Bullseye Funnel (5 Layers)

Porting scope (source: /opt/algorithms/functional/funnel_evaluate.py, 133 lines, sha256 353f0369…)
· Aligned: LAYERS — the 5 layers × 12 metrics and all the green/yellow thresholds (L19-57); the colour judgement ≥green 🟢超线 / ≥yellow 🟡偏线 / 否则 🔴迷路 (L88-93); missing data is recorded as “no data” and does not count towards that layer's colour (L85-87); the layer colour takes the worst metric in that layer (🔴 if there is one, otherwise 🟡 if there is one, 🟢 only if all are green, and ⏳ no data only if nothing was filled in) (L97-104); the break condition and the 4 prescriptions (FAULT_DIAG L59-64 and L119-126).
· Equivalent but looks different: the source code reads the data with data.get(指标名, data.get(指标名.replace(' ', '_'))), and none of the 12 metric names contains a space, so that fallback is a no-op —— this page reads straight by key name.
· Not aligned ①: all 4 pairs of adjacent layers are in FAULT_DIAG in the source code, so the L121 fallback ('未知断层', '综合优化') is never actually reached; this page keeps that fallback and flags it here.
· Not aligned ②: if the data the source code gets from the API contains strings, the comparison at L88 raises a TypeError; this page forces a number conversion at the input layer and errors out on anything non-numeric (the source code has no such conversion).
· Not aligned ③: the source code's layers is a dict (order = definition order); this page outputs it as an array in the same order, field for field; the data_bus write (L132) and the info metadata attribute are not ported (a pure front-end has no bus, and the page does not need them).
· No extensions: the source code judges only the 4 pairs of breaks between adjacent layers out of the 5. Cross-layer collapses like “entry 🟢, conversion 🔴, layers in between not green” the source code does not judge —— this page copies it without extending.
· Why the two items under “next step” are set by hand: the types this page produces are layer_verdict / fault_diagnosis, and none of the 4 existing slices consumes these two types, so build.py's automatic wiring cannot draw edges (it is not a broken link, it is types that do not connect). The two below are hand-set business downstreams:
　① → network effect: the funnel tells you which layer the content breaks in, and network effect tells you how far user scale is from the critical point;
　② → scale conversion: this page is a single-period snapshot and cannot show a trend. If you have historical multi-period data, first lay each period's metrics out on a 0~10 scale, align them with the engine's ±10 scale using “scale conversion”, then go into “curve diagnosis” to see the movement —— this page does not do that conversion.

[Evidence section · where to check every number]
① Source: the algorithm-port machine /opt/algorithms/functional/funnel_evaluate.py (133 lines, read-only): the 5 layers × 12 metrics and all the green/yellow thresholds LAYERS = lines 19-57; the break prescriptions FAULT_DIAG = lines 59-64; run() = lines 74-133; the colour judgement raw_val >= ind["green"] = lines 88-93; the layer colour takes the worst in that layer = lines 97-104; the break condition and the join = lines 119-126.
② Three categories
· [Verifiable] The 12 green/yellow thresholds, the three-tier colour judgement, missing data recorded as “no data” and excluded from the layer colour, the layer colour taken from the worst metric, the break judged only as “upper layer 🟢 and lower layer 🟡/🔴”, and the 4 prescriptions —— all checkable line by line.
· [This page's display convention] The 5 layer colour blocks and the order text are page layout; “the thresholds are proportions, 0.05 = 5%” is a warning that the source code “compares the value straight against the threshold, with no percentage conversion”. This is not source-code output — the source code has no unit note.
· [No source found] The 12 thresholds (CTR 5%, completion rate 60%, save rate 5%…) have no data source or literature citation noted in the source code —— the thresholds themselves have no source found (that is how the provider handed them over).
③ How to check it yourself: fill in only the entry layer's CTR 0.05 and leave the rest blank → the entry layer should be 🟢 and the rest ⏳ no data, with 0 breaks; then keep the entry layer's CTR at 0.05 and put the reception layer's completion rate at 0.10 (below the yellow line of 0.30) → one entry→reception break should come out, with the prescription “title promise = content delivery”.