Silencing the Span

Manhattan Bridge noise investigationusage › procurement comparison

What this would have cost to buy

A scope-matched comparison of this investigation against published consultancy rate cards and against what federal buyers actually paid for noise studies. Three instruments, reported side by side and never averaged.

Rates are GSA Multiple Award Schedule ceiling rates, fetched from the government's own index. Award values are obligated dollars on real contracts. The hours are this repository's own estimate and are the weakest number on the page.

The headline this model refuses to print

Written out so that it can be argued with rather than implied.

This investigation cost $236.61 in inference. A consultancy would have charged several hundred thousand dollars for the same work. The saving is three orders of magnitude.

Every clause is defensible and the sentence is still wrong, for the same reason three earlier claims in this repository were withdrawn: it compares a measured number against an estimated one and quietly changes the scope between them.

The honest statement is narrower and duller: a desk corpus of this size prices like an upper-quartile federal noise study, and this one was produced for the cost of a takeaway meal plus about three days of one person's attention.

What was actually delivered

Counted from the repository, not estimated. Injected data payloads are excluded from source-line counts because they are generated rather than written.

On the source count. This corpus cites by name and quoted locus rather than by hyperlink: there are distinct external URLs in nine documents and lines carrying a 1/5-to-5/5 source rating. Counting hyperlinks would have reported seven sources for a body of work built on several hundred. The rule used here over-counts, because the rubric's own definitions carry ratings, and under-counts where one row appraises several documents. Neither is corrected.

Three rate rungs, from one published instrument

All three come from the same GSA index, so they are comparable to each other by construction. A ceiling rate is the maximum a vendor may charge under the schedule, not a price anybody was observed to pay.

The index was checked against a schedule holder's own published card

A large schedule holder publishes its rate card as a PDF attached to its GSA MAS contract. The GSA index reproduces it to the cent for the current contract year, on every category checked.

Contract year 10 runs 12 July 2026 to 11 July 2027, which is the year the index was built in. Prices in the card include the 0.75 per cent Industrial Funding Fee. OFFSITE is the contractor's own premises, which is where desk research happens.

Instrument B — hours multiplied by a rate 1/5 for the hours

Each package converts a measured quantity into an hours range. The measurement is counted; the divisor is invented. Both ends are printed so a reader can substitute their own.

Which assumption actually moves the answer

Band width is what that package's hours uncertainty is worth in dollars at the whole-schedule median rate. Sorted widest first.

The hours are the weakest number on this page. They are not derived from a timesheet, a productivity study or a published estimation model. They are a stated guess carried as a band. If you disagree with them, the package table above gives you every input needed to disagree precisely.

Instrument A — what buyers actually paid 5/5 VERIFIED

Federal prime awards for noise and acoustic studies, selected by keyword against the award description and by Product/Service Code. Obligated dollars, not list prices, and no hours are assumed anywhere in this instrument.

The distribution

What the money bought, at and above the lower quartile

The most relevant buyer is missing. USASpending records federal awards. The MTA, NYCDOT and NYCDEP are state and city bodies and do not appear in it at all, so the one organisation that would actually commission a Manhattan Bridge noise study is absent from the only dataset that records what such studies cost. Nothing here was hand-removed either: a bombing-range noise study is not a bridge, and it is still in the population, because deleting the rows that spoil an average is the failure this repository exists to avoid.

What a design practice charges to define a problem 5/5 VERIFIED

Asked what an architecture or design-build practice would have charged for a study like this one. The answer needed a different instrument, and the reason why is the more interesting half.

Instrument B is not available for this rung, and not by accident. Every other discipline on this page is priced by multiplying hours by a published GSA awarded ceiling rate. Architectural and engineering services cannot be bought that way. The statutory note to 40 U.S.C. §1103 reads that A-E services shall not be offered under multiple-award schedule contracts entered into by the Administrator of General Services … unless such services … are awarded in accordance with the selection procedures set forth in chapter 11 of title 40 — and those procedures select on qualifications, with compensation negotiated only after the firm has been chosen (§1103(d), §1104(a)). There is no A-E ceiling rate to multiply by, because price is not what the buyer chooses on.

So this rung is priced with instrument A instead: obligated dollars on A-E awards whose description names a study rather than a building. It is a second population, kept separate from the noise-study population and never pooled with it.

Where the population came from

The population is dominated by the design charrette — a practice being paid to sit with a client and work out what the problem is before anything is designed. That is the closest commercial analogue this study has, so the dominance is the population converging on the right thing rather than a contaminant. It is broken out above so a reader can judge that rather than take it.

What the money bought

And a published rate ladder, since the schedule has none

An upstate county panel is not the New York City market. These rates are 5/5 as published rates and 2/5 as a guide to what a signature design practice in Brooklyn would charge. The direction of the difference is obvious and its size is not, so the ladder is used for one purpose only: as a published reference point beside a single stated rate. It is not multiplied by anything.

The human term, and one rate that is not a market observation

Every figure in this card is derived at build time from the usage ledger and the rate files. Three of them used to be typed into the document by hand, and they went stale the moment the engagement continued past the day they were written.

The same hours, at four different rates

One of those four rows is not evidence. The operator's rate is a statement by the person who did the work about what that person's hour is worth. It is 5/5 as a stated rate and 0/5 as a market observation, and it is times the highest rate on the published five-firm A-E schedule above. It is carried here so the gap is visible rather than absorbed, and it is never averaged with the three rows that are observations.

