9. What this would have cost to buy

A procurement comparison for the Manhattan Bridge / DUMBO rail-noise investigation. Companion to Document 8, usage and cost, which measured what this corpus cost to produce. This one asks the other half of the question: what would the same deliverable have cost to buy.

Interactive: procurement-dashboard.html

Data: rates.json - awards.json - procurement-data.json

Scripts: fetch_rates.py - fetch_awards.py - build_procurement_data.py

The claim this document exists to refuse

There is an obvious headline available here and it is wrong.

This investigation cost $437.35 in metered inference. A consultancy would have charged six figures for the same work. Therefore the saving is roughly three orders of magnitude.

That is withdrawn before it is made. It fails in three independent ways, and each failure is large enough on its own to invalidate the ratio.

It compares a measured number to an invented one. The $437.35 is a reconciled per-request ledger rated 5/5 VERIFIED. The six-figure comparator is an hours estimate multiplied by a rate. The rates are 5/5. The hours are 1/5 INVENTED. Nobody has ever built a comparable corpus under a timesheet that this programme could read. Dividing a measured numerator by an invented denominator produces an invented ratio wearing a measured number's clothes.

It omits the human term entirely. The billing ledger prices model inference. It does not price the person who specified the work, rejected outputs, caught the errors, and made every judgement call recorded in plan.md. That person's time is the single largest cost in this project and it appears in the ledger as exactly zero.

It compares different deliverables. What a firm sells includes things this corpus does not contain: professional liability, a licensed signature, a named partner who can be deposed, and - critically - field measurement. See what was not delivered, below. Comparing a desk study to a contracted engagement and calling the difference a saving is comparing a manuscript to a building.

This is the same over-claim this repository has already withdrawn three times: in Phase 9 (the park-phase finding), Phase 10 (the weekday-morning worst case) and Phase 12 (the fitted propagation model). It is recorded here in advance because the fourth occurrence would have been the easiest one to publish.

Three instruments, deliberately never averaged

They are reported side by side and never combined into a single figure. A and B disagree by a factor of about three, and that disagreement is the result of this document, not a defect in it. Averaging them would destroy the only informative thing here.

Instrument B: what the delivered scope prices at

The rate source

GSA publishes awarded ceiling rates for Multiple Award Schedule contracts at a login-free endpoint. Index ceilingrates-2026-08-04_02-00-02. The rates are awarded ceilings on federal schedule contracts, inclusive of the Industrial Funding Fee, for the current contract year.

Cross-validated against a large schedule holder's own published rate card (SIN 54151S, current contract year). The card carries the note:

Prices include the 0.75% Industrial Funding Fee (IFF).

Year 10 OFFSITE, from the holder's own PDF, matched against the GSA index:

Identical to the cent across seven independently checked categories. That raises the index from a convenient aggregate to a 5/5 VERIFIED instrument and removes any need to parse vendor PDFs to build the model.

The rate ladder

Three rungs, and one deliberately missing.

There is no nearshore rung, and its absence is deliberate. Eight sources were retrieved for nearshore and offshore blended rates. Every one of them was a marketing page published by a firm selling nearshore delivery. Not one carried a method, a sample size, or a definition of what was being averaged. A number of that provenance rated against this repository's own rubric is 1/5, and 1/5 numbers do not go in tables next to 5/5 numbers without a warning larger than the number. The whole-schedule 10th percentile is substituted and labelled as exactly what it is: the cheapest decile of firms that hold this specific schedule, not an offshore quote.

The scope, as work packages

Six packages, each sized from something countable in the repository - words of document, lines of code, appraised sources, artifacts - and each given a productivity band rather than a point rate.

Plus two overheads applied as percentages: review and quality assurance at 10-20% (101 - 524 h, priced as a subject matter expert), and engagement management at 12-22% (122 - 576 h, priced as a project manager).

Total: 1,235 to 3,717 hours. A span of three to one. That span is honest - it is the width of the productivity assumptions - and it is the reason no point estimate appears anywhere in this document.

The result

What moves it

Sensitivity, widest band first, as a share of that package's own midpoint:

The largest single uncertainty is not the research and not the code. It is how much review the work receives. That is worth stating plainly: the term that most changes what this would have cost to buy is the term that most changes whether it would have been right.

What was not delivered

The comparison above prices a desk study. This corpus contains no measured acoustic data of its own, no counted pedestrians, no licensed design and no legal opinion. A firm asked for "a noise study of the Manhattan Bridge in DUMBO" would deliver those, and a client would expect them.

