The research
Twelve documents and roughly 130,469 words, in the order they were written. Each asks one question, answers as much of it as the evidence allows, and ends with a section on where it is most likely to be wrong. They are meant to be read out of order — start with whichever question is yours.
Documents
12
each one question
Methods specified
49
7 executed or partly
Questions derived
68
Q1 to Q68, none rhetorical
Claims withdrawn
20
left standing in the text
The investigation
Repository README 18,765 words
The argument in brief, the document index, and the method register with its honest status column.
The register with its honest status column: 49 methods specified, 7 executed or partly. It does not hide the ratio.
1. Idea and concept 20,979 words
What is the problem? Defines it from agency evidence, establishes who is responsible under what law, and derives the questions nobody has asked of this site.
The MTA measured the noise, in DUMBO, at named addresses, and published the levels. The levels are high and nothing followed.
2. Precedent and materials 11,470 words
What has the world already built? Elevated-transit noise mitigation in Japan, China, Sweden, Germany, Hong Kong, Australia and Chicago - and what actually transfers to a 1909 suspension bridge.
Floating slab track, constrained-layer damping, rail dampers and parapet treatments all exist and work — on structures that were designed to carry them. This one was built in 1909 and its load budget is the constraint everything else answers to.
3. Williamsburg comparator 12,757 words
There is a second bridge with the same owner, operator, rolling stock and statute. What does it already tell us, and what would measuring it establish?
A second bridge with the same owner, operator, rolling stock and statute is the cheapest control this problem will ever get, and nobody has measured the pair.
4. Visual model framework 12,458 words
Every argument in the first three documents is an argument about a cross-section nobody has drawn. Can that drawing be built from open data - and made to admit what it does not know?
A drawing can be built from open data, and the useful part is what it refuses to draw: turn every provenance filter off and the viewport goes empty. That is the state of public knowledge.
5. Field capture protocol 7,406 words
Every acoustic claim beyond the published levels is invented. Can a consumer phone fix that this month?
Five captures, a phone already owned, and no permission needed. C2, the temporal envelope, is the one measurement that could prove this repository wrong.
6. Community evidence audit 10,300 words
The people who live under it have been complaining since 2008. What have they already recorded, and why can nobody find it?
Within 500 m of the MTA's own measurement, residents filed 4,055 noise complaints since 2020 and not one of them can be about the train. There is no category for it.
7. Data collection 10,779 words
How many trains, how many people, and for how long? Runnable scripts against MTA and NYC open data - and the traps that each silently produce a plausible wrong number.
Six scripts, and four traps that each silently produce a plausible wrong number. The turnstile feed's entries field means people leaving, not arriving, and reading it the natural way inverts the day.
10. Field media 8,251 words
The first material here that was not retrieved from somebody else: two days of phone video, stills, by-product audio and stopwatch laps under the bridge - and a v1.1 that withdraws v1.0's headline in place.
The first material here that was captured rather than retrieved — and v1.1 withdraws v1.0's headline, because two correctly-computed datasets were joined that had never been joined in the field.
11. Observation protocol 5,188 words
Everything about trains here is measured and almost everything about people is invented. Ten things a person walking through DUMBO can count that would move a number - each with what it moves, what it would be rated, and what a result of zero would mean.
An absence, counted, is data. An absence, remembered, is not. Ten countable observations, each naming the invented number it would replace — and a submission path built so the provenance fields cannot be skipped.
The work about the work
Two documents that study this investigation rather than the bridge: what it cost to produce, measured from the tool's own per-request log, and what the same deliverable would have cost to buy. Both are held to the same evidence standard as everything else here, and both withdraw a headline claim on their own front page.
Looking for the interactive pieces instead? They are on the
front page, where a reader who has never heard of this problem will find them first. This page is the evidence underneath them.
