Field media: what a consumer phone actually recorded under the bridge

Page: media.html · Data: media-data.json · Analysis: build_media_data.py · Renderer: build_media_page.py · Harness: verify_media.js · Next session: FIELD-KIT.md

This is the first material in this programme that was not retrieved from somebody else. On 3 and 4 August 2026 an observer stood under the Manhattan Bridge in DUMBO with a Samsung Galaxy S23+ and filmed the buildings that wall the corridor on both sides of the alignment. Separately, an hour later, he stood with the phone's stopwatch and tapped a lap every time train noise started or stopped.

Every other document here argues from records: MTA memoranda, GTFS feeds, Socrata datasets, GSA rate schedules, a HAER survey. This one argues from fourteen taps on a screen, about two minutes of by-product audio, and a set of photographs of a place this repository had until now only modelled.

0. What was corrected in v1.1, and why it matters more than what was kept

Version 1.0 made a cross-instrument claim that is withdrawn. It is quoted in full in section 5 rather than deleted. The short form: the audio duty cycle was used to refute one of the two possible readings of the stopwatch, and it cannot, because the two are independent samples 62.4 minutes apart with no overlap.

The correction came from the operator, and it arrived as three statements of intent:

The stopwatch is an independent sample. There is no correlation between it and any video or audio also supplied.

It is somewhat imprecise because a human operator decided when to start and stop the stopwatch. It was an indicator that there is a very small amount of time between noise from the train tracks. I will redo this with better documentation.

The video was primarily used to show the buildings that create an echo chamber around the bridge near the water. That was the main value of them. It was not to be used for audio precision.

These are rated 5/5. Every other piece of testimony in this document is rated 2/5, because a recollection about what happened can be wrong. These are not recollections about events. They are statements about intent from the only person in a position to know it, and no analysis performed afterwards can overrule them.

A measurement can only be read for the purpose it was taken for. The analysis that produced v1.0 never asked what the captures were for. It treated a video shot to photograph buildings as an acoustic instrument, and it treated a hand-timed indicator as a data series to be reconciled with that instrument. Both files are real and both numbers were computed correctly. The inference joining them was invented by the analyst, not supplied by the field.

The net effect: this document's centre of gravity moves from the audio to the photographs. The buildings forming the echo chamber are the thing the session was for, and they are the thing it delivered.

Better capture is already scheduled. The operator is re-running the stopwatch with documentation of which tap was which, and capturing audio with a microphone shield and an audio meter later this week. The acoustic content here is superseded rather than supplemented when that arrives, which is the reason to demote it now rather than defend it.

python pedestrian-site-visits/make_derivatives.py    # masters -> web/
python pedestrian-site-visits/build_media_data.py    # -> media-data.json
python pedestrian-site-visits/build_media_page.py    # -> media.html
node   pedestrian-site-visits/verify_media.js        # both themes

1. The refusal, first

Nothing in this directory is a decibel. That is not modesty. It is the condition under which the material is admissible at all.

A phone running Android's default capture path applies automatic gain control. A compressor that pulls the level down during a loud passage and lets it back up afterwards writes its own release curve into the file. Anything computed from the absolute value of those samples is a measurement of Samsung's compressor. The microphone response is unknown and is not flat, and phone MEMS capsules are routinely high-passed in hardware — which is exactly where bridge structural radiation lives, so a spectrum from this material would be missing the part that matters with no way to tell from the file that it is missing. There was no calibration and no windscreen, and Document 5 named both as the things that decide whether a session is worth anything.

What survives all three is timing. AGC changes how loud a sample is; it does not change when the sample was written. Onsets, durations, gaps and rates are recoverable from a file whose absolute scale is worthless. Every quantitative claim below is one of those four, and the unit on every chart is dBFS — a number about the file — never dB(A).

The MTA's dB(A) figures used everywhere else in this repository come from a calibrated meter. Nothing here is comparable to them, and the temptation to put the two on one axis should be resisted permanently.

2. What was captured

The masters total 224.7 MB and are not committed. One of them is 114 MB, which is over GitHub's hard per-file limit, and putting a quarter of a gigabyte of phone footage into a research repository serves nothing. What ships is a poster frame and an extracted audio track per video and a resized still per photograph: 4.1 MB, a factor of 55. Every master is fingerprinted by SHA-256 in media-data.json and on the page, so a master supplied later can be checked against the one that was actually analysed.

Every capture carried GPS and full EXIF and none of it was stripped.

