Quality debt
Where the system produces output that is shallower than it should be.
Not a bug list — everything here works. It exists because a thin implementation and a finished one look identical from the outside, and the risk this project runs is a polished portal whose output is quietly poor with nobody able to say which stage caused it.
Rules for this file, from rules/depth.md:
- Anything shipped below its proper depth is recorded here the same day
- Each entry names what will actually be bad about the output, not just "could be better"
- An entry leaves only when the real version ships, never because it stopped being mentioned
Open
| # | Capability | Shipped as | Should be | What is actually bad |
|---|---|---|---|---|
| 4 | Layout library | Two hand-written satori specs | Higgsfield render, or 6–8 designer layouts | Only matters for the composited path. Superseded in principle by higgsfield-as-art-director.md, which is built but not yet the default. |
| 7 | Insight and angle critics | None — only copy and scripts are judged | Adversarial check at each stage | A weak insight travels three expensive stages before anything challenges it. Preconditions now catch the worst case, but not a plausible-and-wrong idea. |
| 8 | Format selection | The angle suggests, nothing chooses | Agent selecting on semantics and performance | format_fit is an advisory list from the angle worker. Nothing weighs it against what has actually worked, because no performance data exists. |
| 10 | Video production | Scripts queue a task | Scripts → Higgsfield video presets → clips | A script is a brief, not a deliverable. Queue-for-production creates a task; nothing renders. |
| 11 | Atlas Brain depth | One-pass answer over aggregates | Multi-dimension analysis once performance, finance and utilisation exist | Honest about its dark dimensions, but the basis is thin — counts and spend, no trends, no per-client margins. |
| 12 | Content converters | Images only | Video, pdf and webpage converters into extracted_text | Non-image library items sit pending untagged. |
Closed
| # | Was | Now |
|---|---|---|
| 7 | No insight critic | Runs in the insights stage; a named competitor is a kill |
| 9 | Hardcoded Higgsfield shortlist | Live catalogue over MCP, cached, with fallback |
| 1 | Copy generation was one pass | Ecosystem: write wide, deterministic slop gate, adversarial critic that can reject everything |
| 2 | Scene prompt was a static field per angle | direct-visual decides what each ad shows and why, with must_not_show required |
| 3, 5 | Angles written by hand | define-angles proposes from an insight, differing by framework, awareness and proof |
| 6 | Scripting not built | write-script plus a script-specific critic and deterministic timing checks |
Attribution status
Live and verified. Every generated artefact carries produced_by — run, worker, skill, skill version and model — and skills are versioned by content hash, so an edit is automatically a new version. Verified on dev: every feature written by an ingest carries its stamp, so a complaint about a specific fact traces to the exact methodology version that produced it.