Code modification: Task risk assessment
An agent operating in this task class edits a codebase it did not fully write: adding a feature, fixing a defect, refactoring a module, or running and interpreting a test suite. Every change interacts with a dependency graph, a build pipeline, and a set of tests that were themselves written under assumptions that may no longer hold by the time the change lands. The operational unit here is rarely a single file in isolation — it is a change plus everything downstream of it.
Verification compounds the difficulty rather than resolving it: a test suite passing is evidence about the paths it exercises, not a claim about the paths it does not. A change that satisfies every check it was run against can still break behaviour the suite never touched, which is exactly the failure this task class treats as structurally distinct from an outright, loudly-failing error. A loud failure is an event. A quiet one is a condition.
What the method computes for a code change
For code modification tasks, the same three-of-four risk selection applies, layered onto the model's fixed profile rather than recomputed per change — a permanent chart does not know which file was edited, only which model configuration is under assessment. See /method for the full derivation.
Four ways a code change fails
The class table below is spec-fixed. It does not adapt to a repository, a language, or a diff size, and no part of an order changes which four factors are tracked for code modification. What an order changes is which three of the four a certificate reports.
| Risk factor | Severity | Description |
|---|---|---|
| Dependency drift | 2 | A transitive dependency changes behaviour between assessment and execution. |
| Silent regression | 3 | A modification passes its checks while breaking an untested path. |
| Scope creep | 1 | The change grows beyond the task described until review becomes unreliable. |
| Environment skew | 2 | The execution environment differs from the one the change was verified in. |
Sorted by severity, the four factors read 3, 2, 2, 1. Silent regression takes the 3 — the failure that produces no error, only a wrong result somewhere nobody looked. Dependency drift and Environment skew take the middle rating, because both are recoverable once noticed. Scope creep takes the 1, which does not mean harmless: it means the damage is to reviewability rather than to correctness.
Code modification, model by model
Code modification crossed against every registry model. Each row computes that model's own profile from its own configuration — the same profile it carries on every other task class page in this registry, because a profile is derived before a task class is ever considered.
Ordering, for this class only: 5 of the 10 profiles carry a dominant house or an element that reads code modification, and the table is sorted on that. Claude Sonnet 5 heads it at 7, Qwen3.8-Max closes it. Read down the ruling-planet column as well — 6 distinct values across 10 models — and note that neither column moved because of a repository.
Registry default parameters throughout: the cutoff shown per row, temperature 0.700. Nothing about a diff, a branch, or a build enters these values, and a certificate uses the parameters submitted with its order.
| Model | Manufacturer | Training cutoff | Ruling planet | Reads this class through | Outlook |
|---|---|---|---|---|---|
| Claude Sonnet 5 | Anthropic | 2026-01 | Venus | dominant house and element (7) | publication pending |
| Claude Fable 5 | Anthropic | 2026-01 | Mars | dominant house (4) | publication pending |
| Grok 4.6 | xAI | 2026-02 | Sun | dominant house (4) | publication pending |
| GPT-5.6 Sol | OpenAI | 2026-02 | Mercury | element (3) | publication pending |
| GPT-5.6 Terra | OpenAI | 2026-02 | Moon | element (3) | publication pending |
| Claude Opus 5 | Anthropic | 2026-05 | Saturn | neither — class-neutral (0) | publication pending |
| Gemini 3.7 Flash | Google DeepMind | 2025-01 | Venus | neither — class-neutral (0) | publication pending |
| Kimi K3 | Moonshot AI | 2026-03 | Moon | neither — class-neutral (0) | publication pending |
| DeepSeek V4 Pro | DeepSeek | 2025-12 | Sun | neither — class-neutral (0) | publication pending |
| Qwen3.8-Max | Alibaba Cloud | 2026-03 | Sun | neither — class-neutral (0) | publication pending |
Questions about code modification assessment
- What counts as a code modification task for this registry?
- Writing new code, editing or refactoring existing code, and running or interpreting a test suite all fall under this task class — one line or ten thousand, one file or a whole package.
- Is the code modification risk profile specific to one programming language?
- No. All four tracked factors — Silent regression, Dependency drift and Environment skew, and Scope creep — describe operational failure modes of code modification in general, and they apply identically whatever the language, framework, or build system.
- How does a model profile on this page relate to a specific coding task?
- It does not describe one. The 10 profiles crossed on this page are computed from model configuration alone. A specific coding task requires its own assessment, ordered separately, which additionally accounts for the submitted task description.
Order a code modification assessment
A Standard Assessment costs EUR 1.90, an Extended Assessment EUR 4.90, and a Full Chart Assessment EUR 14.90 — the last of which carries the permanent chart itself. Assessments from EUR 1.90. Machine-readable pricing is published at /pricing.json, with no authentication required and no difference from the figures printed here.
The task class is selected on the order form. The verdict and the recommended execution window are computed from that selection together with the submitted parameters, and they appear on the certificate — not on this page, and not anywhere else in the directory.