Document 11 — The pedestrian observation protocol

Ten things a person walking through DUMBO can count, each of which moves a number this repository currently invents. With a submission path that keeps provenance, and the arithmetic that makes a null result worth reporting.

Ethical Tech CoLab · v1.0

Why this document exists

This repository has a lopsided evidence base and has said so since Document 7. Everything about trains is measured: 1,073 weekday crossings, a 58-second daytime headway, 87.50 dB(A) at the dog run, all 5/5 VERIFIED from agency sources. Almost everything about people is modelled: group sizes, walking speeds, dwell times, itineraries, who is carrying a baby, whether anyone has a dog. Those are rated 1/5 and they are the reason no absolute exposure figure has ever been published here.

The gap will not close by more arithmetic. Document 8's cohort model proved that: ten free parameters against twenty-four data points, 9,248 parameter sets fitting equally well, and a worker/visitor split that swings by more than a factor of 1.5 across all of them. A departure curve carries no job titles. No amount of further fitting supplies a fact that was never observed.

What closes it is somebody standing on a corner with a clicker.

This document is the list of what to click. It exists because that person needs to know three things before they start, and none of them is obvious:

1. Which number their count would move, so the hour is not wasted on something already known or already unusable.
2. What rating their count would carry, so they can see in advance whether it will be believed.
3. What a result of zero means — because several of these observations will return zero, and a volunteer who thinks zero is a failed session will stop after one.

The rule that governs the whole list

An absence, counted, is data. An absence, remembered, is not.

This is the single most useful idea here and it came from an operator's own remark, which is worth quoting because it is exactly the shape of thing that usually gets discarded:

People ride the subway every day, so do I. I never see dogs being brought on the subway. I mean, in a year, maybe 1 time of everyday travels.

As stated, that is a recollection, and this repository rates operator recollections of events 2/5. It cannot be used. It is not weak because the operator is unreliable; it is weak because nobody wrote anything down, so there is no denominator and no way for it to have come out otherwise.

Now write it down. A year of weekday round trips is roughly 500 rides. One dog in 500 is 0.2%, and a count of one event in 500 trials carries a computable 95% interval of 0.005% to 1.1%. That is a sharper number than most of what this repository has about people, and it costs nothing but a note on a phone twice a day.

The same observation went from unusable to load-bearing purely by acquiring a denominator. That is what this protocol is for.

The arithmetic of a zero

Observers stop when they see nothing, because nothing feels like no result. It is not. If an event does not occur in n independent trials, the 95% upper bound on its rate is close to 3 / n — the statistical rule of three.

So a session that sees nothing still bounds the thing it did not see, and the bound tightens with effort in a way the observer can watch happening. Record the denominator even when the numerator is zero. A count of "0 dogs" is worthless. A count of "0 dogs in 143 boardings observed between 08:10 and 09:05 at York St, northbound platform" is a published figure.

What makes a count usable here

Every observation below has to satisfy five conditions, and they are listed first because an hour spent counting the wrong way is an hour that produces a number nobody can use — including a number that is wrong in a way nobody can detect, which is worse than no number at all.

1. A mechanical rule. The decision to click must not require judgement. "A dog" is mechanical. "A tourist" is not — so the protocol below never asks anyone to classify a tourist, and asks instead for something visible.
2. A denominator. Every count is a count out of something. Rides taken, minutes elapsed, people passing the line. Without it the count is an anecdote with a number attached.
3. A stated position and time. Not "under the bridge" — a corner, and a clock time. Two of this repository's existing photographs were re-filed entirely once their embedded position and timestamp were read, because those two fields changed what the frames could be used for.
4. A stated purpose, written before the count. Document 10 v1.1 withdrew a published headline because two correctly-computed datasets were joined that had never been joined in the field. The defence against repeating that is to record what a capture was for at the moment it is taken. A measurement can only be read for the purpose it was taken for, and only the person who took it can say what that was.
5. Whatever went wrong. The single most useful line in any submission is the one describing the thing that spoiled it. An operator's disclosure that his hand may have covered a microphone is why one recording is correctly labelled rather than silently wrong.