Published here as a distinction between the two columns:

None of these four disciplines has a matching category on the holder's own rate card. [...] For the upper-quartile rung the model substitutes the whole-schedule 75th percentile for that discipline and says so in the data file. That is a weaker figure than the delivered-scope upper-quartile column, which is drawn from matched, cent-for-cent verified categories.

That is withdrawn. There is no such distinction, because the delivered-scope column is not drawn from matched categories either. The model carried a lookup that consulted one holder's own published categories before falling back to the whole-schedule 75th percentile, and that lookup never returned a match once — not for any of the eight delivered disciplines and not for any of the four below. Every figure in every upper-quartile column, in both tables, has always been the whole-schedule 75th percentile.

The cause is the same taxonomy problem this document already reports from the other direction: holders publish internal job titles, not discipline names. The categories on an individual card read Cyber Programmer 1 and Business Functions Consultant 1, which no discipline regex can match, so the fallback was not a fallback — it was the only path.

No number moves. Every rate in both tables is the figure it always was; what changes is the description of where it came from, the name of the rung, and the field name in the data file, which is now rate_upper_why and reads whole-schedule 75th percentile for every discipline. The dead lookup has been removed rather than left in place, because a branch that never returns is a claim the model does not honour. The upper quartile remains a genuine 5/5 figure — it is drawn from the same GSA awarded-ceiling index as the other two rungs, and that index is independently cross-validated above. It is simply not, and never was, a vendor-specific number.

The model refuses to emit an all-in figure. Adding the two columns would produce a number describing a deliverable that does not exist, and it would look more authoritative than either column alone. The two are reported separately, permanently.

This is also the honest answer to "what did the money buy". The delivered column is what the metered inference substituted for. The not-delivered column is what it did not, and cannot.

Instrument A: what buyers actually paid

Bottom-up estimates are the weakest form of cost evidence, because the analyst choosing the hours already knows what answer they want. So the model carries a second, independent instrument that assumes no hours at all.

USASpending publishes obligated dollars on real federal contracts. Filtering on noise and acoustic study keywords across product-service codes B (special studies), C (architect and engineering) and R (professional services), award types A-D, 2015-01-01 to 2026-08-01:

These are dollars that changed hands. No hours are assumed, no productivity is invented, and no analyst chose them.

What a design practice charges to define a problem

Asked directly what an architecture or design-build practice would have charged for a study of this shape, the obvious move is to price it the way this document prices everything else: hours multiplied by a published schedule rate.

That move is unavailable, and the reason is the more interesting half of the answer. Architectural and engineering services are not bought on price. The statutory note to 40 U.S.C. §1103 is explicit that the schedule route is closed to them:

Architectural and engineering services (as defined in section 1102 of title 40, United States Code) 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, United States Code.

And chapter 11 selects on qualifications, with price entering only afterwards. §1103(d) directs the agency head to rank at least three firms by competence; §1104(a) then has the agency negotiate compensation with the firm already chosen. Price is not a selection criterion, so there is no awarded ceiling rate for architecture the way there is for a data engineer or a project manager.

This is the fifth time this corpus has run into the same shape: an absence in a classification system that makes a real thing unaskable. The others were the noise code's missing rail category inherited by 311 and SONYC, the blank cell on form 1204-a, the twelve thousand project managers against seven acoustical engineers, and the schedule holders who publish job titles rather than disciplines. This one is the strongest of the five, because it is not an oversight — it is deliberate federal policy, and it is written down.

So the rung is priced with instrument A instead

A second population of USASpending awards: product-service code C, award descriptions naming a study rather than a building, same period and award types as the noise population. It is kept separate and never pooled — the two answer different questions, and a combined percentile would answer neither.

3.6 times the median federal noise study. A design practice is not paid to write a report; it is paid to run a process that ends in a decision, and that costs more.

The population is dominated by one award description — the design charrette, 108 of 123 — which is a practice being paid to sit with a client and establish 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 on the dashboard so a reader can judge that rather than take it. The charrette sub-population medians at $238,452, within 2.3% of the whole.

And a published rate ladder, since the schedule has none

Public bodies that appoint an A-E panel publish the rate schedule as an exhibit to the appointing resolution. One such resolution — a New York public authority's 2025-2027 A-E appointment, five firms, each with its own ladder — gives:

5/5 as published rates. 2/5 as a guide to the New York City market. An upstate county panel is not a Brooklyn signature practice; the direction of the difference is obvious and its size is not. The ladder is therefore used for one purpose only — as a published reference point beside a single stated rate — and it is not multiplied by anything.

The disagreement, quantified and unreconciled

Convert the awards to implied hours at the model's own blended rate of $147.31/h, against a bottom-up midpoint of 2,846 hours:

The bottom-up low lands at about 99% of the award 75th percentile. On the population of federal noise studies actually purchased, a corpus of this size prices like an upper-quartile engagement.

