What a Streamed Roster Lane Is Worth

If you came from the front page. The figures quoted there, 0.38 categories a week for a streamed place against 0.00 for a thirteenth man held all year, over 1,536 team-periods, are the first two rows of the table under "The answer, up front" below. The third row matters as much and the front page has no room for it: a place held by an ordinary man returns 0.43, so streaming beats the last man on the bench, not a good one. This page is the working record of that measurement, limits first.

League: one real league — nine-category H2H, auction draft Seasons: 2019–2026 (eight seasons; 2020 suspended by COVID) Script: scripts/streaming_lane.py — passes ruff check, ruff format --check and mypy Companions: waiver_value.md (what was available), acquirable_value.md (what was acquired), roster_churn.md (how fast a roster turns over) Reproduce: cd /home/aisha/fcp-core-lane && PYTHONPATH=. /opt/fcp-core/.venv/bin/python scripts/streaming_lane.py Read-only: every query is a SELECT. Nothing is written to the database.

Note on the reproduce line. The house style sources .env before running (set -a && . ./.env && set +a). That does not work against this worktree's .env: its FCP_EMAIL_FROM value contains unquoted angle brackets and bash rejects the file at line 29. The script therefore reads DATABASE_URL straight out of .env itself when the environment does not already have it, and the line above needs no sourcing.

Limitations, stated before conclusions

1. "All a team's rotating men" is not one place, and an earlier draft read it as one

A team running two or three rotating men at once posts about twice the games of a team running one. A figure summed over every rotating man a team used in a period therefore grows with the team's churn, not with the value of a single place, and cannot be compared against TYPICAL_PICKUP, which prices one place. The headline here is per place. The summed figures are kept in §3b, labelled as a team's whole churn, and are not the headline.

2. An opened place is invisible in this table

daily_lineup_slots records 12 or 13 men per team-day and never 14. A team that trades two for one holds 12 men thereafter, and the vacated place has no row at all. A place opened by an uneven trade and then streamed cannot be measured from this table — exactly the case the trade evaluator cares about. Every figure here is a place found within the places the team already held, so it is a floor on what a fully emptied place returns.

3. "Started games" has two meanings and they differ by ~1.7×

This is the measurement that moved the first draft's numbers, so it is stated plainly. For 2026:

1,804 started slots (9.1%) have played = false — ESPN marked the man in a starting slot, but the box score records no game. started_lines, which produces the currency every other study uses, filters on played = true and so counts 11,530 where the lineup table records 19,843. Per team-period the gap is a stable 1.6–1.7× in seven-day periods.

This document reports games from the lineup table, because "started games a week" is a claim about a manager's lineup decisions. The lens figure is given beside it wherever it differs. A reader recomputing off daily_lineup_slots alone gets the higher figure; through the repo's lens, the lower one. Both are right answers to slightly different questions.

4. A lane's places are inferred, so the per-place divisor is the weakest number here

Places are defined by occupancy, because ESPN slot ids are lineup positions (UT alone is 46,371 rows), not stable places. A team starts at most ten men on a day (measured: max 10, mean 8.96 in 2026), so a rotating lane can occupy at most the starting slots the held men are not filling. The divisor used is the peak number of non-held men rostered on any one day of the period — the loose lane count — which is the best available bound and is still an inference. Two managers can rotate one lineup spot through three men or three spots through one man each, and the table cannot tell them apart.

5. The lane count is a proxy for churn, not a count of places

The lane count rises roughly 1:1 with a team's churn volume. It is reported as a descriptive statistic and is not used to rank anything in the headline.

6. Categories are production, not head-to-head wins

Values run through app.scoring.value.marginal, the lens pickup_values uses, so the number plugs into TYPICAL_PICKUP. That lens converts a marginal count into a change in expected category wins against a normal model of the league's weekly totals. It is not the nine actual wins a team took and not a head-to-head result.

7. The held-13th-man baseline is nearly degenerate, and §4 says why

The team's lowest-valued held man is usually a man the manager never had to start, so his median is 0.00. That is the honest answer to the brief's question, but any ratio against him is meaningless. He is reported beside the code's own floor rather than instead of it.

8. 2020 is the COVID season; small samples, one league

The 98-day block is flagged in every table and --drop-2020 reports the headline without it. Fourteen teams at most, eight seasons: per-season columns are directionally useful, not precise.