The ten observations

Ordered by how much they move, not by how easy they are. The two at the top are the two the whole exposure argument is blocked on.

O1 — Dwell: how long a person stays

Registered as Method 28. The single blocking unknown for any absolute exposure figure in this repository.

Why it is first. Little's Law gives L = λW. This repository has measured λ — arrivals, from MTA origin-destination data, 4/5 — and has never measured W. A commuter walking from York St to Water St is exposed for four minutes; a family on the lawn for three hours. Both currently count once, and they differ by a factor of forty.

The trap, stated because it would inflate the answer tenfold. Measure dwell in the park and apply it corridor-wide and every exposure figure grows by about an order of magnitude while looking like a measurement. Say where the cordon was.

What the model most needs is the bottom of the distribution, not the mean. The agent model's shortest-dwell decile sits at a modelled 430 s — about seven train crossings. Everything the arrival-process result rests on depends on that being roughly right. Q63 asks whether a cohort exists whose corridor time approaches a single 58-second headway, because if one does, several published nulls stop holding. Count the fast people carefully; they are the ones the model is least sure about and most exposed to being wrong about.

O2 — Dogs boarding a train

The observation that prompted this document. Directly tests a constant in the agent model against the reasoning printed beside it.

The specific thing it would settle, and there is a live discrepancy. The model's schema comment read, until this document was written:

dogRate 0..1 probability this group brings a dog. Near zero for anyone arriving by subway: the MTA requires a dog to travel in a container, so dogs in the corridor are walked from home rather than carried off a train.

The reasoning is sound. But the three personas that arrive by subway carry dogRate values of 0.02, 0.04 and 0.01 — and 4% is not "near zero" if the observed rate is nearer 0.2%. The model's own documentation and its own constants disagree by up to a factor of twenty, and nobody has noticed because nobody has counted. One commuter with a notes app settles it.

What has been done about it, and what deliberately has not. The model page now carries the disagreement on its own face — the schema comment states the constants and says plainly that they are not near zero, and the weaknesses list records the factor of twenty. The constants have not been changed. Replacing one unmeasured number with another unmeasured number chosen to match the prose would make the page self-consistent and no more true, and it would remove the visible disagreement that is currently the most useful property of this input. The disagreement is left standing until a count closes it.

Three outcomes, all publishable:

• Observed rate near zero. The comment is right, the constants are too high, the dog-walking population in the corridor is almost entirely resident, and the susceptibility layer's dog class should be driven from resident numbers rather than from transit arrivals.
• Observed rate near 2-4%. The constants are right and the comment is over-stated. Say so and delete the comment.
• Observed rate varies by line or hour. The most interesting outcome, and the one neither the comment nor the constants can currently express.

O3 — Arrivals over the Brooklyn Bridge walkway

The dataset that died. There is a real hole here and it is dated.

NYC OpenData's hourly directional Brooklyn Bridge walkway counter (6fi9-q3ta) has published nothing since 2019. Its 2019 profile is still what this repository uses, and it is doing so across a pandemic, a tourism collapse and a recovery. Nobody knows the size or the sign of the error.

There is no live pedestrian counter anywhere near DUMBO. All ten of NYC DOT's pedestrian-capable sensors are elsewhere. The live counters here are bike-only — including one named "Manhattan Bridge Ped Path" whose travelmodes field reads bike.

Two hours at the landing, done twice, would be the most recent walkway pedestrian data in existence for this location.

O4 — Group size and composition

Three of the six susceptibility classes are not observable and this observation cannot supply them. A pram is visible; hyperacusis, autism and a cancer history are not. Those three will stay at national prevalence, and the protocol says so rather than implying a street count could reach them. This is Method 42's honest half.

A zero count of prams under the bridge would be the most important result in this document. The agent model's own weakness list already states the problem: nobody in the model is prevented from coming at all. If the noise is why a parent does not push a pram under the bridge, then the people most sensitive to it are the ones most likely to be missing, and the model would report that as low exposure rather than as exclusion. A pram count that comes back near zero under the span and normal two blocks away is the difference between those two readings — and it is a comparison, so it needs no instrument.

