Feedback Bridge — Engagement to Roster Suggestions

social/feedback-bridge reads the posts log + engagement snapshot history (both produced by research/viral-gate-status) and emits a suggested roster delta — topics ranked by max-views, formulas weighted by win rate against a baseline. Suggestions are appended to ~/.fabric/roster-suggestions.jsonl.

Why HITL?

A roster change during a campaign is a confound. If the feedback bridge auto-promoted a topic mid-week and the next week’s growth spiked, you can’t tell whether the new topic worked or whether it was random. Human-in-the-loop preserves the experiment.

For longer-running operations (post-30d), an operator can apply suggestions weekly and treat the roster as a living document — just not during the validation window itself.

Input

cohort: validation-cohort-1
posts_log_path: ~/.fabric/posts.tsv
snapshots_path: ~/.fabric/post-engagement-snapshots.jsonl
roster_path: examples/character-rosters/validation-cohort-1.yaml
suggestions_path: ~/.fabric/roster-suggestions.jsonl
window_days: 14
baseline_threshold: 5000

Field	Required	Description
`cohort`	no (default `default`)	Cohort label to evaluate.
`posts_log_path`	no	TSV posts log. Empty = `~/.fabric/posts.tsv`. Same shape as viral-gate.
`snapshots_path`	no	JSONL snapshot history written by `research/viral-gate-status`.
`roster_path`	no	Optional roster YAML. When set, untouched topics from `topics_pool` appear in the rank with `max_views=null`.
`suggestions_path`	no	JSONL output. Empty = `~/.fabric/roster-suggestions.jsonl`.
`window_days`	no (default 14)	Mirrors viral-gate’s evaluation window.
`baseline_threshold`	no (default 5000)	Per-run views bar for “win”. 5k is plan 073’s failure-cluster floor.

Running it

Most natural pairing — run viral-gate first to refresh snapshots, then the feedback bridge:

fab-workflow research/viral-gate-status \
  --input posts_log_path=~/.fabric/posts.tsv

fab-workflow social/feedback-bridge \
  --input cohort=validation-cohort-1 \
  --input roster_path=examples/character-rosters/validation-cohort-1.yaml

Output shape

The workflow output mirrors what’s appended to the JSONL log:

{
  "cohort": "validation-cohort-1",
  "posts_log_path": "/Users/me/.fabric/posts.tsv",
  "snapshots_path": "/Users/me/.fabric/post-engagement-snapshots.jsonl",
  "suggestions_path": "/Users/me/.fabric/roster-suggestions.jsonl",
  "suggested_at": "2026-05-06T12:34:00+00:00",
  "posts_evaluated": 21,
  "posts_with_metrics": 19,
  "topic_rank": [
    {
      "topic": "The exact prompt I use to draft cold emails…",
      "max_views": 67000,
      "posts_evaluated": 2,
      "formula_used_most": "youre_doing_it_wrong",
      "rank": 1
    },
    {
      "topic": "Stop optimizing your morning routine…",
      "max_views": 1200,
      "posts_evaluated": 1,
      "formula_used_most": "reframe",
      "rank": 8
    },
    {
      "topic": "(untouched pool topic)",
      "max_views": null,
      "posts_evaluated": 0,
      "formula_used_most": "",
      "rank": 22
    }
  ],
  "formula_weights": [
    {
      "formula": "youre_doing_it_wrong",
      "posts_evaluated": 8,
      "posts_winning": 5,
      "win_rate": 0.625,
      "delta": 1
    },
    {
      "formula": "validation",
      "posts_evaluated": 5,
      "posts_winning": 1,
      "win_rate": 0.20,
      "delta": -1
    }
  ],
  "rationale": "Top topic 'The exact prompt I use…' hit 67,000 views (55.8× the bottom-ranked topic with data). Formulas above cohort win-rate avg (vs 5,000 views): youre_doing_it_wrong (62%). Below avg: validation (20%)."
}

delta:

+1 — formula’s win rate is ≥10pp above the cohort average. Promote.
-1 — formula’s win rate is ≥10pp below. Demote or remove from allowed_formulas.
0 — at average. Leave alone.

How to apply a suggestion

Open the roster YAML at roster_path.
Reorder topics_pool to put high-rank topics first. Drop topics that consistently rank in the bottom quartile across multiple weekly suggestions (one bad week is noise).
Adjust allowed_formulas ordering — promote delta=+1 formulas to the front so slot_offset rotation hits them more often.
Commit the roster change with a note explaining which week’s suggestion drove it. This makes the campaign auditable.

The committed example roster examples/character-rosters/validation-cohort-1.yaml is your starting point.

Validation campaign integration

This workflow closes the loop in the 30-day validation playbook. Run it weekly (every Monday) alongside research/viral-gate-status and social/account-growth. Apply suggestions only at Day 14 and Day 21 checkpoints — frequent reroll dilutes the signal.