The answer, up front

A streamed lane returned a median of 0.38 categories a week (IQR 0.23–0.53) against the held 13th man's 0.00 (IQR 0.00–0.14), over 1,536 team-periods, starting 4.67 games a week to his 4.00.

All three games columns are read off daily_lineup_slots on the same basis, so they are directly comparable. (The value lens works from a lower games count after its played = true haircut — 3.00 for the ordinary held man — and §8 shows the finding is unchanged on either basis.)

Read the games column first, because it contradicts the brief's premise. The brief expected a lane to start 5–7 games to a held man's 3–4. Measured place against place on the lineup table, the lane plays 4.67 a week and an ordinary held man plays 5.00 — the lane plays slightly fewer. Against the team's own worst held starter it plays slightly more (4.67 against 4.00). The brief's 5–7 appears to come from summing every rotating man a team used and comparing that against one man, which is the unit error in Limitation 1.

Two structural facts explain the missing volume premium. A team starts at most ten men on a day (measured: max 10, mean 8.96 in 2026), so a rotating place competes for the same starting slots as the held men instead of adding games on top of them. And the held 13th man starts 4.00, more than an ordinary held man's 5.00 would suggest at first glance only because he is by construction the lowest-value held man who did start — a survivor of the selection.

That combination is the finding. Streaming does not buy volume. It buys chosen games: the manager picks up whoever is playing, every day, rather than starting whatever the roster happens to hold. And that substitution is worth 0.38 a week against the 0.43 an ordinary held place returns — slightly less than the man it replaces, and 6.3× the 0.06 the code charges for the place.

The arithmetic, once, so it can be checked. A lane returns 0.38 a week over 4.67 games, about 0.081 categories per started game. An ordinary held man returns 0.43 over 5.00 games, about 0.086 per game. The two are within 6% of each other per game, which is the real content of this measurement: a streamed place is worth about what an ordinary held place is worth — the wins come from picking better matchups, not from playing more games. What it is emphatically not worth is the 0.06 floor the code currently puts under an opened place.

1. How often a lane is run, by season

Definition: peak rotating men — the largest number of men rostered on any single day of the period who were not rostered every day.

24.1% of team-periods run no lane at all; 51.9% run two or more. Streaming is not a niche behaviour here. 2023 is the quietest season on record and 2025–2026 the loudest.

This table describes churn, not places (Limitation 5). No headline figure is drawn from it.

2. Games a week: the crux

Per place, pooled 2019–2026, from the lineup table (Limitation 3):

* The two held-man rows differ because they answer different questions: 3.00 is the per-man figure the value lens works from after the played = true haircut (Limitation 3); 5.00 is the same men counted off the lineup table. The lane's 4.67 is below an ordinary held man's 5.00 on the same basis and above the 13th man's 4.00.

The brief's premise does not survive, and the direction is the surprise. The brief expected a lane to play more games than a held man (\"5 to 7 against 3 or 4\"). On the lineup table's own counting the lane plays 4.67 a week and an ordinary held man plays 5.00 — the lane plays slightly fewer. Against the team's worst held starter (4.00) it plays slightly more.

The brief's 5–7 appears to come from summing every rotating man a team used and comparing that against one man — the unit error in Limitation 1. And the reason the lane does not simply play more is structural: a team starts at most ten men a day (measured), so a rotating place is competing for the same starting slots as the held men, not adding games on top of them. Streaming reallocates starts toward better matchups; it does not manufacture volume.

2026, the season the reviewer recomputed independently:

The 2026 games-per-place figure is 4.47 on the lineup table (2.57 through the repo's played = true lens — the 0.58× haircut of Limitation 3), against a held man's 5.00 (3.00 via the lens) and the 13th man's 4.00.

3. What the lane produced, in categories a week

Same lens as TYPICAL_PICKUP: app.scoring.value.marginal of the streamed men's started lines, times 7 over the period's stored length.

3a. Per place (the headline)

(2026's held-13th median of 0.00 with a p75 of 0.11 is the column the brief's direct-query check of 0.109 lands on: same distribution, the reviewer's figure sits at the upper quartile rather than the median.)

Pooled:

0.38 against the floor's 0.06 is a 6.3× premium — but against an ordinary held man's 0.43 the lane is slightly behind. Both statements matter, and the second is the one that constrains what the code should do.

