Seeds of the Throne

type: qa
status: active
updated: 2026-09-06

Shared Skill Tests

Test method

Each skill was initialized with the standard skill scaffold, checked with the official structural validator, and forward-tested by a fresh agent on a representative read-only task. Tests evaluated workflow behavior rather than agreement with a predetermined creative answer.

The 2026-08-18 modular prose expansion additionally used deterministic linter regression cases and representative non-canon prose smoke tests. Those checks were recorded separately because a full fresh-agent run across the then-current ten-case prose benchmark suite had not yet been completed.

The 2026-08-19 controlled-variance integration added a dedicated causal-variance module, a scene-level attention map, an expression-variance control, rubric and revision hooks, behavioral benchmark B11, and descriptive linter geometry. The linter deliberately does not warn or score prose for low or high numerical variance; statistics remain prompts for human inspection.

Results

2026-08-18 prose regression details

Deterministic cases now cover:

1. clean prose stays unflagged;
2. em dashes fail;
3. repeated contrast templates warn;
4. repeated three-word sentence openings warn;
5. one-sentence narrative paragraph clusters warn;
6. normal dialogue paragraphing does not trigger the narrative-cluster warning.
7. repeated It wasn’t X. It was Y. templates using a typographic apostrophe warn;
8. scattered one-sentence narrative paragraphs do not trigger a cluster warning when longer paragraphs break the run.
9. sentence-length summaries report descriptive spread without creating a variance flag;
10. uniform sentence lengths do not automatically fail or warn merely for being uniform.

The prose benchmark reference now defines eleven broader behavioral cases covering close-POV anomaly, political grotesque, archive reconstruction, technology as infrastructure, unequal-knowledge suspense, restraint on already-good prose, anti-AI pattern diagnosis, unresolved canon, continuation inheritance, exposition under conflict, and controlled variance without voice drift.

Structural validation

All seven original skill folders passed the official skill validator after initialization. Their metadata contains valid names and descriptions, their instructions are under 500 lines, and detailed checklists are loaded progressively from one-level references/ files where needed.

The 2026-08-18 write-seeds-prose expansion preserved the same skill folder and frontmatter shape. After desktop reconciliation, the official structural validator passed and the eight deterministic Python regression cases passed under standard unittest discovery.

The 2026-08-19 controlled-variance refinement also passes the official structural validator. Ten deterministic Python regression cases pass under standard unittest discovery, including two safeguards that keep variance metrics descriptive rather than prescriptive. The full fresh-agent suite now contains eleven behavioral benchmarks and remains pending when an authorized external-agent QA environment is available.

Follow-up

Use the skills on real work and revise them when repeated friction appears. A successful isolated test demonstrates that the workflow is legible; it does not make the skill permanently complete. For prose specifically, promote repeated author feedback through the observed -> proposed -> tested -> active lifecycle rather than turning one reaction into a global style rule.

2026-09-06 Compound Writing adaptation experiment

The author approved four narrow additions to write-seeds-prose: a Tracks pass, pressure-turn placement diagnosis, visible development-edit versus line-edit selection, and explicit preference routing.

Initial restraint checks used two existing forms:

1. The public atlas homepage correctly keeps its development-status and uncertainty language because that framing serves the page's audience. The Tracks pass should not erase it as machine residue.
2. The Compound Writing assessment contains headings, option lists, implementation notes, and decision scaffolding appropriate to a session record. The pass should remove those only if someone incorrectly copies the note into manuscript or public narrative prose.

These checks establish the intended boundary: detect residue according to the destination, rather than banning procedural language everywhere. Behavioral benchmarks B12 and B13 now preserve that distinction and prevent pressure-turn diagnosis from forcing premature disclosure. A future real-prose test should record whether the new passes find a consequential issue that the existing structural and anti-AI passes miss.