Where the investigation stands
Eleven research documents, ten interactive artifacts, nine runnable scripts and roughly 130,469 words — against 49 specified methods of which six have been executed and one partially. The gap between those two halves of the sentence is the honest summary of this programme, and it is the reason this section leads the research rather than the front page: the artifacts are the most finished thing here and the least load-bearing.
| State | Count | What it means |
|---|
| Executed | 7 of 49 methods | Method 27, Method 29, Method 30, Method 32, Method 33, Method 37, Method 38. Every one of them changed something, and two contradicted claims this repository had already published. |
| Tooling built | 1 | Method 26 — the code is written and verified against the live feed, but the run has not been done. |
| Not started | 41 | Proposals. Several of the cheapest are also the most load-bearing, and their being undone is why so many claims here are hedged. |
| Withdrawn | 20 explicit retractions | Claims this programme published and then disproved. They are left visible in the text rather than deleted, which is the point. |
| Open questions | Q1–Q68 | Numbered, attributed to a document, and none of them rhetorical. Q42 is the highest-value one and a lawyer could settle it in a day. |
The one thing to carry away from everything above. Nobody from this programme has stood in Brooklyn Bridge Park with an instrument. Every acoustic level quoted here was measured by someone else and published; every acoustic level modelled here is synthetic. No option in any document is recommended for procurement, and none of it is peer-reviewed.
Work: done, and to be done
A mark here means the work was finished, not that the question it was meant to answer is settled — in three cases below the finished work is what established that the question is harder than it looked. The ordering of the open queue is a judgement; the state of every numbered method is read out of the register rather than asserted here, and this page will not build if the two disagree.
DoneThe work was done, the result was not obtainedNot started
Done
11 of 31 items on this page
Done
Method 38 - the first captured field material
One afternoon, a phone already owned
Two days of video, stills, by-product audio and stopwatch laps under the bridge - the first material in this repository that was captured rather than retrieved. It delivered the first photographic record of the echo chamber the repository had until then only modelled from surveyed footprints, the geometry of every capture, and a rate that agrees with the timetable inside an interval wide enough to say so plainly. It also produced this repository's fastest withdrawal: v1.0 used an audio duty cycle to refute a reading of the stopwatch, and v1.1 withdrew that one day later, because the two are independent samples 62.4 minutes apart. Every number in the withdrawn claim was computed correctly - what was wrong was joining two datasets that were never joined in the field, which no check here can catch.
Done
Method 33 - the cost of the study itself
One SQLite read, free
A per-request ledger of what producing this repository consumed, priced by billing channel from the tool's own log. It opens by withdrawing its own first conclusion - that no such data existed - which had been reached from the wrong file and was three minutes from being published as the finding. It reports no measured energy at all, because none reaches a client.
Done
Method 32 - the corridor geometry
Two open datasets, no key, about a minute
The canyon under the bridge, drawn from 960 surveyed building footprints and 2,826 OpenStreetMap ways rather than sketched or traced off a proprietary basemap. It produced an independent check on the bridge alignment - 2.3° off the bearing this repository had digitised by eye - and established that the one object in the frame nobody has surveyed is the bridge itself.
Done
Method 29 - the cohort survival model
About six minutes of arithmetic
Fits four population cohorts to the observed departure curve. It reports its own failure: the sweep found many parameter sets that fit equally well and imply materially different exposure, so the model establishes that dwell time cannot be inferred from arrival and departure counts alone. That is why Method 28 sits at the top of the list below.
Done
Method 30 - the agentic population model
Built, not run as a measurement
Personas, family groups, itineraries, ingress and egress points, and dose accumulated along a path. Built as a mechanism demonstration and labelled as one on its own face, because every itinerary in it is invented. Its most useful output was a negative result: the propagation model over the four MTA points could not be fitted. A fifth rung now models who is standing in the dose and applies no decibel penalty to any of them - a per-class adjustment would be a fabricated exposure-response function wearing the costume of a measurement.
Done
The acoustic demonstration
Synthesised, and labelled synthetic in the interface
A train approaching, passing and departing at the measured decibel difference against each receptor's own background, then running continuously at the real headway. It is synthesis, not a recording, and the page says so before it says anything else.