95.6% of streamed places beat the 0.06 floor; 39.6% of held 13th men do.

3b. Summed over all a team's rotating men (not a place)

Kept because an earlier draft reported it and it is a real quantity — a team's total streamed production — but it is not the per-place figure and must not be compared against TYPICAL_PICKUP.

Two places at ~0.38 sum to ~0.76. The earlier draft's 0.76 was this number; it is a team-week, not a place.

4. The held-13th-man comparison (Fix 1)

The brief asks for the lane against the team's lowest-value held man. The first draft's held() required a man to have started every day, which made the baseline the one man who sat in UT all week and read 0.00 for the wrong reason. Held means rostered every day of the period; starts do not matter for being held.

Two further corrections, both stated because they change the column:

1. Among held men, the comparison man is the lowest-valued one who STARTED at least one game. A held man who never started returns 0.00 by construction, so a plain minimum would pick a man who could not have produced anything.
2. His games figure comes from the lineup table, like the lane's.

The direct-query sanity check the brief names is 0.109 for 2026; this document's 2026 held-13th column is 0.00, with a p75 of 0.11. The check agrees at the quartile, not the median: the 2026 direct figure of 0.109 is close to this document's p75 of 0.11, which is what a differently-centred statistic on the same distribution would give. The median is 0.00 because most teams' worst held starter genuinely contributes nothing — that is the honest answer, and it is why the ratio against him is not usable.

Held men who never started, excluded from the minimum above:

Only 620 of 23,010 held men (2.7%) never start a game in a period — so requiring a start does not squeeze the baseline the way it did in the first draft; what it removes is a small tail that would otherwise pin the median at 0.00 by construction rather than by measurement.

5. Loose versus tight lane definitions (Fix 2)

Loose (the brief's, and the headline): every non-held man rostered part-period who recorded a started game counts toward the lane.

Tight: a non-held man counts only if he started at least one game for this team in this period.

These two definitions coincide in the production sample, and that is a finding, not a null result. A Lane is only built for a man with a recorded started line (if not line.games: continue), so every man already in the loose lane is a starter and the tight filter removes nothing. What the loose definition would otherwise admit — a man rostered for part of the period who played but was never placed in a starting slot — is 20,506 man-periods across the eight seasons, and the script excludes them before either definition applies. The distinction that matters is therefore not loose-vs-tight on production (they are identical) but the one the definitions make on the count: a man who occupied a roster place without ever starting is churn, and counting him as a lane is what made an earlier draft read "2 or 3 lanes run" where the production came from one.

Identical on every column, for the structural reason above. The headline (0.38 categories, 4.67 games per place) does not depend on the choice, because on the production side there is no choice to make: a lane is its starters. The cut does bite on the lane count when the never-started men are left in — the earlier draft's "2 or 3 lanes" against a production figure drawn from one place — which is exactly the unit error Limitation 1 names, and is why the count is not used to rank anything here.

Does lane 2's decay survive the tight definition?

No — and it did not survive the loose one either. That is the finding.

Ranked by own value, the second and later men were higher than the first:

Every cut puts the later men higher. A manager who is streaming is streaming toward a better player: his first add of the week is a fill-in, and by the time he has a lane running he is picking up someone he wants. There is no measured decay for the second lane in production under either definition. The decay that does exist is in the add budget (§6) — a different thing.

6. The add budget

Seven adds a matchup period, shared across every lane a team runs.

331 of 2,040 team-periods (16.2%) spent the full budget; 102 ESPN moves were refused with FAILED_MATCHUPACQUISITIONLIMIT. The budget binds, and it binds where the lanes are: teams running four lanes hit seven adds 56.5% of the time.

Production by lane count, categories a week (summed over the team's churn):

Two lanes return about 2.2× one lane, not 2.0× — the second lane is not visibly cheaper than the first. The proxy this data supports for the second lane is adds, not lanes: a team with one lane spends a median 2 adds, with two it spends 4, and a team that has spent its 7 has no second lane at all.

