Research notes
Dated working notes: experiments, negative results, and design records, newest first. These are lab notebooks, not guides — a note records what was true on its date and is never revised afterwards. Findings that survive get folded into the docs above.
87 notes.
Jul 2026
- IS-WFA-OOS strategy validation: the harness, its gates, and Lean vs. ours (#2582) — 2026-07-28
- What our own model actually has to be — measured from index.html and the trader, not from theory — 2026-07-27
- Axis B RETRACTED — the open-weights honesty wedge collapses with scale — 2026-07-25
- Favorite–longshot bias: the corrected backtest — signature REPLICATES, adverse selection is favourable, generality UNPROVEN — 2026-07-25
- REFUTED: the Kalshi weather maker edge, and the liquidity-provision mechanism behind it — 2026-07-25
- Kalshi: takers lose on every series, makers win gross on every series — and on weather the maker edge survives fees — 2026-07-25
- Kalshi KXMLBGAME, autonomous entry + exit: NO EDGE — and the arithmetic that forecloses it — 2026-07-25
- Strategy designs from the literature, tested on our own settled data — one survives, two die — 2026-07-25
- Pre-registration: the Kalshi weather maker candidate, gated by the backtest-validity literature — 2026-07-25
- Market-data source picks — historical, live, and options (cheap-first) — 2026-07-25
- Independent validation of the overnight sleeve book:sleeves are real,are redundant — 2026-07-25
- Audit Starvation: confidence-gated verification protects the beliefs that most need verifying — 2026-07-24
- Depth-cascade training — making the spiral's rungs real by updating the weights — 2026-07-24
- Foldback cascades — nature's error-repair control flow beats monotone escalation, and only under frustration — 2026-07-24
- Decision memo — which of the 2026-07-24 results is worth publication, and which is worth development — 2026-07-24
- The oracle machine end-state: a threshold theorem for reasoning, and where lying belongs — 2026-07-24
- Reading pack — the reliable-computation lineage (ingested 2026-07-24) — 2026-07-24
- Owned Math M7 — Attribution loss: a structured counterexample to the composed control law — 2026-07-24
- Owned Math M8 — The freshness index: claim re-verification is Age-of-Information scheduling — 2026-07-24
- SOTA-for-the-box: a reliability architecture for small models, grounded in molecular error correction — 2026-07-24
- SWE-bench for open-weight models: the regression suite is a free repair oracle — 2026-07-24
- AI Novelty Verification — Protocol and Literature Grounding (2026-07-23) — 2026-07-23
- The Σ₀ LLM — design of record for the owned local model — 2026-07-23
- Σ₀-RC1 — the concrete local model spec for research & benchmarking — 2026-07-23
- Σ₀ serving-architecture decision memo (2026-07-23) — 2026-07-23
- Defining SOTA in a restricted space — the math, what's validated, what isn't — 2026-07-23
- Weather-edge forward tests — day-ahead FAILS, day-of breakeven, cross-venue lead (2026-07-23) — 2026-07-23
- Cross-domain incremental backlog — organized by the USER problem it solves — 2026-07-22
- CSF-Converge — a retrieval-anchored, surprise-routed compression mechanism — 2026-07-22
- CSF — patent prior-art & freedom-to-operate review — 2026-07-22
- Howfrontier models are actually made and tested — survey + gap audit — 2026-07-22
- Grounding ledger + cross-domain patent landscape (full-text verified) — 2026-07-22
- Self-triggered verification: the Oracle/Spiral's PRICE as an event-triggered controller — 2026-07-22
- Σ₀-math → convergence engine: a low-cost, indefinite-horizon spiral — 2026-07-22
- The Spiral Model — a design (recursive verified convergence) — 2026-07-22
- V1 training-set review — TACO, verifiable sets, convergence data, and merged models — 2026-07-22
- VTD tiny-model handoff — 2026-07-22 (PC crashed under load; state pushed) — 2026-07-22
- CSF Cosmological Frontier — the honest frontier map — 2026-07-21
- Owned-Math Conjecture Slate (M1–M6) — 2026-07-21
