Lumisonde — The Brief

The LLM Edge in Finance Is a Division of Labor

Wed, 26 Aug 2026 00:00:00 GMT

THE QUESTION: Where does an LLM actually add value in a financial system—numerical forecasting, unstructured-information understanding, macro interpretation, portfolio decisions, or quantitative research—and where do traditional systems remain stronger?

WHY IT MATTERS: Financial systems already combine text, market data, macro releases, research code, portfolio construction, and execution. Those stages do not carry the same consequences: an inspectable feature can remain advisory, while a portfolio decision can move capital.

THE ANSWER: The evidence supports a division of labor. LLMs look most useful with semantic, heterogeneous inputs and inspectable, reversible outputs; statistical, machine-learning, and deterministic systems remain the stronger default for numerical prediction, calibration, optimization, constraints, and execution. Each module therefore needs a matched control and an endpoint suited to its claim, with point-in-time data, equal search budgets, repeated runs, and executable costs. No independently reproducible matched system in this evidence set shows persistent value across every role, so authority should contract as an output approaches a capital decision.

Receipts and full analysis →

Can This Checkpoint Still Learn?

Tue, 25 Aug 2026 00:00:00 GMT

THE QUESTION: Under which regimes can a network retain useful performance yet lose the ability to learn the next target, how is that distinguished from nearby failures, and which interventions have controlled evidence?

WHY IT MATTERS: Repeatedly trained systems must choose how to use new data: keep the full checkpoint, reset mutable state but keep the weights, or start from scratch — while protecting behavior already needed.

THE ANSWER: Present accuracy cannot settle it. Measure future learnability and retained performance separately, because plasticity loss is a history-dependent deficit on a new-target learning curve while forgetting is damage to previously learned behavior; controlled long-horizon experiments establish later-task deficits in bounded settings, but not a field-wide law or the fully matched conjunction of stable retention and impaired new learning.[1] The comparison is three arms — aged weights plus training state, weights with optimizer state reset, and scratch — under matched budgets; the report gives the promotion rule for each. If old-target retention cannot be measured, stop: the experiment cannot support a claim that the updated checkpoint still remembers.

Receipts and full analysis →

The Memory Was Right. The Decision Was Wrong.

Mon, 24 Aug 2026 00:00:00 GMT

THE QUESTION: Why can a long-running language-model agent make a worse decision after retrieving an accurate, relevant memory—a record correct as history, yet the wrong premise for what it should do now?

WHY IT MATTERS: Persistent memory turns a stateless responder into a system whose earlier interpretations can steer later actions; once they survive across sessions, each write, update, retrieval, and reuse can affect a consequential choice.

THE ANSWER: Retrieval quality is only one link in a state-maintenance and decision loop. A useful memory must also be current, applicable to this task, valid in this scope, sufficiently independent of other support, and assembled so that the model uses it correctly; failure at any link can make a true historical record the wrong premise for action. The practical rule: admit records selectively, turn accepted events into versioned current state, retrieve the smallest decision-sufficient bundle, revalidate volatile premises before action, and update memory from external outcomes. The clean case—a memory that passes every condition—remains open: existing failure cases each leave at least one condition uncontrolled.

Receipts and full analysis →

How SPADE turns self-generated worlds into a bounded training loop

Sun, 23 Aug 2026 00:00:00 GMT

THE QUESTION: Can a pretrained LLM agent become more capable by generating and solving its own tasks, with no human-written QA pairs and no stronger-teacher demonstrations—and when does that loop improve rather than collapse?

WHY IT MATTERS: Any training loop needs a continuing supply of useful tasks, a trustworthy way to judge outcomes, and a rule for deciding whether an apparent gain is real.

THE ANSWER: Current results support a qualified yes. In SPADE, a Qwen3 policy writes executable environments, plays them, and updates its designer and solver roles from checked outcomes; its best games checkpoint beat its strongest fixed-environment baseline across eight held-out benchmarks, under matched training iterations rather than matched rollouts or FLOPs.[1] Whether it improves or collapses turns on the controls, and the report specifies them: fixed task supply and designer weights, a swapped verifier, multiple seeds, predeclared holdouts. If the project cannot run those controls, withhold the co-evolution claim and stop before deployment. The evidence remains a short, mostly author-reported test; independent reproduction and sustained truthful frontier expansion remain unestablished.

Receipts and full analysis →

How a Reasoner Learns When to Stop

Sat, 22 Aug 2026 00:00:00 GMT

THE QUESTION: Why can more reasoning tokens or search depth make a model's delivered answer worse, and what full-parameter post-training procedure makes inference-budget use adaptive and reliable?

WHY IT MATTERS: Inference budgets already change model quality, latency, and cost, and a deployed system must decide whether to keep its current answer or spend compute on another turn, candidate, or verification step.

THE ANSWER: More compute helps only when it preserves a choice among actions. On SVR's GSM8K trajectories, a correct answer appeared within ten turns for 0.925 of examples, but forcing the same policy to return its tenth answer reduced accuracy to 0.736.[1] The directly supported response is narrow: for an existing post-trained reasoner on tasks with exact outcome checks, train full-parameter, fixed-horizon RLVR and let verdict and confidence control retention only at inference — 0.563 on the All-7 macro-average against 0.450 for the best forced fixed-turn point, averaging 2.99 turns, with premature-stop error still 29.9%.[1] The cited studies do not yet validate an end-to-end full-parameter adaptive-compute pipeline across math, code, factual, and open-ended reasoning.