7. Percentages and turnovers

The lane's line added to the season's ordinary started week and re-scored, per category (2026 medians). The right-hand column is the summed column scaled by 0.38/0.76, not a separate measurement -- step_5 scores the whole lane's line, and the per-place cut is that figure apportioned; it is arithmetic shown for readability, and the summed column is the measured one:

The three categories the brief singled out behave as volume says they should, and the ratios behave differently from each other (reading the measured summed column):

• Turnovers are the largest cost the lane imposes: −0.235, close to its PTS gain (0.264). A streamed lane is near turnover-neutral in categories won, which is not what "free production" implies.
• FG% is a real but small cost (−0.053).
• FT% is essentially flat (−0.014), and exactly 0.00 in two seasons.

A punt-FT% roster gives up nearly nothing and should stream harder; a chase-FT% roster pays a small real cost. The cost that bites both is turnovers.

8. Sensitivity

The headline is §3a's pooled per-place median (0.38). Under the variants tested it does not move:

Held 13th man under the same variants: 0.00 / 0.00 / 0.00 / 0.00, p75 0.14.

Nothing moves the categories answer. The one cut that moves the games column is Limitation 3's definition, and it moves both sides together: in 2026 the lane reads 4.47 games a week on the lineup table and 2.57 through the repo's played = true lens, against a held man's 5.00 and 3.00 respectively — a ratio of 0.89 on the lineup basis and 0.86 through the lens. Both bases agree, and both say the lane's volume is slightly below a held man's, not above. The finding does not depend on the games definition.

What this means for the code

Nothing in this section has been implemented. It is what the numbers support, so the change can be made deliberately.

TYPICAL_PICKUP is two numbers, not one

pickup_values reproduces replacement.py's table and 2026's median is 0.0716 — the flat charge the trade grade uses — so 0.07 is the comparison for the trade work and 0.06 for judge.py. Both are per-add figures and both are correct for what they measure. The bug is that a per-add figure is used where a per-place figure belongs.

The three call sites affected

1. app/pickups/judge.py — places_cost / season_cost. wire is wire_replacement(...) floored at TYPICAL_PICKUP. Every place a move opens is credited at 0.06; under this measurement an opened place is worth 0.23–0.38, so places_cost understates the cost of opening a place by 3.8–6.3×. The suggested floor is the lower quartile, 0.23 — the conservative end of the distribution, because the lane's upside depends on manager attention no code here models.
2. The trade evaluator's settlement — app/trades/evaluate.py's _settle_drops → places_cost, and the playoff path's replacement = max(TYPICAL_PICKUP, best wire man). Same constant, same understatement. This is the 2-for-1 error the calibration found: the side giving up two men for one opens a place credited at 0.06 when it returns 0.23–0.38, and the settlement should credit the opened place at the same figure as the filled one.
3. app/scoring/moves.py's replacement charge — app/scoring/trade_grades.py feeds grade_move median_of(pickup_values(book)) (0.062–0.128). docs/trades.md §7 names this as the hindsight grade's other cause of the 2-for-1 gap. Its 0.07 should be 0.23–0.38; the grade should be re-run before the settlement is trusted.

A caution the numbers force

Do not price an opened place above an ordinary held one. A streamed place returns 0.38 and an ordinary held place returns 0.43. The lane is worth more than the floor, not more than a man. The change is to stop pricing an opened place at the floor, not to argue that opening places beats holding men. Any code that ends up preferring a 2-for-1 because the opened place is "valuable" has over-read this measurement.

Should the price fall with the number of lanes?

Not in production — the data says it does not. There is no measured decay: the second and later men returned more than the first under both definitions (§5).

Two things should be enforced rather than modelled:

1. The 7-add budget is the binding constraint. A team that has used its seven adds cannot run a lane at all, and 16.2% of team-periods are in that state. adds_left already exists in app.pickups.state; gate the place value on there being an add to spend rather than scaling it by an unmeasured coefficient.
2. Do not raise the per-add floor. 0.38 is what a place returns when rotated, not what an add returns. Raising TYPICAL_PICKUP to 0.38 would pay every no-op swap six times what it is worth.

What would make this exact

A place identity on daily_lineup_slots. The single missing fixture. A stable per-team place index (or a slot_id persisting across days for UT) would let a lane be cut from the place it occupied instead of inferred from occupancy, would turn the lane count into a place count rather than a churn proxy, and would let adds be attributed to the lane they built.