The .gitignore nearly shipped a page of broken images

The patterns excluding the masters were written as grassy-knoll-*.jpg and similar. A gitignore pattern with no leading slash matches at any depth, so those patterns also matched the derivatives inside web/ and silently excluded them. The page would have been committed with every image broken and no error anywhere in the build.

Caught by git add -A -n pedestrian-site-visits before committing, and fixed by anchoring each pattern with a leading /. The slashes are load-bearing and there is a comment in the file saying so.

3. Where the captures actually stood

Placed in the frame the rest of this repository already uses: distance along the fitted track axis, and perpendicular distance from it. Chainage zero is the DUMBO Archway, which is itself one of the four points the MTA measured.

The axis is imported from build_carousel.py rather than reimplemented, so these numbers cannot drift from the drawings in the noise canyon. It is fitted to 21 OpenStreetMap nodes tagged railway=subway on the bridge and is straight to within 12.5 m over the 1,757 m it was fitted across.

Two of the video captures stand 26 m and 27 m from the track. Three of the four points anyone has ever measured are further from it. That is not a claim these captures are better sited — they carry no calibration and the MTA's do. It does mean geometry is not the reason nothing here is comparable with a decibel.

Two guards that exist because the first version was wrong

The compass label was typed and it was backwards. The first version of the geometry table hard-coded "north-east" if t > 0. Probing the fitted axis showed positive offset actually points 246.9°, west-south-west, toward the river. A refit that flipped the sign would have silently inverted every row. The bearing is now derived from the fitted axis at runtime on a 16-point compass.

An axis is only a description where the fit lives. A straight line fitted to a railway says nothing about that railway outside the span of the fitted points. geo_context() records the fitted span and asserts every capture falls inside it. All do; captures_within_fit is true. An offset quoted outside that range would be measuring distance from an imaginary line.

4. The stopwatch, and the inference the whole reading rests on

Fourteen laps, 12:59:00 to 13:07:19, one observer. The stopwatch records a lap every time the observer judged train noise to start or stop, so the file contains fourteen durations alternating between noise and quiet.

It does not record which kind the first lap was, and nobody wrote it down. Both interleavings are arithmetically valid and they give opposite answers.

The cycle, and therefore the rate, is the same either way — 63.8 s, 56.4 an hour. Everything else inverts.

The tie is broken on a property of the railway, not a preference. Headway is scheduled and event duration is not. Trains are dispatched to a timetable, so gaps between them should cluster; how long a given train sounds loud depends on its speed, its length, which of four tracks it is on and where the observer is standing, so events should scatter. The reading whose quiet intervals cluster more tightly is the reading in which the quiet intervals are really quiet intervals.

The two differ by 4.67× in exactly that respect. Under the accepted reading the gaps have CV 0.227 and the events scatter at 1.060. Under the rejected reading those swap, which would mean the railway runs to no timetable while every train sounds loud for almost exactly the same length of time.

That is a strong argument and it is still an inference. It is the load-bearing assumption in this document and it is listed first in section 9.

It now rests on that one argument alone. Version 1.0 gave it a second and apparently independent prop — an audio duty ceiling that seemed to rule the other reading out. That prop is withdrawn in section 5, so the confidence attached to the pairing, and to the event durations and duty cycle hanging off it, drops with it. The data file marks the pairing "confidence": "WEAK" and "sole_support": true.

The operator's own reading points the other way

The operator describes the session as an indicator that there is

a very small amount of time between noise from the train tracks.

That is ambiguous between two things this stopwatch cannot separate. It may mean the gaps are short, which would favour the rejected pairing, where gaps average 12.8 s. It may mean the cycle is short, which is pairing-independent and is already established at 63.8 s — a train about every minute, which anyone might reasonably describe that way.

It is recorded, not resolved. Reading it either way would be choosing the answer. The ambiguity is itself the argument for re-running the stopwatch with documentation, which the operator has said he will do.

Only the cycle is pairing-independent

The cycle, and therefore the event rate, is identical under both pairings. It is the one quantity here that does not depend on the tie-break, and it is the only one carried forward into section 8.

One event does not fit and it is not hidden

Lap 2 is 41.09 s against 18.63 s for the next longest of the seven and a median of 7.3 s — a 5.6× outlier on the median. It is the first event of the session, which is where a settling error would sit: a stopwatch started on a train already passing, or a missed tap that makes one recorded duration span two real intervals.

The lap was not dropped and the reading was not re-chosen to remove it. The mean of the accepted reading is 12.8 s and the median 7.3 s, a gap this one lap almost entirely accounts for, which is why the median is quoted beside the mean throughout.

