Prerequisites IV: Two Spaces of Improvement
There are two places an AI system's performance can live, and therefore two places improvement can happen. Confusing them produces muddled strategy: teams argue about fine-tuning versus prompting as if they were substitutes, when they operate on different substrates, at different costs, with different risks. This page separates the two spaces, then explains why all three November papers live in the second one, and how the two compose in a serious program.
Weight space: the model itself changes
In weight space, improvement means changing the model's parameters. The canonical modern methods:
- Supervised fine-tuning and preference tuning shape behavior on curated examples and comparisons.
- GRPO (Group Relative Policy Optimization, from DeepSeekMath, 2024) is the standard reinforcement learning optimizer of the current era: for each prompt, sample a group of outputs, score them, and update the policy toward the above-average ones, using the group's own mean as the baseline. No separate critic network; the group is its own control group.
- RLVR (Reinforcement Learning with Verifiable Rewards, named by Tulu 3) supplies the reward: a mechanical verifier's binary verdict. See Verifiers and Graders.
Weight-space improvement has distinctive properties. It is baked in: the improvement travels with the model, needs no runtime machinery, and works even where no harness can be customized. It is expensive per iteration: training runs, data pipelines, evaluation, and the risk of degrading capabilities you were not measuring. It is opaque: when behavior changes, the reason is distributed across billions of parameters; there is no diff to review. And it is general: a capability gained in weights tends to apply everywhere the model goes, for better and for worse.
LoRA adapters are the lightweight end of weight space: small trainable matrices layered onto a frozen base model, cheap to train and cheap to swap. One practitioner pattern makes them relevant here: train a LoRA adapter inside the exact harness the model will run in, with reinforcement learning from verifier outcomes and rollouts executed by the unmodified harness. The adapter then learns the harness's specific demands: its tool formats, its failure modes, its turn economy. The weights become harness-shaped on purpose. That is the two spaces deliberately fused, and it is the bridge this page ends on. (Doctrine: this is a general practitioner pattern, not a result from the November papers, and not a claim about any particular program.)
Artifact space: everything around the model changes
In the second space, the model is frozen and improvement happens to its surroundings:
- Prompts and instructions: the text the model is given.
- Harness code: context assembly, control flow, tool interfaces, verification hooks (The Loop).
- Memory: structured records the system reads at run time, from Reflexion's reflection buffer in 2023 to Paper 2's edit ledger.
- The measuring instruments themselves: benchmark definitions, graders, thresholds, the subject of Paper 3.
Artifact-space improvement has the opposite profile. Iterations are cheap and fast: edit a file, rerun the evaluation. Changes are legible: every improvement is a diff a human can read, revert, or argue with. The substrate is fragile in a specific way: an artifact tuned to one model or task set may not survive a change in either, which is the negative transfer problem Paper 2 measures. And the gains are real but bounded by the model: no harness edit supplies knowledge the model does not have. Paper 1's diagnosis stage is required to abstain when failures trace to a knowledge gap rather than a harness fault, because no edit to the surroundings fixes a missing capability.
Why the three papers all live in the second space
Three reasons, in ascending order of importance.
Access. Frontier models are served frozen. Their users, including large companies' internal teams, can change everything around them and nothing inside them. Harness space is the space practitioners actually control.
Iteration economics. The strongest published comparison is GEPA (2025, see the reading path, abstract-level): reflective evolution of prompts outperformed GRPO by about 6% on average across six tasks while using up to 35 times fewer rollouts. Paper 1 adds a deployment-side data point (paper-reported, one run per method): its gated harness ran at 46.7k tokens per task against a full rewrite's 451.6k, at similar held-out accuracy (40.0% versus 42.5% combined over 40 tasks, a one-task gap; the gated harness led in-domain and trailed on the cross-repository set). That is runtime cost. The gated search itself used more rollouts (100 versus 24), which the paper says the runtime savings repay in under 100 deployment tasks. When iterations are cheap, search becomes practical, and search is what these loops are.
Legibility under change. Models turn over fast. When the model changes, weight-space work tied to the old model is stranded, and artifact-space work is stranded differently: the artifact survives, but its fitness may not. Paper 2's central finding is about that second case. When the setting changes, carrying the evolved harness together with its record (which edits were made, why, with what predicted and observed effects) beat carrying the harness alone in every setting it tested. Paper 2 did not compare against weight-space methods; the claim is about artifact plus record versus artifact alone. Artifact space is where a program can keep its lessons in a form that outlives any particular model. That is a strategic property, not just an engineering convenience.
How the two spaces compose
They are complements with a natural division of labor, and the composition rule falls out of their failure profiles. What follows is the site's doctrine, inferred from the papers' findings rather than tested by them:
- Structural, recurring failures belong to the harness. Premature exits, context exhaustion, tool misuse, unverified completion: fix them in code, once, legibly. Do not train a model to stop doing something a hook can simply forbid.
- Recurring behavioral failures belong in the weights. If the model reliably reasons badly in a situation the harness cannot detect mechanically, that is training data. The harness's job is to find these failures at scale and record them; the training run's job is to fix them.
- The harness is the discovery instrument for the weights. A gated improvement loop with honest records is, among other things, a failure-mining machine. Its diagnosis clusters and rejected-edit records are a curated map of where the model is weak, which is the scarcest input to any training program.
- Train into a frozen harness, then re-adapt it. The LoRA-inside-the-harness pattern works because the harness is held still during training, so the adapter learns a stable target. Then the harness loop resumes, now around a better model, and its records say which old edits the new model has made unnecessary. Paper 1 lists model and harness co-evolution as unsolved future work; this staged alternation (adapt harness, train weights into it, re-adapt) is a practical approximation, not a result. Its cadence and its returns are unmeasured.
The November papers are sometimes framed as an argument that harness space beats weight space. Read carefully, they support a narrower and more useful claim: harness space is underpriced. It is cheaper to search, legible to audit, and, with records kept properly, more portable across model generations than an artifact carried alone. A serious program budgets for both, sequences them as above, and keeps the verifier, the one component both spaces share, frozen and honest. The synthesis turns that into doctrine.