Two readings survive this, and neither is chosen:

1. The hours are too generous. Nine documents and seven artifacts is a lot of output, but output is not effort, and the productivity bands were set by the same person who wanted the answer to be large.
2. The awards buy narrower deliverables. A $68,441 federal noise study is very likely one site, one campaign, one report - not a nine-document corpus with six interactive artifacts and a method register.

Both are plausible. Resolving between them requires a timesheet from a firm that has built a comparable corpus, and no such document was found. The disagreement is therefore published as the finding rather than smoothed into a range.

The transparency asymmetry

One large schedule holder's published GSA Advantage price file is five pages long and contains no rates at all. The file's final page carries a legal notice restricting internal use of the rate card.

The same contract's awarded labour categories appear in GSA's own ceiling-rate index — dozens of them, retrievable without a login, an account or a key. Both facts are true simultaneously. The restriction is real and the data is public.

This is not an accusation; it is a note about where public information lives. A researcher who accepts the vendor's own published file as the authoritative public record concludes that the rates are not disclosed. A researcher who queries the awarding agency's index concludes that they are. The second researcher is right, and would have had no reason to look if the first document had not so specifically declined to help.

The same shape as the SONYC taxonomy finding in Document 6 and the 1204-a blank cell in Document 1: the absence is inherited, and it is findable elsewhere if you stop asking the party with the least interest in answering.

Twelve thousand project managers and seven acoustical engineers

Labour categories on the GSA schedule matching each discipline:

Seven. Across the entire federal professional-services schedule, from five firms, ranging $68.95 to $172.65.

This is a procurement-side instance of the same taxonomy blindness this programme documented on the complaint side. A federal buyer who wants an acoustical engineer on schedule has seven categories to choose from and will very likely buy a "subject matter expert" or an "environmental scientist" instead - because those categories exist in abundance and the specific one barely does. The work then gets done, and gets recorded as something else.

Consequence for Instrument A: the 56-award population is almost certainly an undercount, because acoustic work bought under a generic category is invisible to a keyword-and-PSC search. Direction of bias is known; magnitude is not. Stated, not corrected.

Note also that a project manager costs more per hour than an acoustical engineer ($151.95 against $131.21). The scarce specialist is cheaper than the abundant coordinator.

The human term

What the billing ledger does not price. Every figure in this section is derived at build time by build_procurement_data.py from the usage ledger and the rate files. Three of them used to be typed here by hand, and they went stale by close to a factor of two the moment the engagement continued past the day they were written.

The active-attention figure is 2/5 UNVERIFIED. It is derived from request timestamps and inter-request gaps, not from a stopwatch, and it cannot distinguish a person reading output carefully from a person who walked away. The band is the same measure at idle cutoffs of 120 s and 300 s.

The same hours, at four different rates

One of those four rows is not evidence, and it is the first one. The operator's rate is a statement by the person who did the work about what that person's hour is worth. It is rated 5/5 as a stated rate and 0/5 as a market observation — the same way this repository rates every operator statement of intent. It is 3.64 times the highest rate on the five-firm published A-E schedule above, and 3.91 times the subject-matter-expert upper quartile on the GSA index. Those ratios are printed rather than smoothed away, because a stated rate sitting nearly four times above every published comparator is a claim a reader is entitled to discount, and burying it in a blended average would deny them that.

The four are never averaged. Averaging a valuation with three observations launders the valuation, which is the same failure this document refuses at the level of the three instruments.

What survives whichever rate is picked

At every rate tested, including the lowest, the human hours cost several times more than the model:

The direction of this work costs between 4.8 and 30.5 times the entire metered inference bill of $437.35, and the multiple never falls below five.

Any framing of this project's economics that reports only the inference cost is understating it by close to an order of magnitude at best, and by a factor of thirty at the operator's own valuation — before any allowance for the fact that the direction required domain judgement the model demonstrably did not supply. Several claims in this repository were withdrawn because a human noticed they were wrong.

Where this document is likely to be wrong

1. The hours are invented. This is the largest weakness by a wide margin and it is first deliberately. Six packages, each with a productivity band, each band chosen by the author. The bands are wide (three to one overall) precisely because they are not knowledge, but a wide invented band is still an invented band. Nothing in the delivered-scope table is stronger than this term, and every dollar figure inherits it.

2. The scope map is arguable. Whether "interactive artifacts" is one package or four, and whether a research corpus needs 10% or 20% review, are judgement calls that swing the answer by more than $150,000 combined. A different analyst with the same rates would produce a different number.