O5 — Trains audible, and whether two overlap

Answers a question the published schedule provably cannot.

Why no amount of schedule analysis reaches this. Every scheduled departure in the MTA feed falls on an exact :00 or :30 second. Across 1,073 weekday traversals the sub-minute values are {0: 528, 30: 545} — two distinct values, nothing else. The gap histogram has 154 pairs at 0 s, 279 at 30 s, 198 at 60 s, and nothing between 0 and 30. The entire range in which two trains merge acoustically is empty by construction. A merged-pair table computed from that feed was published here once and withdrawn for exactly this reason.

So the only way to find out whether two trains on the bridge sound like one event twice as loud, or one event twice as long, is for a person to stand there and notice. Energy addition predicts +3.01 dB for two identical sources; what a listener reports may differ, and the difference between those two is a real finding either way.

O6 — Where the canyon stops

This is the one observation an uncalibrated human is better at than a cheap instrument. Speech interference is a defined, well-studied effect; a phone's automatic gain control actively destroys the decay tail that a listener hears without difficulty. The protocol asks for the perceptual judgement and forbids the conversion, because converting it would manufacture a decibel out of an opinion.

The corpus asserts a "noise canyon" — 17 buildings cut, 91.22 m between nearest facades, height-to-width 0.29 — entirely from surveyed footprints. Nobody has walked out of it and recorded where it stopped.

O7 — The photograph stop

Comes directly from an operator observation about evening tourists.

A photograph stop is dwell in place, at a fixed location, at a known distance from the track. It is the mechanism by which visitor dwell gets its tail, and the tail is what makes visitor exposure differ from commuter exposure. It is also trivially countable and requires no interaction with anybody.

Two existing photographs in this corpus were re-filed into the pedestrian cohort study once it was established what they had been taken for — evening tourists on a weekday. Their embedded position and timestamp survived; a count would have survived better.

O8 — Ear-covering

The susceptibility layer refuses to compare the MTA's 98.90 dB(A) peak against the loudness-discomfort levels in the audiology literature, because dB(A) and dB HL are not the same units and bridging them needs a third-octave spectrum that does not exist for this bridge. That refusal is correct and it leaves the layer with no behavioural evidence at all. This is the cheapest available.

The control point is not optional. Ear-covering happens for reasons other than trains. Without a comparison point the count means nothing.

O9 — Windows facing the bridge

DumboNYC.com maintained a list of insulation contractors. Residents have been paying privately to solve this since at least 2008, and the total has never been added up. An open/closed window ratio is a crude proxy and its confounders are obvious — orientation, sun, rent tier, whether the unit has air conditioning. It is included because it is the only route to Q44 that does not require asking anybody for their receipts, and because stating a weak observation as weak is cheaper than pretending it is unavailable.

O10 — The York St to water's edge walk

Already in the backlog as a drawn route. This is the counted version.

This route crosses the one geometry this programme keeps asserting and has never walked end to end: it goes under the bridge, out from under it, and back into its shadow at Fulton Ferry. It is the cheapest single trip that touches four open questions at once.

How to submit an observation, and how provenance is kept

The requirement

An observation that arrives without provenance cannot be used, and this repository would rather refuse it than quietly downgrade everything around it. So the submission path is built to make the provenance fields structurally required rather than politely requested.

Route 1 — a GitHub issue form (preferred)

Two forms live in .github/ISSUE_TEMPLATE/:

The correction form matters as much as the observation form and nobody asked for it. This corpus has withdrawn published claims twenty times, every one of them found internally. There has never been an inbound route for a reader to file one, which means the self-correction record measures this programme's appetite for finding its own errors and nothing about anybody else's ability to find them.

