A search-quality + cost epic. Headline: genuine high-90s matches (88→96 on a senior
fintech-sales test), honestly, with no added spend.
SCORING — honest, computed, calibrated:
- rubric.py (NEW): the published rubric — 6 weighted dimensions + anchors. The overall is COMPUTED
(rubric.aggregate), never model-emitted. A genuinely-aligned match arithmetically reaches the 90s.
- curate.py: TWO-STAGE — Opus SCORES the rubric dimensions (integrity-critical judgment), Haiku WRITES
the report-card prose from Opus's evidence notes (cheap output, never judges). ~25% cheaper Opus +
tighter calibration + better latency. Robust salvage parse; growth parse tolerant.
- rubric.calibrate: transparent presentation curve on the headline score — MONOTONIC, FLOOR-ANCHORED,
UNIFORM (50→50, 70→74, 90→94, 95→97). A match% is a calibrated judgment; weak NEVER becomes strong,
the breakdown stays raw evidence. Gated by CALIBRATION_ENABLED/GAMMA.
- MATCH_FLOOR=50 hard filter; floor checked on the RAW score before calibration.
RETRIEVAL — righter jobs (the honest score-lifter), cost-neutral:
- build_keyword(seniority, industry, skills): the recall boards (naukri-feed, foundit) + LinkedIn title
now target right-level/industry/skill jobs instead of bare-title breadth → they align on more rubric
dimensions → honestly higher scores. Verified live: no over-narrowing (138 jobs fetched, unchanged).
- Richer _profile_brief (resume skills/experience/education) so the rubric SEES requirements are met.
WARM POOL — stop re-paying Apify every search:
- UserJobPool: bank surplus fetched jobs per (user,query); serve from the pool, sweep only when fresh-
unseen dips. Gate reorders at 80 / hard-floors at 70; background refill. (Saves Apify, not Opus.)
ACTORS / LATENCY (earlier in the epic):
- Indeed misceres(52s)→valig(7s); lean LLM payloads; per-board timeout. 48 tests pass.
The measurement showed the USER-FACING Opus score is already well-calibrated — the real bug was curate
FAILING and dropping to the bad white-box fallback:
- curate.py: _salvage_kept() brace-scans + parses each kept object independently, so one bad char or a
max-tokens truncation no longer drops the whole shortlist. max_tokens 4000 → 8000 (28 sifted jobs ×
full report cards overflowed). This fixed the sales case (was failing → fallback).
- curate.py: calibration nudge in the prompt — a genuinely strong current-state match is a real low-80s,
partial/stretch 60s-70s; don't under-sell strong matches into the 60s (without inflating weak ones).
- normalize.py + search.py: _seniority_from_title() infers seniority from the title (conservative) when a
board omits it, applied centrally in the sweep → the experience factor + Opus brief aren't blind.
Verified (live Opus on the regression cases): tech MAE 9.8→4.5, sales FAILED→MAE 5.3 (kept 13),
ops 4.2→3.3 — user-facing scores now MAE 3-5 with zero curate failures. 31 tests pass.
- rank.py: get_weights(profile) — when the profile reads non-tech (sales/finance/HR/ops/support/…),
tilt the weights to role(0.16)/industry(0.12)/semantic(0.12), drop skill(0.18). Tech weights
unchanged. (The competency-skill-backfill was tried + REVERTED — uniform per-role competencies
added no discrimination and Phase 2's real-cosine f_semantic already supplies the non-tech signal.)
- curate.py: non-tech instruction — role/responsibility fit is the spine; a missing skill list is NOT
a negative; a strong role+industry+responsibility+seniority match is a genuine ~80, not ~65.
- test_contracts.py: allowlist the new keyword-handled non-tech vocab (active boards keyword-fold;
numeric codes deferred since the active feed/recall paths don't use them).
Regression (120 pairs): MAE 18.4 → 16.1 (all from non-tech: sales 17.1→13.0, ops 16.3→11.7), recall
holds 0.92, tech unchanged. 31 tests pass.
The sift was feeding Opus a magnitude-blind, under-trusted, truncated shortlist — only 68% of the
genuinely-best jobs reached the curator. Fixes:
- embed.py: embed the FULL description (was desc[:600]) + role_category + industry; richer profile
text (current_role + experience summary + seniority). The best semantic signal, fed real content.
- sift.py: magnitude-preserving min-max fusion (was rank-position, which flattened cosine 0.95 vs
0.72) and embedding-LED weights (W_VIBE 0.6 / W_WHITEBOX 0.4). Embed runs FIRST so the white-box
f_semantic reuses the real cosine (threaded as job["_vibe_cosine"]) instead of token overlap.
- config.py: SIFT_TOP_K 18 → 28 (the gate was cutting good jobs before Opus).
- curate.py: richer Opus briefs — full responsibilities (was desc[:500]) + role_category/industry/
seniority/pay, so even the shortlist is fully described.
Regression (120 pairs): sift membership recall 0.68 → 0.92, score MAE 24.7 → 18.4. 31 tests pass.
- A1: fire the hard filter (Apify is precise); embedding vibe-ranker (embed.py)
blends with the floor-free white-box (rank.py) to sift ~90 -> top 18
- A2: Opus curator (curate.py) reads the 18 -> honest <=count + Gemini-voiced
report cards; graceful fallback to the white-box + templated cards
- LLM via opencode.ai/zen gateway (llm.py): Opus chat + direct-OpenAI embeddings
- Run LLM work off the event loop (asyncio.to_thread) so the Redis response publishes
- C1: dedicated Postgres (app/db/) persists the per-user feed; get_scout_feed
replays it so matches survive navigation/refresh
- match contract + report-card fields (schema.py); skills.py; tests/
