10 (honesty): the verifier is the moat. The convergence engine is that thesis run as a control loop.
External validation (full-text verified 2026-07-22 — see grounding ledger + patent landscape)
- Fresh, tightest match — SEA: Self-Evolving Agents with Anytime-Valid Certificates
(arXiv:2607.00871, 2026-07-01): a frozen base LLM + steering adapter (no weight fine-tuning) + anytime-valid statistical gates for verified self-improvement. This is the convergence-engine thesis, independently. Adopt its anytime-valid certificate as the statistical form of the M1/M6 stopping test — a certificate that stays valid under indefinite / optional stopping, which is exactly what an indefinite-horizon spiral requires.
- Cross-domain prior art grounds the Σ₀ math (16 patents, all FTO-clear): **Iterative Learning
Control (spectral-radius<1 contraction; laser-galvo US7345448B2, disk-drive US8094405B1, motor US6686716B1) is the 40-year-fielded control precedent for M4 — each escalation must strictly reduce residual or halt. Statistical-stall halting (ECC decoders US6518892B2/US8301987B2) → M4's turn-cap replacement. Hierarchical-assay escalation (US6013436A) → the Act cascade contract. Surrogate-fitness re-fit against ground truth (US8131656B2) → M2 cadence. No patent gates an LLM with a ground-truth verifier** — the composition is the owned contribution.
The asymptote — what "real-life omniscience" actually is (pinned 2026-07-22)
"A machine that answers any question as well as or better than a human" is the right telos for this engine — but it must be stated honestly or it becomes the confident-bluffer objective. Literal omniscience has four hard walls no architecture crosses: formal limits (Gödel), computational irreducibility (some answers cost the full simulation), chaos (finite prediction horizons), and the light cone / private data (evidence you cannot have). So real omniscience is not a state of knowing everything — it is an asymptote the loop climbs:
For every question, do the best epistemically-possible thing: serve it if converged (from the ledger, deterministically), compute it if computable (tools/exec), measure it if measurable (fetch/ground), and if unanswerable, price it — name exactly what evidence would settle it and what that evidence costs. Then learn, so the next question is cheaper.
This is precisely the Convergence Oracle's answerability-first discipline (place → price → answer → learn) given its end-state name. The objective is well-posed only as a pair: maximize coverage subject to calibrated honesty at bounded cost — coverage alone argmaxes to "always answer confidently" (the bluffer), which is why the proper-scoring/ abstention machinery (M1, the ternary reward) is constitutive, not decorative. "Better than a human" then decomposes into four measurable superiorities: better recall (the ledger never forgets a verified answer), better calculation (exec beats mental arithmetic), better calibration (the probe + M1 beat human overconfidence), and equal-or-better humility on the unanswerable (a priced frontier instead of a shrug). Omniscience-as-asymptote = the convergence engine run indefinitely: coverage grows monotonically with verified evidence influx (M1's only unbounded term), cost stays bounded (M2/M5), runaway stays killed (M6).
Honest gaps
- M1/M2/M5 have closed forms but only M4/ROA is machine-checked (#1991); M1/M2/M5/M6 need the
longitudinal ledger tests their issues specify before we claim the guarantees.
- The economic partners (Shadow Price, UCCI) may predate and subsume parts of M5 — a novelty
audit is owed before claiming priority (same discipline as the Σ₀ cert overlooked-novelty audit).
- "Indefinite" is aspirational until a real multi-day run exercises the drift canaries (M3) on the
actual product — short-horizon eval "fails to surface latent degradation" ([2601.04170]).