Methodology
The current V4 baseline, the pre-registered V5 research target, acceptance gates, platform support, uncertainty, and limitations.
Deterministic traction baseline
The score currently visible in the graph is the production deterministic index and rollback baseline. It is not a calibrated probability, a causal estimate, or a learned prediction of company quality. V5 will replace it only after compatible longitudinal data, leakage-safe held-out evaluation, calibration, reproducibility, subgroup, and runtime-parity gates all pass.
V5 validated platform coverage is currently none: every platform remains v4-only until it has enough compatible, licensed, longitudinal examples to clear the same held-out acceptance gate. An unsupported V5 row is unscored rather than routed through another platform's parameters.
Favorite score is a separate 1–100 signal about the conviction a YC partner expresses toward any startup in the selected batch. It uses already-ingested public commentary. Explicit superlatives, strong endorsements, and specific reasoning about a team, market, product, or technology carry much more weight than a short tag or congratulations.
The strongest attributable statement sets the pair's foundation. Additional independent posts add a bounded, diminishing-return bonus, and duplicate or cross-posted copies count once. Skeptical language can reduce the result. The score does not measure company quality, investment merit, ordinary traction, or the number of posts by itself.
Confidence is separate from Favorite score. It reflects unique supporting posts, independent contexts, platform breadth, attribution quality, date completeness, and verified source links. No commentary is not treated as evidence of dislike; it simply produces no attributable ranking signal.
This explains returner-traction v4.3.1, not the intended learned V5 formula. The model starts with verified native evidence whose configured visible metrics are normalized to canonical aliases. It maps each raw count through the table value, then applies a platform-specific logarithmic reference. After that evidence value is clamped, the configured score-level calibration retains 95% before integer rounding. The current monotonic patch uses 100% reference-anchored absolute signal and 0% evidence-level cohort midrank, so changing one row cannot lower an unchanged same-platform peer.
Publication date and post age do not raise or lower an evidence score: identical visible metrics receive the same score regardless of when they were published. Duplicate physical posts count once; only the strongest 2 posts per platform contribute, at 95%, 5% by slot. Posts are therefore not simply summed without limit.
The strongest platform supplies 95% of the entity score. The remaining 5% is a fixed-share corroboration slice using the configured shares below; missing platforms contribute zero only inside that bounded slice. Adding platforms therefore cannot lower a score, while platform breadth can change it by at most 5 points. That raw absolute score remains the auditable benchmark input (100% absolute and 0% cohort-percentile signal). The displayed headline uses one ratio shared by every supported batch: the strongest current company's absolute score maps to the configured global target of 95, and every other company receives the same global calibration. There is no per-batch min/max stretch, and platform visibility filters never recompute this canonical factor.
| Platform | Share | Log reference | Raw metric weights before normalization |
|---|---|---|---|
| GitHub | 15% | 250,000 | stars × 1.5 · forks × 4 · issues × 0.5 |
| X | 21% | 120,000 | views × 0.04 · likes × 1.4 · replies × 4.5 · reposts × 6 · quotes × 6 |
| 15% | 18,000 | views × 0.04 · reactions × 1.4 · comments × 4.5 · reposts × 6 | |
| 21% | 80,000 | views × 0.04 · likes × 1.1 · comments × 4.5 · shares × 5 · saves × 4 | |
| Product Hunt | 7% | 4,000 | upvotes × 2 · comments × 3.5 |
| YouTube | 10% | 35,000 | views × 0.025 · likes × 1 · comments × 3.5 |
| 4% | 4,000 | upvotes × 2 · comments × 3.5 | |
| Hacker News | 5% | 2,500 | upvotes × 2 · comments × 3.5 |
| Bilibili | 2% | 35,000 | views × 0.025 · likes × 1 · comments × 3.5 · shares × 4 |
V4 confidence is separate from score: its heuristic starts at 20%, then uses evidence depth (55%, scale 4), platform breadth (5%), publication-date coverage ( 12%), and verified links (8%). Medium/high labels begin at 50%/75%. Publication-date coverage affects this separate confidence metadata, never the score. These weights, references and slot shares are product heuristics; they are not fitted V5 parameters.
V4 summarizes verified, visible platform-native traction evidence on a bounded index. Its current company headline target is 95, not 100. The proposed V5 target is a pre-registered future platform-native performance outcome observed after a genuine t0 measurement; it will be labeled as a probability or percentile only if held-out calibration supports that interpretation.
The versioned source registry separates incorporated, rejected, unavailable, and license-restricted benchmarks. Citation alone does not count as incorporation: data, protocol, baseline, failure-mode test, or acceptance gate must actually enter the reproducible pipeline.
The pre-registered design keeps canonical physical posts in one split, places training before validation and validation before final test, and reserves an entity holdout for future unseen-company evaluation. Planned leave-one-batch-out checks are separate development analyses. The current rejected artifact has no rows, so it reports no unseen-company or unseen-batch result and the final test has not selected a model.
V5 candidates use only features genuinely available at the observation timestamp, including missingness and post age. Separate native signals are not collapsed with a hand-selected exchange rate. Monotonic candidates ensure that increasing a genuine positive signal cannot lower a prediction. Signal-family ablations remain a held-out acceptance requirement; none is reported for the current zero-row rejected artifact.
No hand-picked coefficient or top-post slot vector is accepted for V5. Linear coefficients, spline effects, calibration maps, temporal curves, and company pooling parameters must be fitted under the frozen search protocol and survive held-out evaluation. Nonlinear marginal effects depend on platform, age, and context.
V5 does not assume that platform breadth is beneficial. Cross-platform combination must be learned and validated, or platform scores remain separate. Coverage may narrow uncertainty without secretly adding score points. Visibility filters never recompute a canonical score.
The deterministic calculation retains a 0–100 safety bound, then retains 95% at the evidence level and uses a 95-point company headline target. Publication date and post age do not affect the score, so an older post is not discounted and a newer post receives no freshness bonus. Date completeness may be reported separately as confidence metadata. Any future learned age effect would require its own held-out validation before it could change a score.
Score, predictive uncertainty, evidence coverage, link verification, date quality, and source reliability are separate fields. A completeness heuristic is never called a statistical interval. Unsupported platforms and out-of-distribution inputs remain visibly unscored.
Public engagement is selected, platform-dependent, missing-not-at-random, and potentially manipulated. The model does not establish causality, company quality, valuation, or investment outcomes. Fairness and transfer claims are limited to the subgroups and platforms actually tested.
Lines explain only the relationship types present in the map, such as shared industry context or a shared group partner. They do not imply company interaction and never add score points. The map legend exposes the exact relationship explanations available in this graph.
A future accepted V5 artifact will state: “The parameters were fitted on versioned benchmark and longitudinal data under a frozen evaluation protocol. They are predictive associations for the stated target, not causal estimates of company quality or investment outcomes.” Until that gate passes, the product makes no such fitted model claim.
Read the visible model version before comparing scores, and use the native evidence, observation cutoff, coverage, and limitations alongside the number. See data sources for provenance and corrections for disputed records.