How the Streamwake reliability score
is computed.
A editorial derivation of the quarterly per-stream reliability composite — the probe distribution, the synthetic failure shapes the fleet injects, the four-axis weighting that turns raw verdicts into a published percentile, and the vendor-neutral guarantee that keeps the benchmark from inflating toward whoever sponsors it.
Three axes, read off each declared window.
The probe fleet spans three axes — codec mix, CDN mix, ISP mix — and every share is declared for the observation window. No axis is hand-weighted toward a partner, a sponsor, or a preferred vendor; the methodology explains how the declared mix drives the signal, so a reader can evaluate any future published percentile.
- H.264 / AVCDeclared per report
- H.265 / HEVCDeclared per report
- AV1Declared per report
Production-share weighting is a rule for an approved dataset, not a current result. Any floor or inclusion rule should be named alongside the published window and source.
- CloudflareDeclared per report
- FastlyDeclared per report
- AkamaiDeclared per report
- Amazon CloudFrontDeclared per report
- Edge origin shieldDeclared per report
Pull-share weighted so the declared egress mix drives the report signal. The methodology does not lock the mix; each published window should identify the source and period for its shares.
- Residential broadbandDeclared per report
- Mobile carrierDeclared per report
- Enterprise transitDeclared per report
Viewership weighting is a rule for an approved dataset, not a current result. Any weighting rule should be named alongside the published window and source.
Five shapes the fleet injects.
Five per-cadence synthetic failure shapes are introduced into the probe stream so the percentile can be read against a known distribution rather than against only the natural failures the cohort happens to produce in a quarter. Each row names the public probe identifier a reader will find on the agents feed.
Egress saturation. TTFB p95 drifts from a ~360ms phased baseline into the 1–2s range.
~1 in 42 cadences
fetch.ttfb_drift · scoring impactScores against the early-warning posture — fires before any 4xx lands, so the agent can route the affected cohort away from the saturated egress without paying the rebuffer-report cost first.
CDN rotation / stale cache-tier. Same URI returns a different SHA-256 byte stream from the backup CDN than the primary.
~1 in 87 cadences
segment.hash_drift · scoring impactCounts against the verification-side composite. Catches a CDN rotation that came up with a stale cache tier — protects the cohort from handoff artifacts that a header-only probe misses.
Outbound fanout pegs at the pool ceiling during a peak. Player retries pile on top.
~1 in 31 cadences during peaks
fetch.concurrency · scoring impactHits the detection-time weight directly. The agent auto-throttles probe fan-out and meta-classifies as join-storm vs regional-skew, then routes the cohort to a healthier egress.
A regional roll lands on a CDN where the signed-URL region key disagrees with the player GeoIP. 4xx spike on the alternate edge.
~1 in 124 cadences
playlist.reachable · scoring impactCounts against restore-time. Agent re-derives the signing rule from the player’s claimed region and surfaces a routing suggestion rather than a flapping roll that re-triggers the same 4xx.
PART-HOLD-BACK drifts away from the delta cursor after a partial-segment rewrite. Latency sits at 8–12s instead of the expected 2–3s.
~1 in 58 cadences in LL-HLS streams
llhls.part_holdback_drift · scoring impactCounts against detection-time in the low-latency arm. Agent replays the last three partials to clear the cursor and reasserts PART-HOLD-BACK on the publish clock.
Four axes, one anchor.
The composite weights fast detection plus verified remediation over zero-event totals. A stream that survived a quarter without a verdict pays no bonus — the signal rewards the streams whose agent loop closed an incident, not the streams whose failures happened to stay quiet.
| Axis | Weight | Why it carries this weight |
|---|---|---|
| Restore-time wins | × 1.0 (anchor) | A verified remediation event (typed action applied + next probe cycle clear) is the strongest signal that the agent loop closed. Anchored at 1.0 — every other axis is read against it. |
| Detection-time wins | × 0.75 | Catching a verdict under one of the heuristic families before any header-level probe fails is high-leverage — but only because the matching remediation still gets to run. Lower weight than the verified remediation because detection without remediation is just an alert. |
| False-positive cost | × −0.40 | Each detected heuristic that closed on the next cadence without a real remediation is subtracted from the composite. The program is read by on-calls comparing their SLOs against it — a noisy signal punishes every reader of the percentile. |
| Vendor-grading | × 0.00 | Set to zero by the vendor-neutral guarantee below. A stream’s CDN-of-record, encoder vendor, packager vendor, or sponsor status cannot lift or lower its individual percentile — the signal carries no vendor pitch. |
The composite lands on a 0–100 per-stream scale so an on-call can compare a single SLO number against a single published figure without unwinding four underlying weights themselves. weight values are read at the published percentile, not re-derived per reader — the rule is the rule, and the rule is published here.
What a high score vs a low score actually means.
The percentile figure is only as useful as the operational translation a reader can draw from it. Two columns — the operator who survived a noisy quarter with the agent loop closing incidents, and the operator who carried avoidable detection and remediation latency because the loops never closed.
Example: join-storm and CDN-rotation signals stay in one recovery record.
A high-composite quarter means the probe fleet's three heuristics — TTFB drift, fetch-concurrency exhaustion, segment hash drift — fired, classified, and closed under the autonomous loop: classify, remediate, verify. This illustrative example keeps the alert feed useful by showing which signals block an incident and which remain contextual; it does not claim a measured intervention rate.
Example: passive-monitor lag leaves detection and remediation work visible.
A low-composite quarter means verdicts landed but the loops did not close. Alerts fired after rebuffer reports — too late to claim the early-warning posture — and remediations waited for a human to clear the runbook step a heuristic could have taken on its own. The cohort paid the rebuffer cost first, then the recovery cost, then the postmortem cost.
A passive monitoring pipeline logs the same verdicts and waits for a rebuffer report — by which point you are already paying the cohort-trust cost and writing the postmortem. The agentic-ops layer flips the polarity: classify the verdict under one of the heuristic families, remediate by replaying from origin or re-routing the affected CDN, then verify that the next probe cycle clears. The composite is read against that polarity, not against the missing-event count.
No vendor grade, no sponsor lift.
The benchmark's vendor-neutral posture is a written rule, not a positioning line. It is published on the methodology page so the percentile breakdown stands on the rule alone.
Weighting for "CDN-of-record", encoder vendor, packager vendor, or DRM tier is set to × 0.00 on the weighting table above. Sponsor status cannot lift a stream's individual percentile; the cohort should be built from an identified dataset — we do not invite vendors to sponsor, take ownership of, or co-author the breakdown. Any published numbers must stand on the named dataset and methodology.
Raw probe data and the per-stream breakdown are available under NDA to enterprise prospects — write to the public contact form.
Read the methodology in the rest of the library.
The methodology explainer is one of four sibling surfaces that share the cohort signal. The landing page, the quarterly report, the agent loop landing page, and the Incident Lab library are four ways to read the same rule.