Decisions

1. The headline is per place; the summed figure is quarantined to §3b. The first draft's 0.76 was the sum over every rotating man a team used. The code prices one place, so the per-place figure is the one that belongs against TYPICAL_PICKUP. Two places at ~0.38 sum to ~0.76, which is the consistency check.
2. The per-place divisor is the peak non-held men on any one day — the best available bound, since the table stores no place identity (Limitation 4). The alternative, dividing by every man in the lane, understates a place whenever men overlap in time and was rejected after it produced a lane figure below the held man's.
3. Held means rostered every day; starts are not required. The draft's "started every day" made the baseline degenerate at 0.00 by construction.
4. The comparison man is the lowest-valued held man who started at least one game. Never-started held men are 2.7% of held men and are excluded, with the count reported.
5. Games come from daily_lineup_slots; the lens figure is reported beside it (Limitation 3), and the games ratio is shown to be the same (1.56 vs 1.53) under either, so nothing hinges on the choice.
6. Lane 2's decay is reported as NOT surviving, under either definition, with all four cuts shown. Reporting the decay the budget model predicts would have been the more satisfying answer and it is not in the data.
7. The recommended place value is the lower quartile (0.23), not the median (0.38), for a floor in judge.py and the settlement — because the lane's upside depends on manager attention no code models.
8. The document states plainly that a lane does not beat an ordinary held man (0.38 vs 0.43), so the code change cannot be read as an argument for more 2-for-1s.
9. The per-add floor is left alone, and the budget is enforced instead of a discount being invented.
10. 2020 is pooled and flagged; --drop-2020 and --seasons expose the variants.
11. Only docs/streaming_lane.md and scripts/streaming_lane.py are touched. Nothing in app/ changed: this is a measurement and a recommendation, not an implementation.
12. The script is read-only. SELECT-only, no writes, no alembic.
13. The document was rewritten, not amended, after the per-place reconciliation — the headline figure and its unit both changed, and leaving both readings side by side would have invited the reader to take the larger one.

Applied

2026-09-22, f603775. The headline was taken into the code as revision R2 (docs/trades.md §7b, declared before the change was written). OPENED_PLACE = 0.38 now sits beside TYPICAL_PICKUP in app/scoring/replacement.py, with the lower quartile recorded as OPENED_PLACE_LOWER, and app.scoring.replacement.opened_places is the one rule three call sites read: app/pickups/judge.py's places_cost (so season_cost, judge and the trade evaluator's per-man number all follow), the trade evaluator's own headline in app/trades/evaluate.py (a scalar top-up over the best free agent's week, which is zero whenever that man already clears the lane), and app/scoring/moves.py's grade_move, which had been charging a flat ~0.07 in hindsight. An opened place is priced at the better of the man and the lane — never their sum, per the caution above — and the median was taken rather than Decision 7's conservative lower quartile, because the number is used as a floor under the best free agent rather than as a forecast (docs/trades.md §8). The second opened place takes Decision 6's finding at its word: no decay is measured, so it is priced at a single ordinary add rather than at a second lane.

What it was worth. The trade calibration re-ran once, the same day (docs/trades.md §7). The error it was aimed at — consolidating deals over-rated by +0.389 categories a week — is +0.103 with this figure in both engines, and re-pricing the forecast alone left it at +0.404, so the over-rating was mostly the hindsight grade charging 0.07 for a place the forecast filled with a real man. §2's limitation stands and is now load bearing: an opened place is invisible in daily_lineup_slots, so 0.38 is a floor measured on places found within a roster, applied to places that are genuinely empty.

Not taken. The recommendation in "Should the price fall with the number of lanes?" to gate the place value on adds_left is not built. 16.2% of team-periods have spent the budget and can run no lane at all, and the code still prices their opened places as though they could.

As built, 2026-09-22: 0.38 is this league's number, not the code's

OPENED_PLACE = 0.38 is no longer a constant every league is charged. It is this league's opened_place row in league_calibrations (docs/intake.md), measured by exactly the chain above -- measure_season per season, the per-place samples pooled, the median with its IQR beside it -- and it is what the code still uses for any league that has not been measured, which is the honest thing to call a default rather than a fact.

What changed here: load_occupancy and measure_season take a league_id, because a database now holds more than one league and a lane measured across two of them is a measurement of neither. Nothing else moved, and the number this document reports is the number the seeded row carries, to two decimals.

A second league's own lane is measured by intake_lane, in about 96 seconds over eight seasons here, and the payload keeps the IQR, the held 13th man and the ordinary held place beside the headline -- because the caution this document draws is from the comparison and not from the number.