Done
The community evidence audit
Searching, and finding nothing
Reddit, Freesound, the NYU SONYC corpus, NYC Open Data, local press and petitions, searched for crowd-sourced recordings. None exist - and the audit found the structural reason: three separate instruments for recording city noise and no rail category in any of them.
Partial
Method 27 - count the denominator
Four API pulls, about a minute
Arrival rate, walkway flow and resident count from four public datasets, two of which agree to within 1.77% by different methods. It disproved this repository's own claim that 08:00 is the worst hour; the real peak is 14:00. Marked partial because it produces λ and not W - the arrival rate exists, the dwell time does not.
Partial
Method 26 - the traversal census tooling
Written and verified against the live feed
The poller is built and correct. The week-long run has not been done, so this is a mark on the code and not on the result - which is why the census still appears in the queue below.
Partial
The field capture protocol
Written, and now partly attempted
A complete protocol for a consumer phone targeting the four things the MTA's five-number table discarded. One session has now been run against it and scored honestly: C1 not satisfied, C2 partly and not usefully, C3 partly, C4 and C5 not attempted. The gap between what the protocol asked for and what a first outing produced is itself the useful part.
Done
Method 37 - price the same deliverable against the market
Two free federal APIs, no key
What this would have cost to buy, from three instruments reported side by side and never averaged: 56 federal noise-study awards, a bottom-up build at GSA awarded ceiling rates verified cent-for-cent against a vendor's own card, and the metered inference ledger. Its result is the disagreement between them, not a number - and it withdraws the obvious headline before making it, because the numerator would be measured and the denominator invented.
To be done
20 of 31 items on this page
Blocking
Method 28 - measure dwell time
A few hours with a clicker, repeated over several sessions
Presence is L = λW. Method 27 produced λ. Without W there is no absolute exposure figure, and the cohort model proved that no amount of further arithmetic will substitute for measuring it. This is the single blocking unknown for any design-build case. Method 42 rides along with it at no extra cost — the same observer, at the same cordon, tallying prams, dogs, mobility aids and apparent age band. Record individual crossings, not only a total. The arrival-process result in the agent model turns on the bottom of the dwell distribution rather than its mean, and a running tally throws that away.
Cheap & high value
Method 48 - re-derive every digitised position, and write the rule
A desk afternoon; the surveyed ground is already in the repository
The walkable model tests every inherited coordinate against 2,342 surveyed building footprints. Two places the population model marks as outdoors fall inside a building, one of them inside an 81 m tower, and one MTA point named for a street intersection resolves inside a footprint. For an outdoor place that is not an approximation — it is a coordinate that cannot be right. Nothing has been moved, and moving the points is not the method. The method is writing down a rule — something like the centre of the publicly accessible footway area nearest the named feature, snapped to the network — which is checkable, reproducible and can be wrong in a statable way. An eye-placed point is none of those, and every position in this repository is currently eye-placed.
Cheap & high value
Method 42 - count the corridor by attribute, not just by head
The same session as Method 28, one extra clicker
Every susceptibility share in the agent model is a national prevalence applied flat to a corridor nobody has ever counted by any attribute at all, and each is rated 1/5 on whether it applies here. A flat rate is knowably wrong at a place whose purpose selects for a class, and this corridor is full of them: a dog run, a carousel, a lawn, and the residential blocks at Farragut Houses. The honest half of the method is what it cannot do: an age band, a pram and a lead are observable; hyperacusis, autism and a cancer history are not. It must publish which shares it moved and which it left national, or it will read as having verified all of them.
Cheap & high value
Method 44 - the crowding half, and only if the whole of it is run
The feeder headways are one constant and a minute; the rest is a study
Three arrival processes were compared in the agent model at a fixed population, and multiplying peak bunching by 2.9 times moved the mean dose by 0.006 dB — never more than 0.047 dB on any seed — because dwell dominates headway. Contention is the one quantity that does respond, up about 18%. So this method has been demoted by a result rather than by a judgement, and it is listed here mostly to say so. bridge_schedule.py already computes the feeder headways if one constant is changed, and running that alone would produce a real number that reaches nothing. The part that matters is behavioural: whether a full bench sends someone toward the bridge or away from it.