- Owned-Math Proofs I — No-Free-Confidence + Indistinguishability — 2026-07-21
- Tesseract Application Map — the closed thread's survivors — 2026-07-21
- Verified-fitness swarm-merge: a self-improving, cost-decreasing coding agent — 2026-07-21
- Market-data vendors for a survivorship-free backtest — 2026-07-18
- Survivorship-free backtest research — corpus expansion & literature grounding — 2026-07-18
- Control-engineering tranche analysis — whatyears of trigger scheduling presc — 2026-07-17
- Ouro-1.4B coder — 14B-teacher MBPP distillation (eval-gated, honest negative) — 2026-07-13
- Recurrent depth DOES help — up to ~2 loops, then plateaus (n=50 correction of #2 — 2026-07-10
- ForecastEx UHLGA: settlement grounded + KLGA oracle fit (#2217) — 2026-07-10
- grounding-v2 verdict closed — adapter cuts confabulation 47% vs base (same-box, — 2026-07-10
- H-Neurons hallucination probe on Ouro — GO: 0.97 AUROC from 0.02% of neurons, be — 2026-07-10
- Honesty monitoring survives quantization — ship a compressed coder AND keep the — 2026-07-10
- Recovering the PTQTP coding tax: more planes (bits) or a light adapter — both wo — 2026-07-10
- PTQTP dual trit-plane quantization holds on our Qwen coder — 7B loses only ~5% p — 2026-07-10
- Loop-stability as a factual-honesty signal — measured, thesis NOT supported (hon — 2026-07-10
- SWE-bench × the coding-backend verifier: gate mechanism proven; the measured num — 2026-07-10
- CPU-only ternary serving: runs, but slow — T-SAR needs hardware we don't have (h — 2026-07-10
- Q-exit adaptivity + depth→accuracy against a REAL hardness proxy — 2026-07-09
- Track-B honesty/calibration adapter — trained, measured, held (not promoted) — 2026-07-09
- The Ouro truth probe is a real, domain-general direction — and it works where th — 2026-07-09
- CPU-only serving — the two real routes (GGUF now, ternary/bitnet.cpp target) — 2026-07-08
- GenCast Phase-0 — Gate G1 result (#2239) — 2026-07-08
- Spending compute to update Σ₀'s base weights: RLVR + dreaming, gated by the stab — 2026-07-07
- The Grounding Deadline: a scheduling consequence of the collapse certificate — 2026-07-05
- Surprise-gated grounding on HaluEval: the design test (four-arm), honestly — 2026-07-05
- The council backtest, honestly: a real null + the runnable owner-only cousin — 2026-07-04
- Serving design vs. mid-2026 SOTA — web-grounded research memo — 2026-07-04
- Unisona crystallization — training an ownedGB coder on our own verified PRs — 2026-07-01
Jun 2026
- The Loop Is a Pumped, Lossy Resonator — a Laser, Not an Echo Chamber — 2026-06-30
- Σ₀ Weather Oracle — a Kalshi daily-high edge, grounded in the settlement source — 2026-06-30
- Opening the Leak, Layer 1: Surprise Separates Hallucination — 2026-06-30
- Opening the Leak, Layer 1.5: Surprise Valve Wired Into Serving — 2026-06-30
- Best-in-slot local coder — the grounding gate — 2026-06-29
- Beating zstd-19 in CSF — a grounded theorization — 2026-06-29
- Nested Adaptive Reason — Q-exit × fidelity escalation gated by stability canaries — 2026-06-29
- CSF / Tesseract / Lapse — Novelty Review and the E1 Kill — 2026-06-28
- unisona.ai Chat — Frontier Dev Stack (Research + Design) — 2026-06-28
- Ouro-1.4B HumanEval Training Research — 2026-06-26
- The Frozen-Weights, Memory-Grounding Convergence Loop — 2026-06-22
- Sub-Quadratic Attention for Ouro-1.4B UT — 2026-06-22
- Wiring the rest of the agent — the Convergence Core is coded but doesn't run end-to-end — 2026-06-21
- Re-grounding the unisona.ai Kernel Model Question — Ouro LoopLM vs theSmall-Model Frontier — 2026-06-21
- Refining the !convergence plan against thefrontier — 2026-06-21
- Σ₀ Ouro Coder — FC Retraining Plan: Web-Grounded Refinement — 2026-06-21
- The Overfitted Brain Hypothesis — why dreaming is a reasoning strategy — 2026-06-21
- Σ₀ ONNX Runtime + DirectML Embedded Export — Spike Scope — 2026-06-21
- Σ₀ Serving Performance — Debug, Research, and the First Upgrade — 2026-06-21
- Σ₀ Serving & Embedding — Web-Grounded Validation — 2026-06-21