5. The audio, which is a by-product and is treated as one

The audio analysed here is the sound track of video shot to photograph buildings. It was not captured for acoustic precision and is not used as though it was. It is used for exactly one thing that it can carry: showing that one recording's microphone misbehaved, which is a claim about the recording chain rather than about the corridor.

The detector was written and run before the stopwatch file was parsed. An excursion is a stretch at least 1.5 s long sitting at least 6 dB above the clip's own rolling median, with excursions less than 1.5 s apart merged.

It is a level excursion, not a train. The detector cannot distinguish a train from a truck, from the operator's own footsteps as he walked, from clothing against the handset, or from wind. On a corridor carrying heavy road traffic under the same deck, recorded by someone walking through it, this is not a marginal caveat.

Across 108.4 s of usable recording: 4 excursions, longest 10.10 s, and the duty cycle never exceeds 44% at any detection threshold tested.

Those are statements about three files. They are not statements about the corridor, and the next section withdraws the claim that treated them as such.

The withdrawal

Version 1.0 of this document said:

That refutes the rejected reading robustly. It does not confirm the other. The rejected stopwatch reading requires the corridor to be under train noise about 80% of the time with each event lasting about 51 s. The audio ceiling is 44% and the longest excursion of any kind is 10.10 s.

That is withdrawn. It was the strongest-sounding result in the document and it was not a result at all. Two independent reasons kill it, either of which is sufficient.

First, the two instruments do not overlap. The operator states there is no correlation between the stopwatch and any video or audio supplied. The file timestamps agree without being asked:

A duty cycle measured in one window places no constraint on a duty cycle in another window an hour away. The arithmetic was right; the inference joining the two was invented here and was never in the data.

That the operator's statement and the file timestamps agree is the strongest form this evidence could take, because they could have disagreed. The check is in build_media_data.py as separation(), and it hard-exits if the two ever overlap.

Second, and independently, the video was never an acoustic instrument. It was shot to record the buildings forming the echo chamber near the water. Its audio track is a by-product of a visual record, captured while walking, and cannot characterise an acoustic environment however carefully it is processed.

What survives

Each instrument constrains itself and nothing else. Concretely:

• The duty ceiling has no consumer. Its only use was the refutation above. It is retained in the data as a property of three files and is used for nothing.
• The threshold sweep is kept for the opposite reason it was built. On the same audio it runs from 44% down to zero depending only on where the threshold is put. That is worth publishing as a caution against anyone — including this repository, which did it — quoting a single duty figure from uncalibrated audio.
• The obstruction finding in section 6 is unaffected, and section 6 says why.

6. A microphone was obstructed, and finding that out was worth more than the clip

After the recordings were made the operator volunteered two recollections about the lawn clip:

I may have had my hand over the microphone for the drop out period.

Turned phone upside down at some point.

Both are rated 2/5, OPERATOR RECOLLECTION. Recording a recollection as evidence would be circular. Recording it as a hypothesis and then testing it is not, and both tests were written after the recollection and were free to disagree with it.

The mono mixdown had to be undone first

The pipeline was mixing to mono. That destroys the only evidence that separates a quiet site from an obstructed microphone, because both produce the same envelope. A stereo decode path was added specifically to make the question answerable.

The first result was negative, and checking it found a blind spot

channel_asymmetry() found no sustained divergence in |L-R| and no sustained sign change in (L-R). Before accepting that, the detector was checked for power: channel correlation runs 0.83–0.92, so the channels are genuinely separable and the test could have fired. Printing the full envelope showed no drop-out at all — the clip sits inside a 5.9 dB band for its entire 10.2 s.

That looked like a clean negative. It is not, because the test has a blind spot: it measures each clip against its own median, so an obstruction that never lifts moves the baseline with it and becomes invisible.

The between-clip test, on three indicators that disagree by design

A hand over a microphone raises the noise floor, compresses the dynamic range, and decorrelates the channels. A genuinely quiet site lowers the floor and leaves the channels agreeing. The indicators point opposite ways for the two explanations, which is what makes them a test rather than a description.

All three picked the same clip, unanimously. Three indicators over three clips cannot carry a p-value and none is claimed. The consequence does not depend on resolving the cause: a clip that looks like that is unusable either way.

The clip is excluded at source, before any statistic is computed, not caveated afterwards. A caveat below a number does not stop the number being quoted. The audio span in this document is 108.40 s rather than 118.68 s because of it.