3. GSA schedule rates are ceilings, not prices. Awarded ceiling rates are the maximum a holder may charge on that schedule. Actual task-order pricing is routinely discounted below ceiling, and the discount is not public. The consultancy rung is therefore an upper bound on that rung, not a quote.

4. The award population is filtered by keyword and PSC. As set out above, acoustic work bought under generic labour categories does not appear. n=56 is a floor. It is also federal-only: this is a municipal problem, and city and state procurement is not in this dataset at all.

5. Federal rates are not New York City rates. The GSA schedule prices work for federal buyers under federal contracting overhead. A private client in New York, or the MTA under its own procurement rules, faces a different market. The direction of the difference is not known.

6. The active-human-hours figure is inferred from timestamps. 2/5. A gap between requests is not proof of attention, and simultaneous attention to something else is invisible. This term is load-bearing for the human-cost comparison in the section above.

7. "Not delivered" is a list, not an estimate of a designed campaign. The four items were sized by analogy, not scoped. A real field acoustic campaign at this site might cost substantially more than the band shown, because working over a live four-track river crossing owned by the MTA involves access, flagging and insurance that a generic estimate does not capture.

8. This document compares a finished artifact to a hypothetical engagement, and finished artifacts flatter themselves. The corpus that exists is the one that survived. A firm's engagement would have included work this project simply abandoned, and this project's abandoned work does not appear in its own hours estimate either. The asymmetry is not quantified.

The question this opens

Q55. What does a research corpus cost, and against what? This document establishes that the question has no single answer, because the three available instruments disagree by a factor of about three and each measures something different. A per-request billing ledger measures inference and nothing else. A bottom-up build at published rates measures an analyst's beliefs about productivity. Obligated contract dollars measure what buyers with budgets actually chose to pay for deliverables nobody here has seen.

The open part is the one that would make this transferable: is there any published, methodologically stated dataset of hours-to-deliverable for research and analysis engagements? Eight searches for nearshore blended rates returned eight vendor marketing pages. Searches for consultancy timesheet data returned nothing usable at all. If such a dataset exists, it collapses the largest weakness in this document — the 1/5 hours term — into a 4/5 or 5/5 one, and converts a disagreement into a measurement. If it does not exist, that absence is itself worth recording, because it means every published claim about what AI-assisted research saves is resting on the same invented denominator this document refuses to hide. Method 37 is executed; this is what executing it found that it could not answer.

Reproducing this

python procurement/fetch_rates.py           # ~4 min, 346 API calls, no key
python procurement/fetch_awards.py          # ~1 min, no key
usage-calc build                            # reads the local billing store
python procurement/build_procurement_data.py

The first three write rates.json, awards.json and usage/usage-data.json. The fourth combines them, writes procurement-data.json, and injects the result into procurement-dashboard.html.

What a fresh clone cannot re-run. Stated so the reproducibility claim above is not larger than it is. fetch_rates.py sweeps the schedule by discipline and, separately, can sweep it by a named holder. The discipline sweep is the one every figure in this document rests on and it runs from a clean checkout with no configuration. The holder sweep does not: its query terms are read from an untracked procurement/vendor_terms.json, and with that file absent — which is the normal state of this repository — those groups are simply skipped and nothing downstream changes, because no rate published here is derived from a holder-specific query. The cross-validation described in The rate source was checked by hand against a published card, not by that code path.

rates.json carries no vendor name and no contract number for any row. A contract number resolves to a holder in a single search, so the two are one field for this purpose and neither is carried. This is a study of published rates and not of firms, and the file is exactly as useful for that without them: every statistic here is computed over discipline and price, and both survive. It does mean a reader cannot audit which holders a percentile was drawn from — only that it was drawn from every holder on the schedule, which is what the rung now claims.

Two silent failure modes in the GSA endpoint are guarded in fetch_rates.py and both were hit live during development:

• hits.total.value saturates at 10,000 and reports relation: "gte". Taken at face value it undercounts project managers by 2,913.
• page_size caps near 1,000 and default ordering is price ascending, so 1,000 rows of a 12,913-row population is the cheapest thousand, not a sample.

Both are avoided by sweeping the price axis in bands. price_range is inclusive at both ends, so rows priced exactly on a band boundary return twice and are deduplicated on record id - a naive three-band probe reports 12,916 project managers where the deduplicated sweep finds 12,913.

One more, in fetch_awards.py: curl.exe -d $body in PowerShell mangles JSON bodies and returns a parse error from the API. Use a real HTTP client.

← All research · Overview · Source on GitHub · Built from 6f6b90d on 2026-08-11.