Review · VerifiedUpdated 2026-05-03Verified against code 2026-05-03

Verified 2026-05-03 against MEMORY.md (Visibility-analysis active-prompt filter + rotation warning gate, v3.8.2 production state).

06 — Historical Tracking

The sixth stage closes the loop. Every report is generated with full access to the workspace’s report history and produces week-over-week deltas. But “the score moved” is only useful if the comparison is fair — that’s what change-guards enforce.

Week-over-week deltas

Every Visibility Analysis Report is generated with access to the full history of previous reports for the same workspace. Week-over-week deltas in visibility score, audit score, share of voice, and all key Phase 6 metrics are computed and surfaced automatically:

Banner shows current score + delta vs last period
Trend lines for score, mention rate, share of voice, sentiment
Per-provider deltas (catches “ChatGPT collapsed but the headline number didn’t move much”)
Per-prompt history (which prompts gained or lost ground)

Change-guards (the integrity layer)

Comparing this week’s score to last week’s only makes sense if the underlying calculation hasn’t shifted. Two situations break that:

1. Scoring methodology change

When the platform’s scoring code changes between versions (new dimension weights, new normalisation, new providers added to base tier), the next report’s WoW delta would mostly reflect the methodology change, not real customer movement. The change-guard suppresses the delta in those cases and surfaces an explanation banner instead.

2. Active prompt set rotation > 20%

If a workspace adds, retires, or significantly restructures its prompts between collection cycles, the score on the new prompt set isn’t comparable to the old one. The active-prompt filter at fetch_responses_node (PR #318, v3.8.1) ensures only currently-active prompts feed the score. The rotation warning gate (PR #319, v3.8.2) detects >20% rotation and:

Marks the report’s WoW delta as suppressed for 2 post-rotation cycles
Auto-decays the warning state after 2 stable cycles via wow comparison
Surfaces an explanation in the report

This was triggered by the Kess Berlin “vegan beauty” booster leak in April 2026 — 1 of 26 weekly reports was materially contaminated by retired-prompt data. The gate makes that class of bug structurally impossible going forward.

Why this stage matters for sales

The honest answer to “can I trust this number over time?” lives here. Without change-guards, every score movement is suspect. With them:

Customers can run an 8-week trend chart in a board pack and defend it
Engineering can ship scoring improvements without erasing customer trust in the metric
Prompt churn (which is a normal part of using the product) doesn’t poison the trend

It’s the boring infrastructure that makes the Stage 4 report credible.

How the loop closes

After Historical Tracking processes this cycle’s report, the next Stage 4 report generation inherits:

The new historical baseline
The change-guard state (active or auto-decaying)
The full Action history from Stage 5 (completed/pending/dismissed) so Boosters can adapt to “they already did that”

Then Stage 1 collection runs again the next morning. The loop is closed.