Cheap & high value
Method 45 - count dogs boarding, with a denominator
Free, and it rides along with a commute somebody is already taking
The cheapest correction available in this repository. The agent model documents its dog rate as “near zero for anyone arriving by subway” and then assigns 0.02, 0.04 and 0.01 to the three subway-ingress personas. If the true rate is nearer 0.2%, the comment and the constants disagree by up to a factor of twenty — and nobody noticed, because nobody counted. A result of zero is the expected result and is fully publishable: no occurrence in n boardings bounds the rate at about 3/n, so 400 boardings observed across a month of ordinary commuting bound it below 0.75%, which is tighter than anything now in the model. The denominator is the whole method. A tally of dogs without one is an anecdote.
Cheap & high value
Method 47 - does a listener hear two crossings as one event
Ninety minutes standing still, no gear at all
A clicker tally of audible train events against the traversal count from the feed for the same window. This is what licenses comparing any observed event rate to a scheduled one — a comparison this repository has already made once, at 56.4/hr against 57.7/hr, on seven events whose interval runs 13.8 to 99.0. The schedule cannot settle it and never could: every departure in the feed falls on an exact :00 or :30 second, so the window in which two crossings merge is empty by construction, which is what forced the withdrawal of this repository's own merged-pair table. Take the count before looking at the feed, or ambiguous events will be resolved toward the number the timetable predicts.
Cheap & high value
Method 46 - the matched-pair exclusion count
Two observers for ninety minutes, or one observer on two days
The same cordon count under the deck and at a control two blocks outside the canyon, at the same hour. It is the only method specified here that could establish exclusion rather than exposure. Every susceptibility figure in the agent model is a flat national prevalence, which structurally cannot represent a person who does not come at all: a class that avoids the corridor is reported as low exposure rather than as exclusion, and that is the flattering direction. It is also the method most easily made to lie — two blocks in DUMBO differ in footway width, retail frontage and grade, so an unmatched control measures land use. A null is publishable; a positive is not causal.
Cheap & high value
Method 31 - the decay transect
One afternoon, free if a sound level meter is borrowed
Walk outward from the structure with a meter and establish where the affected zone ends. Nobody has ever determined that boundary, yet it sizes the denominator for every exposure figure here. Three possible outcomes and all three are informative, including the one that corrects this repository.
Cheap & high value
Method 21 - the taxonomy query
A database query
Confirm across the full 311 and SONYC taxonomies that no rail category exists anywhere in either. The finding is already evidenced; this closes the last route by which it could be wrong.
Cheap & high value
Method 34 - pin one cohort from outside the curve
One download of a free federal dataset
The cohort model is degenerate by construction, because a departure curve carries no labels: at 14:00 the fitted worker count ranges 3,629-5,490 and visitors 1,460-2,313 across 9,248 parameter sets that fit equally well. No better fitting narrows that. One exogenous number does - LEHD LODES gives jobs by census block, which pins workers and leaves visitors as a residual rather than a guess.
Cheap & high value
Q42 - preemption or merely unregulated?
One competent lawyer, one day
Is elevated rapid transit federally preempted from local noise regulation, or simply unregulated? These have opposite consequences for every remedy in the programme. The highest-value open question here.
Cheap & high value
Method 26 - the traversal census
Leave a script running for a week, unattended
The tooling is built and verified. It is now the only route to the coincidence distribution, since the schedule feed was shown to be quantised to 30 s and unable to answer it.
Fieldwork
Methods 39, 40 and 41 - the instrumented session
One afternoon plus one quiet hour, with gear already owned
Three questions that need no calibration, because all three are ratios or timings and an unknown microphone sensitivity cancels out of both.
