## Extended lag analysis: S24 and W25 through +6 batches

Added 28 October 2026. **A longer window does not reveal a broad, durable delayed rise in this lexical screen.** Both cohorts rise at +4 in the full-description method, but drop below their preceding-batch baseline at +6. The +4 rise is not stable across text specifications. A few individual themes show stronger later alignment.

This extends the existing descriptive analysis rather than estimating whether RFS caused later investments. It reuses the same pinned October snapshot, fitted text corpus, request weighting and threshold rules, so the +0 to +2 results reconcile exactly with the earlier output at published precision.

### Cohorts and available timeline

The fixed cohorts are the 16 requests retained in today's Summer 2024 tab and all nine Winter 2025 requests: 25 request entries throughout every offset. Four categories in the original February 2024 announcement are absent from the retained S24 tab and remain outside this extension. Do not generalise it to the complete original 20-category list.

| Offset | Summer 2024 cohort follows | Winter 2025 cohort follows |
|---|---|---|
| +0 | Summer 2024 | Winter 2025 |
| +1 | Fall 2024 | Spring 2025 |
| +2 | Winter 2025 | Summer 2025 |
| +3 | Spring 2025 | Fall 2025 |
| +4 | Summer 2025 | Winter 2026 |
| +5 | Fall 2025 | Spring 2026 |
| +6 | Winter 2026 | Summer 2026 |

Both +6 endpoints are completed batches in the snapshot. Fall 2026 and Winter 2027 are excluded from outcome windows. Offsets are relative to the named batch, not independently established publication dates. On nominal seasonal batch-start months, +4 is about 12 months and +6 about 18 months. The S24 list was originally announced in February 2024, so these nominal intervals understate its actual time since publication. The study does not claim an exact day-level publication clock.

### Main result: a +4 bump, not sustained cohort-wide growth

Median lift in the full-description/metadata method, with each request's threshold fixed at its three-preceding-batch 95th percentile:

| Fixed request cohort | Requests | Named batch (+0) | +2 | +4 | +6 | Above baseline at +4 / +6 |
|---|---:|---:|---:|---:|---:|---:|
| Summer 2024 | 16 | 1.02× | 0.83× | 1.18× | 0.89× | 11 / 7 |
| Winter 2025 | 9 | 0.95× | 0.83× | 1.20× | 0.77× | 6 / 3 |
| Both cohorts | 25 | 1.02× | 0.83× | 1.18× | 0.86× | 17 / 10 |

For S24, 11/16 requests exceed baseline at +4, declining to 7/16 at +6. For W25, the equivalent counts fall from 6/9 to 3/9. The combined median is 1.18× at +4 and 0.86× at +6. Combining cohorts is descriptive: their calendar follow-ups overlap and their themes are related, so the 25 entries are not independent experiments.

A lift of 1.18× means the share of companies exceeding a lexical-similarity threshold is 18% above its constructed baseline. It does not mean 18% more companies actually satisfy an RFS or that RFS increased investment by 18%.

### Does the result survive changing the text?

The same 95th-percentile thresholds are recalculated separately for each text specification:

| Cohort | Text | +4 median | +6 median | Pooled +3 to +6 median |
|---|---|---:|---:|---:|
| Summer 2024 | Full descriptions + metadata | 1.18× | 0.89× | 1.03× |
| Summer 2024 | Company one-liners | 1.00× | 1.09× | 1.06× |
| Summer 2024 | Generic terms removed | 1.12× | 0.94× | 1.00× |
| Winter 2025 | Full descriptions + metadata | 1.20× | 0.77× | 0.98× |
| Winter 2025 | Company one-liners | 0.80× | 1.03× | 0.82× |
| Winter 2025 | Generic terms removed | 1.30× | 0.86× | 1.00× |
| Both cohorts | Full descriptions + metadata | 1.18× | 0.86× | 1.02× |
| Both cohorts | Company one-liners | 0.82× | 1.03× | 0.96× |
| Both cohorts | Generic terms removed | 1.18× | 0.89× | 1.00× |

At +4, S24's 1.18× full-text median is approximately 1.00× using one-liners. W25's 1.20× becomes 0.80×. At +6, one-liners give 1.09× and 1.03× respectively, while the two richer text specifications sit below baseline. There is no common sustained trajectory across these methods.