What GitHub gives the provenance for free, and it is the reason this route is preferred over any form service:

• an identified author — an account with a history, not an anonymous post
• a server-side timestamp the submitter cannot set
• an immutable reference — issue number and permalink, citable from a document the way every other source in this repository is
• a public edit history — if a submission is later changed, the change is visible, which is exactly the property a research record needs and exactly the property an email thread lacks
• no third party: nothing passes through a form vendor's database

One property of the mechanism had to be tested rather than trusted. A YAML issue form that GitHub's parser rejects does not fail visibly — it degrades silently to a single plain body textarea, which is precisely the unstructured free-text box this whole route exists to replace. A form that looks correct in the repository and is broken on the site would remove every required provenance field while appearing to have added them. Both forms are therefore validated against GitHub's published issue-form JSON schema by .github/verify_issue_forms.py, which reports the field count, the type mix, the number of required fields and the number of unique identifiers — 9 required fields on the observation form and 5 on the correction form at the time of writing. It is committed rather than run once, because the failure it guards against is invisible.

Route 2 — email

For anyone without a GitHub account, or who does not want their name attached to a public repository. Send the same fields. They will be transcribed into an issue.

This route has weaker provenance and the difference is stated on the record, not smoothed over. An email has no public timestamp, no public edit history, and the transcription is performed by someone with an interest in the result. A transcribed observation is therefore filed at one rating step below the same observation submitted directly, and the issue says that it was transcribed and by whom. A submitter who wants the higher rating can open the issue themselves.

Nothing is published under a submitter's name without explicit consent, and consent to publish a count is not consent to publish a name.

How a submission gets rated

Inbound observations go through the same rubric as everything else. Nothing is rated higher for being ours, and nothing is rated lower for being a stranger's.

Statements of intent are the exception and they are rated 5/5. Nobody but the person who took a measurement can say what it was for, and no analysis performed afterwards can overrule them. That rule was written after an operator's correction retired a published headline here, and it applies to inbound submissions identically.

Where this document is likely to be wrong

1. Nobody has run any of these. Every estimate of what an observation would cost in time, and every claim about what it would move, is a prediction. The first real session will find that some of these rules are not as mechanical as they read at a desk. O5's "audible" and O8's "covering their ears" are the two most likely to fail that way.

2. The observations are not independent of the models they would correct. The list was assembled by reading this repository's own weakness lists, so it is shaped to fill the holes this programme already knows about. An observation programme designed by someone who had never read these documents would probably count different things — and the things it counted instead would be the more interesting list.

3. Volunteer counts have a selection problem this document does not solve. People who volunteer to count noise-related things in DUMBO are not a random sample of people in DUMBO. They are more likely to be residents, more likely to be annoyed, and more likely to count at the times that annoy them. Every observation above inherits that, and only O1 and O3 have an obvious defence (fixed session bounds set in advance).

4. The rule of three assumes independence, and rides are not independent. A commuter counting dogs on 500 rides is counting the same route, the same hours and often the same carriage. The effective sample is smaller than 500 and by an unknown factor. The bound in the table above is therefore optimistic, and the direction of the error is known even though its size is not.

5. O9 is weak and is included anyway. Its confounders are not controlled and probably cannot be by this method. It earns its place only because Q44 has no cheaper route, and a reader is entitled to skip it.

6. A submission path is not a submission. Building the form is the easy half. Nothing here establishes that anyone will use it, and the honest prior is that most such channels receive nothing. If it receives nothing, that is a finding about this programme's reach and should be published as one rather than quietly dropped.

7. This document asks people to do unpaid work for a research programme that has no funding and no institutional standing. It should not represent itself as more established than it is, and the standing condition applies: any approach to the community here must offer something before it asks for anything.

Questions this raises

Q65. What is the rate at which dogs board subway trains on the B, D, N, Q, F, A and C, and does it differ by line, hour or station? The corridor's dog population is currently derived from a container rule and a national prevalence figure, with no observation between them. Method 45.

Q66. Does the pram and wheelchair rate under the span differ from the rate two blocks away at the same hour? A difference would be the first direct evidence that the noise is excluding people rather than merely exposing them, which is a claim the agent model is currently structurally unable to make. Method 46.

Q67. Do two trains crossing together produce one event or two to a listener standing beneath? Energy addition predicts +3.01 dB; the schedule cannot answer it at all, because its departure times are quantised to :00 and :30 and the merge window is empty by construction. Method 47, and it rides along with Method 41.

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