Direction and speed from two timecode-synchronised recorders separated 50–80 m along the track axis, read off the order of the two level peaks.
The decay tail — the operator's “noise time on clock versus floor time” — which no instrument here has ever measured, because automatic gain control pushes gain up as a sound fades and so flattens exactly the thing being measured.
Simultaneous crossings, which the published schedule cannot answer at all: every departure in the feed falls on an exact :00 or :30 second, so the window in which two trains merge acoustically is empty by construction. The session card is
FIELD-KIT.md.
Fieldwork
Extend the walk: York Street to Fulton Ferry along the water
One afternoon on foot, the phone already owned
The drawn walk currently stops at the water's edge. That is not the walk people take. They come up from the York Street F platform, go straight down to the water, turn, and follow the shoreline past Jane's Carousel to Fulton Ferry Landing — passing under the bridge, out from under it, and back into its shadow. Extending the corridor drawing and the capture route along that full line puts the receptor path where the population actually is, and it crosses the one geometry this programme keeps asserting and has never walked end to end: where the canyon stops. Method 31's decay transect and this share a route.
Fieldwork
Captures C1 to C5 - the phone protocol
A Galaxy S23+, public ground, no permission and no funding
The only proposal in the programme that requires nothing the programme does not already have. One outing has been made and it satisfied none of C1, C4 or C5 and only part of C2 and C3. C2, the temporal envelope, remains the highest-value item: it is the one measurement that could establish that this repository's own derived result is wrong, and the first attempt showed why it needs a windscreen and a capture path with the compressor disabled. A shielded and metered session is planned; Q56 should be answered before it is taken, not after.
Cheap & high value
Q56 - is any duty cycle computable from consumer capture?
An afternoon on a bench, no equipment beyond the phone
Every duty figure computed from the field audio passed through automatic gain control, and the interaction has a sign: a compressor shrinks excursions above a clip's own median, so any such figure is biased low by an unknown amount. Play a signal of known duty cycle through a speaker, record it on the same handset, and the known input gives the answer. This should be run before the next capture, not after - a shielded, metered microphone still feeds the same compressor unless AudioSource.UNPROCESSED is explicitly enabled, so this test decides whether the next session needs its capture path changed to be worth taking.
Cheap & high value
Method 43 - what licenses comparing dB HL to dB(A)?
One afternoon at a desk, but strictly blocked on capture C1
Loudness discomfort centres near 100 dB HL for normal-hearing listeners and hyperacusis is commonly marked at 90 dB HL or below; the MTA measured 98.90 dB(A) peaks at the dog run. Those look comparable and they are not the same units - one is per-frequency, pure-tone and headphone-presented, the other broadband, free-field and weighted by a single fixed curve. Bridging them needs the third-octave spectrum of the actual sound, which does not exist for either bridge. That is the value of the question: it turns an open-ended plea for more evidence into one named missing measurement. It will not give a clean verdict even then - the defensible output is a bracket with its assumptions named, and a bracket that straddles the measured peak would be as informative as either clean result.
Needs review
Red-team the three newest results
Reading and arithmetic
Issues
#26 (is the cohort model's non-identifiability a finding or an artefact of an arbitrary threshold?),
#27 (does a model whose every input is invented belong in a repository built on quoted loci?) and
#28 (
was the propagation model genuinely unfittable, or merely digitised badly?).
Needs review
Read the five counter-citations
Library access
Five works surfaced during red-teaming that bear directly on Q1 to Q8 and were not read in full. They are listed in section 14 of the concept document. Anyone taking this forward should start there, not here.
Gating
Methods 0 and 1 - the structural and acoustic prerequisites
A records request and an engineer; then a two-season field campaign
The load rating at the track zone gates roughly half the option space, and source apportionment is the programme's stated prerequisite - nothing downstream is non-arbitrary without it. Expensive, unavoidable, and the reason nothing here is recommended for procurement.
The count above is of items on this page, and is not the state of the programme. The register holds 49 methods, of which 7 have been executed or partially executed. Anything that reads as encouraging progress here should be read against that ratio.