The active-hours band is 2/5 UNVERIFIED. It is inferred from gaps between request timestamps, not from a stopwatch, and it cannot tell a person reading output carefully from a person who walked away. It is load-bearing for every figure in this card, so the card is no stronger than that inference — whichever rate is applied to it.

The two instruments disagree, and that is the result

Reported as a number rather than described, and not reconciled.

At the blended midpoint rate of per hour, the median federal noise study buys about hours. The bottom-up estimate for this corpus starts at hours. Two readings survive that gap and both are live:

Distinguishing them needs a timesheet from a firm that has built something of this shape, and no such figure was found. It is recorded as an open question rather than resolved by preference.

The scope that was not delivered

Priced the same way, and rated the same way: the rates are published, the hours are invented. This column is the honest counterweight to the first one.

The part of this work that could not have been done this way is precisely the part that is missing.

No language model can stand under a bridge with a calibrated microphone, count pedestrians at a cordon, seal a structural drawing, or give a legal opinion that anybody may rely on. Those four items are the binding constraint on every absolute claim in this investigation, and they are the items an agency engagement would have delivered.

Twelve thousand project managers and seven acoustical engineers

A count nobody asked for, which fell out of the rate fetch.

These are counts of distinct awarded labour categories across every holder of a GSA Multiple Award Schedule contract, which is the federal government's largest professional-services purchasing vehicle. The whole schedule carries acoustical categories, priced between and an hour.

This is not an argument that acoustics is undervalued. It is an observation about what is procurable at speed. An agency that decides on Monday it needs a project manager has twelve thousand priced options. An agency that decides it needs an acoustical engineer has seven, from five firms. That asymmetry is the same shape as the finding in Document 6: the instruments a city has for buying attention determine which problems get attention, and rail noise falls through the same kind of gap twice.

One firm's rate card is restricted and published at the same time

Found while trying to price the comparison, and worth recording.

One large schedule holder's GSA price list is a five-page PDF that contains no rates. Its final page carries a legal notice restricting internal use of the rate card.

GSA publishes of that contract's awarded labour categories in a search index that needs no login, priced from to an hour. Both statements are true at once: the vendor restricts its card, and the buyer publishes the rates it awarded under it.

Why this belongs in a noise study. It does not, on its own. It is here because this comparison could not have been made from vendor disclosure alone, and a reader deciding whether to trust the numbers above should know which of them came from a firm and which came from a government.

Instrument C — what it actually cost 5/5 VERIFIED

From the client's own billing telemetry. The full derivation, and the one inference it still rests on, are in the usage dashboard.

The wall-clock span is not working time: it is the interval from the first request to the last, across three calendar days, and most of it is the human being elsewhere. Inference time is the sum of model latency and is a better lower bound on machine work than on human work. Neither is a timesheet, and no timesheet was kept.

The human hours are not measured and are not zero. Every question asked, every direction changed, every claim rejected and every withdrawal ordered came from a person. That labour is the input this page cannot price, and any ratio computed from the $236.61 alone silently sets it to zero.

Where this comparison is likely to be wrong

  1. The hours are invented. Named first because it is the largest error term by a distance. The sensitivity table says which packages carry it.
  2. A ceiling rate is not a paid rate. Agencies compete task orders below the ceiling. The bottom-up column is therefore biased HIGH by an unknown amount, probably tens of percent, and no discount was applied because inventing one would trade a published number for a guess.
  3. A federal rate is not a commercial rate. Both directions are arguable: federal work carries compliance overhead that raises rates, and competitive pressure on a schedule that lowers them. No adjustment was made.
  4. The award population is federal only. The MTA, NYCDOT and NYCDEP are absent. This is the largest gap in instrument A.
  5. The award keyword filter is crude. It selects on description text and Product/Service Code. It will miss studies whose descriptions do not say "noise", and it admits work in other domains entirely.
  6. The nearshore rung is missing and was left out deliberately. Every figure found for nearshore delivery was published by a firm selling nearshore delivery. No independent survey with a quotable method was located. Rather than print a vendor's marketing range beside two government datasets, the whole-schedule tenth percentile is used as the low rung, and it is labelled as what it is.
  7. The discipline decomposition is a choice. Six packages and two overheads is one way to cut this work. A firm would cut it differently and the answer would move.
  8. The delivered corpus may not be worth building. The comparison prices producing it. It says nothing about whether producing it was the right thing to do, and several methods registered in this repository would have been better value than any of it.

How to reproduce every number on this page

  1. python procurement/fetch_rates.py — sweeps the GSA ceiling-rate index in price bands and writes rates.json.
  2. python procurement/fetch_awards.py — pulls federal noise-study awards from USASpending and writes awards.json.
  3. usage-calc build — reads the local billing store and writes usage-data.json.
  4. python procurement/build_procurement_data.py — combines all three, applies the hours band, and injects this page.

Two traps on the rate endpoint, both silent

Recorded here because either one produces a plausible wrong answer with no error.

Both are avoided by sweeping the price axis in bands, which return exact counts, and subdividing any band the page cap cannot hold. Bands are inclusive at both ends, so rows priced exactly on a split point are returned twice and are deduplicated on the index document id.