Spaces:
Sleeping
Sleeping
Commit History
-
LLM overhaul: task-specific prompts, junk-claim filter, honest telemetry
Fixes the five audited failures from live E2E: - extraction v2: exclusion rules + few-shot + kind tags (project/skill/metric/ experience); deterministic sanitize_claims net drops GPA/contact/years-of- experience junk and splits compound skill lists (was: 3 of 7 claims junk, 5 skills in one claim) - judge v2: confidence rubric (was uniform 0.50), evidence headers with langs/stars/fork from metadata, runs on Reasoning: medium, optional TINKER_JUDGE_MODEL override - metric claims skip the judge -> honest NEI, budget freed for real claims - summarize v2: structured brief (VERIFIED STRENGTHS / UNVERIFIED / BOTTOM LINE), repo citations, <=120 words - robust complete(): finish_reason checked, truncation retried once at 2x budget, response_format json_object probed and used (endpoint supports it), parse_failures + truncations counted per candidate and shown in Run health (api_successes measured HTTP success, not usable output - the metric that hid silent template fallbacks)…ea9cb18 -
UI overhaul: dark teal theme, editable ranking, tabbed detail, sample demo mode
- Theme: dark blue-charcoal + verification-teal accent, semantic badge hues (green=SUPPORTED, yellow=NEI, red=REFUTED/high, orange=warn), Inter font - Landing: hero + value-prop cards + one-click "Try sample data" (3 fictional candidates covering strong/inflated/no-github archetypes, zero API calls) - Ranking: st.data_editor with inline shortlist SelectboxColumn, ProgressColumn final score, pinned candidate column; Tinker banners folded into a Run-health popover; downloads into an Export popover; 4 bordered KPI metrics - Detail: header card (skills chips + score meter bars), segmented-control decision, tabs (Overview / Claims & evidence / Authenticity / Interview kit); raw JSON blobs replaced with a field grid + Raw-data popover; claim cards sorted REFUTED-first with verdict filter pills - Compare: bordered cards, per-metric winner highlighting, verdict/fraud chips - Methodology: tabbed, badge vocabulary matching the app - New AppTest smoke tests execute all three pages h…a36589b -
Fix 12 audited bugs: PII leaks in blind exports, counter inflation, scoring perversity
- A1: per-report LLM call counts are now per-candidate deltas, not cumulative - A2: blind mode now redacts JSON/CSV/PDF/audit exports (was UI-only) - A3: NOT_ENOUGH_INFO contributes a flat 0.35 to verification - confidence in "we can't verify" no longer raises the score - A4: commit-clustering fallback only fires when the sampled commits are the whole history (weekend bursts on active repos no longer flagged) - A5: star-anomaly gated on deep mode; RepoSnapshot.watchers -> subscribers (shallow watchers_count just mirrors stars and was silently ratio=1) - A6: shortlist keys disambiguated by email (same-filename collision) - A7: US 4.0-scale GPAs normalized to /10 (3.8/4.0 was scored as 3.8/10) - A8: skills-line tokens filtered against tech vocab + JD (junk no longer dilutes the JD-fit denominator) - A9: heuristic SUPPORTED capped at 0.65 confidence and labeled keyword-level; leftover LLM budget now double-checks keyword-SUPPORTED claims - A10: circuit breaker only hard-disables…fc45a20 -
Grifter Filter v0.2: evidence-verified resume screening with fraud detection
Full rewrite of the pipeline into a deployable product: - Tinker integration via the documented OpenAI-compatible endpoint (openai/gpt-oss-20b, reasoning separated server-side, ~$0.003/candidate), with circuit breaker and full heuristic fallback - 7-signal fraud engine (contributor share, commit clustering, star anomalies, identity mismatch, language cross-check, account-age gap, AI-authorship markers) over a cached GitHubSnapshot - FEVER-discipline claim judging: REFUTED requires contradicting evidence, evidence_index resolved locally to kill hallucinated links - Recruiter UX: per-candidate PDF reports, targeted interview questions, side-by-side Compare page, shortlist tagging, evidence deep-links - Trust layer: blind-screening mode, published Methodology page, audit-log export, no-auto-reject disclaimers - PRD.md + ARCHITECTURE.md, HF Spaces front-matter, 29 passing tests Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>b78196e