The cheapest useful contribution is a recording. If you have ever recorded a train crossing the Manhattan Bridge from Brooklyn Bridge Park, DUMBO or the Williamsburg Bridge walkway, that file is more useful to this programme than anything currently in it. The bar is far lower than people assume: spectral shape and event timing survive an uncalibrated phone.
The research documents
Rendered here for reading. The markdown files in the repository remain authoritative, and each rendered page links back to its source.
| Document | What it asks | Words |
|---|
Repository README
README.md | The argument in brief, the document index, and the method register with its honest status column. | 18,765 |
1. Idea and concept
IDEA-CONCEPT.md | What is the problem? Defines it from agency evidence, establishes who is responsible under what law, and derives the questions nobody has asked of this site. | 20,979 |
2. Precedent and materials
PRECEDENT-AND-MATERIALS.md | What has the world already built? Elevated-transit noise mitigation in Japan, China, Sweden, Germany, Hong Kong, Australia and Chicago - and what actually transfers to a 1909 suspension bridge. | 11,470 |
3. Williamsburg comparator
WILLIAMSBURG-COMPARATOR.md | There is a second bridge with the same owner, operator, rolling stock and statute. What does it already tell us, and what would measuring it establish? | 12,757 |
4. Visual model framework
VISUAL-MODEL-FRAMEWORK.md | Every argument in the first three documents is an argument about a cross-section nobody has drawn. Can that drawing be built from open data - and made to admit what it does not know? | 12,458 |
5. Field capture protocol
FIELD-CAPTURE-PROTOCOL.md | Every acoustic claim beyond the published levels is invented. Can a consumer phone fix that this month? | 7,406 |
6. Community evidence audit
COMMUNITY-EVIDENCE-AUDIT.md | The people who live under it have been complaining since 2008. What have they already recorded, and why can nobody find it? | 10,300 |
7. Data collection
data-collection/README.md | How many trains, how many people, and for how long? Runnable scripts against MTA and NYC open data - and the traps that each silently produce a plausible wrong number. | 10,779 |
10. Field media
pedestrian-site-visits/README.md | The first material here that was not retrieved from somebody else: two days of phone video, stills, by-product audio and stopwatch laps under the bridge - and a v1.1 that withdraws v1.0's headline in place. | 8,251 |
8. AI usage and cost
usage/README.md | What did producing this repository consume? A per-request ledger read from the tool's own store, set against the argument that inference is becoming metered infrastructure - and what follows from measuring it. | 7,041 |
9. What this would have cost to buy
procurement/README.md | What would the same deliverable have cost from a large schedule holder or the cheapest decile of the same schedule? Three instruments, reported side by side and never averaged - and the disagreement between them is the result. | 5,075 |
11. Observation protocol
pedestrian-site-visits/OBSERVATION-PROTOCOL.md | Everything about trains here is measured and almost everything about people is invented. Ten things a person walking through DUMBO can count that would move a number - each with what it moves, what it would be rated, and what a result of zero would mean. | 5,188 |
Data and code
Every script runs against live public feeds and can be re-run from scratch by anyone. The schedule figures they produce are rated 5/5 — read directly from the MTA's own published feed.