Receipts and full analysis →

How to Full-Train a Distilled Reasoning Model Without Making It Worse

Fri, 21 Aug 2026 00:00:00 GMT

THE QUESTION: How should a practitioner full-train an existing small reasoning model, why can verified, polished teacher traces make the student worse, and what fixes that at its source?

WHY IT MATTERS: Full-parameter post-training can change every capability in a checkpoint. A successful target-task run can therefore damage behaviors the deployment still needs, and a clean dataset with a passing final answer provides no audit of that trade.

THE ANSWER: Treat the trace as candidate data and the student checkpoint as the object being tested. Freeze the target, out-of-domain, completion, efficiency, sampling, and retention measurements first, then build a provenance-tracked pool with outcomes and intermediate steps verified separately. Run conservative full-parameter cross-entropy SFT first; it is both the default and a valid endpoint. Whether to go further depends on what that checkpoint shows, and the report gives the entry condition, matched control, and stop rule for each branch, with thresholds predeclared per project. It remains unknown whether any student-generated branch reliably beats retention-safe SFT in a fully matched comparison beyond math and code.

Receipts and full analysis →

What Stops Working Before the Million-Token Window Runs Out

Thu, 20 Aug 2026 00:00:00 GMT

THE QUESTION: When a model card advertises a million-token context window, how much actually functions as usable context — and what quietly stops working before the window runs out?

WHY IT MATTERS: Every frontier vendor now advertises a million tokens or more, and the number is quoted in real architecture decisions — retiring retrieval pipelines, sizing agent memory, paying for long-context tiers.

THE ANSWER: The advertised figure is a serving and billing quantity, not a measurement, and no published work identifies what fraction of any window is usable: an effective length needs a threshold on a score-versus-length curve, and none exists at 1M for the latest cohort. The record splits by task family, and the report walks them: literal single-fact lookup runs near ceiling at a full million [10, 14]; non-literal retrieval has thresholded evidence only through GPT-4.1, 16K effective against 1M claimed [3]. The working-set budgets shipped systems run — a quarter to a third of the advertised figure — are engineering choices, not measured optima; validate them on your own sessions. Above 256K most evidence is vendor-run.

Receipts and full analysis →

LoRA vs. Full Fine-Tuning in 2026: What Survived the Re-Measurement

Thu, 20 Aug 2026 00:00:00 GMT

THE QUESTION: Under what conditions does LoRA match full fine-tuning in 2026 — and which parts of the 2024 answer, "learns less, forgets less" and the "illusion of equivalence," are still the answer?

WHY IT MATTERS: Practitioners still choose between LoRA and full fine-tuning on decision rules anchored in two 2024 studies, while the models and training regimes those rules were measured on have since been replaced — and RL post-training appeared in none of them.

THE ANSWER: In supervised fine-tuning the canon survives, re-parameterized: LoRA matches when the dataset fits within the adapter's information capacity [4], when adapters sit on all weight matrices rather than attention only, and when the learning rate is set about 10× the full-fine-tuning optimum [3]. Placement and learning rate, not rank, were the real variables [3, 4]. In reinforcement-learning post-training the answer is new: LoRA matches down to rank 1 [3, 7, 27], multiply reproduced but not peer-review-attested, though the floor is disputed — VERL advises at least rank 32 [9]. "Forgets less" and the "illusion of equivalence" both survive, but only in supervised training, where they were measured [1, 15].

Receipts and full analysis →

The Skill Is Real, the Ritual Is Not Yet Measured

Thu, 20 Aug 2026 00:00:00 GMT

THE QUESTION: Skills frameworks like superpowers load process disciplines into a coding agent's harness. Do they improve what an agent can do, or teach it to perform the shape of a process — and does that depend on model capability?

WHY IT MATTERS: The skills mechanism is a vendor-supported feature of the major coding-agent harnesses, the frameworks built on it are spreading fast, and teams are deciding now whether to standardize on one.

THE ANSWER: Skill loading as a mechanism is real: curated skills lift mean pass rates from 33.9% to 50.5%, largest where the skill carries procedure the model lacks [9]. The process disciplines themselves have no such record: no published same-model ablation shows a statistically supported task-outcome improvement from a superpowers-class framework, on any tier. The closest measurements lean the other way: superpowers indistinguishable from baseline on process quality [13], one discipline delivered as pure instruction making a small model worse [14]. The capability gradient is not one gradient but an inverted-U, which the report walks; how process-discipline layers behave across it remains, precisely, unmeasured.

Receipts and full analysis →

Hard, Verifiable Agent Data: Download for Code, Build for the Rest

Tue, 18 Aug 2026 00:00:00 GMT

THE QUESTION: A team that has deployed a strong agent across code, browsers, and desktops wants to distill it into cheaper models. Which strategies generate the hard, verifiable training data, and what can it use today?

WHY IT MATTERS: Such a team holds the teacher, not the training material — tasks hard enough to be worth learning from and checkable enough to reward against — and whatever it cannot download, it must build.

THE ANSWER: For code the shelf is download-ready — generators, executable per-task verifiers, and permissively licensed trajectory corpora exist now [15, 17, 18] — and verified command-line environments ship with executable checkers attached [36, 38]. Elsewhere the shelf thins: the largest web-task artifact verifies by natural-language criteria and LLM judgment rather than execution [21], the classic GUI trajectory pipelines filter with vision-model evaluators [22]; the environments are yours to build. What is unevenly solved is verification — mined task pools have failed independent audits at measured rates [6, 7, 8], and any checker facing an RL-trained student is an attack surface: shipping agents have already been caught gaming scoring [10, 11]. The durable asset in every domain is the generator-verifier pair, not any dataset it emits.

Receipts and full analysis →