Nine specifications combine these three text definitions with 90th-, 95th- and 97.5th-percentile cutoffs. Across all nine, the combined-cohort median at +4 ranges from 0.82× to 1.40×; at +6, from 0.79× to 1.19×. These ranges describe method sensitivity, not statistical confidence intervals.

### Pooling later batches instead of choosing a single endpoint

For each request, sum threshold-exceeding companies across the chosen window, divide by all companies in those batches, and compare that share with the request's fixed baseline. Then take the median across requests. This weights batches by their disclosed company counts and avoids averaging endpoint medians.

Full text, 95th percentile:

| Cohort | Pooled +0 to +2 | Pooled +3 to +6 | Pooled +4 to +6 |
|---|---:|---:|---:|
| Summer 2024 | 0.98× | 1.03× | 0.94× |
| Winter 2025 | 0.95× | 0.98× | 0.95× |
| Both cohorts | 0.96× | 1.02× | 0.95× |

Across all nine specifications, the combined pooled +3-to-+6 median ranges from 0.84× to 1.08×; the narrower +4-to-+6 window ranges from 0.92× to 1.08×. Pooling does not uncover a large typical delayed increase. This conclusion is descriptive; no formal causal test or power analysis has been performed.

### Individual request results

Pooled +3-to-+6 lift at the 95th percentile:

| Request | Release | Full text | One-liners | Generic terms removed |
|---|---|---:|---:|---:|
| New Defense Technology | Summer 2024 | 2.01× | 1.68× | 2.16× |
| Applying Machine Learning to Robotics | Summer 2024 | 1.71× | 1.80× | 1.50× |
| Manufacture in the USA | Winter 2025 | 2.03× | 1.62× | 2.03× |

These three themes remain above baseline across all three text definitions. Defense already has 2.05× lexical lift in S24's named batch, so its later strength is not evidence of a newly emerging delayed effect. Robotics rises from 1.26× in the named batch to a pooled 1.71× later; US manufacturing rises from 1.07× to a pooled 2.03×. These are lexical results, not validated counts of companies responding to RFS.

For scale, at +6 the full-text threshold flags 27/197 Winter 2026 companies for the S24 defense request, 17/197 for robotics, and 38/229 Summer 2026 companies for W25 manufacturing. The underlying baselines are roughly 5% by construction. These are automated flags; prior company audits found material false positives.

### Interpretation and remaining limits

The longer window weakens the argument that the initial modest aggregate simply reflects insufficient observation time. It does not show that RFS has no effect. Effects could vary by request, change applications without altering the selected batch, or be missed by lexical matching.

A company appearing several batches later may predate the request, pivot into the category, or have been influenced by a newer request. Current descriptions, improved tags, general sector growth, repeated themes and YC's selection all remain explanations. The pre-batch trajectories are uneven too, particularly W25's high -2 observation.

The fixed October TF-IDF corpus includes later descriptions; this is not a held-out forecasting exercise. The comparison has no matched unrequested topics, non-YC counterfactual or applicant denominator. We therefore report no causal effects, significance claims or investment-return conclusions.

### Sources and validation

All 75 overlapping short-window comparisons match the earlier analysis at its published three-decimal precision. Validation checks cover fixed baseline shares, nonempty denominators, fixed cohorts, completed follow-up batches and the pinned source hash.

Sources are the existing [pinned YC-OSS snapshot](https://github.com/yc-oss/api/blob/e071a842f4067f9f8ebf980f49cae96e8d93132e/companies/all.json), the retained [YC request corpus](https://www.ycombinator.com/rfs), and the original [February 2024 announcement](https://www.ycombinator.com/blog/ycs-latest-request-for-startups) documented in the main report. The extension introduces no newly verified company-level classifications.