| Script | What it does | Size |
|---|
bridge_schedule.py | Counts scheduled Manhattan Bridge traversals from the MTA GTFS static feed, by hour, route and direction. | 5 KB |
bridge_realtime.py | Polls GTFS-realtime for actual traversals. Built and verified; the week-long run has not been done. | 7 KB |
build_dashboard_data.py | Assembles the frequency dashboard's dataset, including the coincidence analysis that exposed the feed's 30-second quantisation. | 4 KB |
build_pedestrian_data.py | Derives arrival rate, departure rate, walkway flow and resident count from four public datasets. Disproved this repository's own claim about which hour is worst. | 17 KB |
build_cohort_model.py | Fits four population cohorts to the observed departure curve and reports the range across every parameter set that fits equally well - which is how the non-identifiability result was found. | 43 KB |
fetch_geodata.py | Fetches NYC building footprints with surveyed roof heights and the OpenStreetMap street, park, water and footway network for the corridor. The reproducibility path for every line in the noise-canyon drawings and the walkable model. | 11 KB |
build_carousel.py | Draws the noise-canyon slides from that geodata and emits the page. Slides are declared in visual-review/carousel.json, and the build refuses to run if any of them lacks a source or a caveat. | 60 KB |
build_walkable_map.py | Builds the walkable model: the routing graph, the shortest-path walk from York Street to Pier 1, and the audit of every inherited coordinate against surveyed building footprints. | 88 KB |
make_hero.py | Composites the hero band on this page from a public-domain HAER photograph and a render taken from this repository's own 3D model. | 8 KB |
dashboard-data.json | Traversal counts by hour, route, direction and period. | 33 KB |
pedestrian-data.json | Turnstile entries, origin-destination arrivals, walkway counts, residents. | 8 KB |
cohort-data.json | The admissible cohort parameter family and the presence ranges it implies. | 13 KB |
The trap that took three attempts to find. The MTA's turnstile feed publishes
entries at a station, which sounds like people arriving in the neighbourhood and is the exact opposite: an entry is somebody going
down into the system and
leaving. Reading it the natural way inverts the daily curve, and it will look entirely plausible while doing so. That directional trap and three others are documented in
the data-collection notes.
How this investigation works
This is an open, unfinished investigation, not a report. It is published while it is still wrong in places, because the method depends on that being visible.
The programme applies the methodology from ai-research-question-assistant — AI-Powered Assistance in Formulating Research Questions (Rhodes et al.), §8. Its central move is that gap identification and contradiction detection come before question formulation. You do not start from what you want to prove. You start by reading what the record already contains, finding where it stops, and writing the question it never asked.
68
questions derived this way
Numbered Q1 to Q68 and carried across every document, so a question raised in one place can be answered or killed in another.
49
methods specified to answer them
Each with a cost, a named question and an honest status. 6 have been executed. The register does not hide the ratio.
20
explicit retractions
Every one was published here, disproved here, and left standing in the text with the correction beneath it. That count going up is the method working, not failing.
Hypotheses, not positions
Every question here is written so that it can come back no. That is a deliberate constraint and it has cost this programme several of its more attractive claims — the three-site propagation fit, the peak-hour argument, and the assumption that the tool kept no record of its own cost. Each was a reasonable reading of real data. Each turned out to be wrong, and each is still on the page, because a research record that only shows its surviving claims is not showing its method.
What follows from that is the uncomfortable part: nothing here is a finding until someone has stood under the bridge with an instrument. The questions are sharp. The evidence behind them is second-hand, and every page says so where it applies. Where the investigation stands gives the honest ratio, and what to do next lists the work that would settle it.
How to read anything in here
Five conventions apply across every document and every artifact. They exist because this programme has failed adversarial review repeatedly, and always the same way: over-claiming from abstract-level reading.
- Every number carries its locus. Not a citation - the actual quoted sentence the number came from. If a claim has no locus, it is an inference and is labelled as one.
- Sources are rated 1 to 5 and marked VERIFIED, SNIPPET or UNVERIFIED. VERIFIED means the full text was read. SNIPPET means only an abstract or search result was seen. Most over-claiming in this programme has come from treating a SNIPPET as if it were VERIFIED.
- Errors are quoted in place, not deleted. When something here turns out to be wrong, the original wording is left visible as a blockquote and followed by the words that is withdrawn. A research record that hides its own corrections is not a research record.
- Every document ends by attacking itself. A section titled where this document is likely to be wrong, written by the authors, naming the specific claim they would attack first.
- Synthetic is labelled synthetic, in the interface. The 3D model contains zero measured elements. The audio is synthesised, not recorded. The agent model's itineraries are invented. Each says so on its own face rather than in a footnote somewhere else.
