Source of truth: resources/css/app.css (HSL design tokens). HEX values computed from HSL.
Primary brand font: Figtree (300–700).
Landing (/new) pairing: Newsreader (serif headings), Hanken Grotesk (body), IBM Plex Mono (labels/mono).
| Token | HEX | HSL |
|---|
Sample of 25 changed picks + 5 unchanged for a sanity check. The full pilot output is in vacancy_data.minerva_reclassify_results (1000 rows from the 2026-06-17 run).
Headline: 985/1000 ok, 15 sentinel (invalid/too_broad jobtitles), 0 errors. Of the 985 picks: 420 kept the current minerva, 565 changed (~57%).
For each row mark new ✓ / current ✓ / neither — should be X in the Verdict column.
A = current Stefan approach (2x /api/search no-context + gpt-5.4-nano picks 1 with context) E3 = single /api/search call with context parameter (newly deployed)
Mark A, E3, or neither in the "Best" column. Add notes when both are off and the right answer is something else.
| # | language | jobtitle_ai | jobtitle_cleaned | A pick | E3 pick | current | Best | Notes |
|---|---|---|---|---|---|---|---|---|
| 1 | nl | Wijkcoach | Neighborhood Coach Allowances | 9808 — Welzijnscoach | 9934 — Wijkcoördinator | 9934 |
Calibrated against actual project costs from 2026-04-02 → 2026-05-05 (91 real projects with ≥50 scored candidates), with CrustData realtime discount (−70%) applied and CrustData costs imputed onto LinkedIn-sourced projects so the reference is provider-agnostic.
0-2 agents: $0.07 × scored_candidates + $0.15 × companies
3+ agents: $0.20 × scored_candidates + $0.15 × companies
Period: 2026-04-02 (cost tracking rollout) → 2026-05-05 Scope: projects created post-rollout with ≥50 candidates that ran through scoring (real, working projects, not test stubs) Adjustments applied:
Goal: give a defensible price reference for client conversations — what does a project of size X actually cost us.
The prompt used by AI to decide whether a candidate is relevant to a role. Built in app/Prompts/BatchCandidateEvaluationPrompt.php. Candidates are evaluated in batches of 25 per request.
Prompt texts are shown verbatim, exactly as sent to the model. {...} marks interpolated values.
Endpoint: POST /screener/person/search
Date: 2026-04-21
Reproducible: yes, every call
We pass post_processing.strict_title_and_company_match: true expecting your API to return only profiles where a role at the target company also matches the title filter. Instead, the API returns profiles where:
Project: Learco Personal, Project 1409 (Graphisoft CEO search)
Date: 2026-04-04
Candidates: 359 with full LinkedIn profiles
Scoring factors: 6, all on 0-5 scale (total max: 30)
Runs per experiment: 5
Concurrency: 32 parallel requests (10 for Anthropic due to latency)
Prompt: Identical across all experiments (exported via experiment:export-prompts using production CandidateScoringPromptService)
Dealroom enrichment: Skipped (to eliminate external API nondeterminism)