Every score now ships with the evidence that produced it.

Numbers come with receipts.

"87/100 — no visible rubric. Power users will smell this."

Every score now ships a signals array. Each signal has a name (snake_case), a value 0–10, and an evidence field with the concrete phrase from the input that drove the score. Agents cite the phrase, not paraphrase the number.

"signals": [
  {
    "name": "specificity",
    "value": 9,
    "evidence": "uses 4 concrete numbers (24, $58K, $100K, 12%)"
  },
  {
    "name": "credibility_signal",
    "value": 8,
    "evidence": "'no side hustle' attaches a verifiable constraint"
  },
  ...
]

Provenance is a first-class field.

"Are those view counts real and recent? If they're plausible-but-fabricated, the whole output is sand."

Every response includes as_of + a provenance block. analyze_account exposes data source, fetched_at, video_post_dates[]. trend_pulse exposes per-entry sources, signal_window, signal_strength + top-level cache_age_hours and cache_status.

"provenance": {
  "data_sources": [
    "Google Trends RSS (regional + global)",
    "YouTube Data API (velocity)",
    "Reddit hot (cross-subreddit topic surfacing)",
    "Internal viral-video corpus (100K+ videos)"
  ],
  "cache_age_hours": 3,
  "cache_status": "fresh",
  "rubric_version": "trend_pulse.v1"
}

Scores discriminate. No more 82/85/87 clusters.

"Three different scripts, similar scores, similar reasoning. The scoring model isn't discriminating enough to be useful as a decision tool."

Six calibration anchors at 10 / 30 / 50 / 70 / 85 / 95 are now wired into the system prompt with example hooks per anchor. Temperature dropped from 0.5 to 0.2. The model is required to name the closest anchor in its why field. predict_virality adds a calibration_check field that flags when a score is overstated.

"calibration_check": {
  "anchor": "80",
  "agreement": "agrees — strong hook + scene structure + named CTA,
                missing a second pattern interrupt to hit 95"
}

Voice DNA as math, not vibes.

"Voice DNA labels are vibes, not systems. 'Excited Discovery Evangelist' sounds good but isn't reproducible."

match_voice now returns a deterministic voice_metrics block alongside the qualitative profile: type-token ratio (moving-window 100), filler rate per 100 words (18 filler tokens matched), avg sentence length, and top recurring 2-3-grams with counts. Reproducible across runs, citable as numbers.

"voice_metrics": {
  "vocab_diversity_ttr": 0.62,
  "filler_rate_per_100_words": 1.4,
  "avg_sentence_length_words": 8.2,
  "signature_phrases": [
    { "phrase": "real talk", "count": 4 },
    { "phrase": "you need to", "count": 3 }
  ]
}

Every tool says what would have failed.

"No counterfactual. A real intelligence layer would say 'here's the version that scored 45 and why.'"

score_hook, predict_virality, and analyze_account all return a would_fail_because field. One sentence. Names the closest version that drops 20+ points and the reason. Tells agents what to avoid, not just what to copy.

"would_fail_because": "If this hook said 'crypto secrets' instead
of 'no crypto, no dropshipping', the anti-hype share trigger
disappears and share rate drops roughly 40%."