Behind the scenes
Two pieces of work about this investigation rather than about the bridge: what it cost to produce, measured from the tool's own per-request log, and what the same deliverable would have cost to buy from a consultancy. Both are held to the same evidence standard as everything else here, and both withdraw a headline claim on their own front page.
8. AI usage and cost Document
What did producing this repository consume? A per-request ledger read from the tool's own store, set against the argument that inference is becoming metered infrastructure - and what follows from measuring it.
9. What this would have cost to buy Document
What would the same deliverable have cost from a large schedule holder or the cheapest decile of the same schedule? Three instruments, reported side by side and never averaged - and the disagreement between them is the result.
Usage and cost dashboard Metav0.4.7
What this investigation cost to produce, from the tool's own per-request log: every model call priced by channel, time measured four ways that disagree by an order of magnitude, and an energy bracket that spans a factor of twenty-four because nothing here is a measured joule.
The two bars under the ledger: what an agent reads is most of the tokens and half the money; what it writes is under one per cent of the tokens and a fifth of it. Then the process note at the foot - the first conclusion this dashboard reached about itself was that the data did not exist, and that was wrong.
Procurement comparison dashboard Metav0.4.6
What this same deliverable would have cost to buy, from three instruments that are never averaged: dollars actually obligated on 56 federal noise-study contracts, a bottom-up build at GSA awarded ceiling rates verified cent-for-cent against a vendor's own card, and the metered inference ledger.
The first card, which withdraws the headline before it is made. Then the discipline populations near the foot: 12,825 project managers against seven acoustical engineers on the same schedule - and the project manager costs more per hour.
Why this is published at all: the cost of producing research with these tools is routinely asserted and almost never measured. Both pages exist so that the assertion here can be checked, including the parts that come out unflattering — the human direction time alone costs between 5.0 and 31.0 times the entire metered inference bill, which is the opposite of the usual claim.
How to help
This is a working research repository, not a publication. Corrections are more valuable than agreement.
There is now a way in. Everything here about
trains is measured. Almost everything about
people is invented, and it is why no absolute exposure figure has ever been published on this site.
Document 11 lists ten things a person walking through DUMBO can count that would each replace one of those invented numbers — what each moves, what it would be rated, and what a result of zero would mean.
Submit a field observation File a correction
Why a form rather than an email address. A count without a denominator cannot be used, and a count whose purpose was never stated can be read for the wrong thing — which is how a headline was withdrawn here once. The forms make those fields structurally required rather than politely requested. A GitHub issue also carries provenance an email cannot: an identified author, a server-side timestamp, a citable permanent reference, and a public edit history. Email is accepted as a fallback and is filed one rating step lower, and the record says so.
A result of zero is a result. If something does not happen in n observations, the 95% upper bound on its rate is about 3/n. Nobody has ever counted dogs boarding a train in this city; a commuter with a notes app would bound it inside a month, and that single number would settle a constant in the agent model that its own documentation currently contradicts.
- Count something. The ten observations, ordered by how much they move rather than how easy they are, are in Document 11. Two of them need nothing but a clicker and a stated corner.
- Post a recording. See the field capture protocol for what makes one usable.
- Falsify a “not found” claim. Five have already failed. The prior on others failing is not low.
- Execute any method. Several cost an email, a form, or an afternoon.
- Challenge the arithmetic. The event-duration derivation, the cohort identifiability result and the propagation rejection are each either right or wrong, and none has been reviewed by anyone. Issues #26, #27 and #28 exist specifically to attack them.
- Reclassify a component's provenance. Anyone with structural knowledge of riveted lattice trusses will find misclassifications in the 3D model, and finding them is the point of publishing the classification.
One standing condition. The people who live under this bridge have been asking for help since 2008. Any contact with them must offer something before it asks for anything, must not represent this programme as more established than it is, and must make clear that its central artifacts are synthetic. Over-claiming to a community in that position would be a different and worse kind of error than the ones already made here.