7. A null shorter than a headway is not a finding

The lawn clip is closer to the track than either canyon clip (27.0 m against 37.6 m and 26.4 m) and detected zero excursions. That has the shape of a result. It is not one.

At 57.7 trains an hour a 10.3 s window has an 84.8% chance of containing no train at all regardless of how loud the site is. detectability() computes this for every clip and marks the uninformative ones, so a zero count can never silently become a claim about quiet.

This test was added after noticing the proximity, and it would have caught the error even if the obstruction test had not.

8. The rate agrees with the timetable, within a wide interval

The accepted reading counts 7 events in 446.45 s, or 56.4 an hour. The MTA schedule for the same weekday window puts 8 trains across the bridge, or 57.7 an hour.

Seven events is a small sample. The 95% Poisson interval runs from 13.8 to 99.0 an hour and the scheduled rate sits inside it.

This is the only figure here that touches an external source, and the direction matters: the observed rate could have landed outside the interval, which would have said the stopwatch reading was wrong. It did not. That is weak positive evidence and it is reported as weak. It rules out gross error; it does not establish agreement to a few per cent.

8.1 What was asked for and is not separately here

The capture list requested an audio recording at Adams and John Street as a distinct file. No audio-only file exists in the material handed over.

That material appears instead to be inside the canyon videos. Both were shot at 40.7044, −73.9885 — the John Street block — and the poster frame of the first carries a John St / DUMBO Historic District street sign with the bridge deck filling the top third of the frame. Their audio tracks are extracted and published as .m4a, so the recording exists; it is simply not a separate capture.

This is recorded rather than quietly reconciled because the difference matters for what comes next. A separate audio capture would have been made with the phone held still and pointed. A video's audio track was made by someone walking and turning. The second is worse for anything spectral and no worse for anything temporal, which is all that is claimed from it.

8.2 Measured against the protocol this repository wrote for itself

Document 5 specified five captures before any of this existed. This session satisfies none completely and two partly. That is the honest score, and it is worth stating plainly because the temptation with first field data is to let its existence stand in for its adequacy.

C2 remains the one that matters. Section 1.7 of Document 1 solves event duration under an assumed envelope shape, and the continuous mode of the acoustic demonstration converges on the published L eq because that is the number the shape was solved from. That loop is still closed. This material does not open it, because AGC deforms precisely the rise and decay the assumption is about. A UMIK-1 over USB-C, about USD 100–150 with a serial-numbered calibration file, still does.

The equipment gap is closing this week. The operator has stated he will capture with a microphone shield and an audio meter. A shield addresses the wind and handling noise that make C1 unsatisfiable; a meter addresses the absolute scale that makes everything on this page a ratio. When that material arrives, C1 and C2 become answerable for the first time, and the acoustic content of this document is replaced rather than extended.

9. Where this is likely to be wrong

1. The pairing inference is load-bearing and it is an inference. Everything the stopwatch contributes rests on deciding which laps were noise. The tie-break is principled and the ratio is 4.67×, but it is one observer, fourteen laps, and no independent record of what the first tap meant. It had a second prop in v1.0 and no longer does. If the pairing is wrong, the event durations and the duty cycle are wrong together. The rate survives either way, because the cycle is identical under both.
2. The stopwatch is an indicator, not an instrument, and its author says so. A human decided when to start and stop it, which puts reaction time, anticipation and attention into every lap with no way to separate them. It is being re-run with documentation. Nothing derived from it should be treated as settled until that happens.
3. The audio is a by-product of a visual record. Every acoustic characterisation drawn from it has been withdrawn. What remains is one finding about the recording chain. If even that is over-read — if the between-clip test is picking up something about the lawn rather than something about the microphone — then the audio contributes nothing at all, and the honest position would be that only the photographs and the geometry survive.
4. The observer knew what the study is about. The same person who has been writing about train noise for weeks decided, in real time, when train noise started and stopped. Expectancy effects are not controlled for and cannot be recovered from the file.
5. A level excursion is not a train. Trucks, footsteps, clothing rustle and wind all qualify. Some of the four excursions may not be trains.
6. Four excursions is not a sample. Every audio statistic rests on four events in under two minutes of usable recording. Little survives that being an unrepresentative two minutes.
7. The exclusion rests on three indicators over three clips. No significance is claimed and none is available. With a different third clip the unanimity might not hold.
8. The rate agreement could easily be coincidence. An interval running from 13.8 to 99.0 an hour would have accepted a very wide range of observed rates.
9. The schedule is not the service. Every timetable comparison inherits the known problems of GTFS static: no non-revenue moves, a plan rather than an observation, and departure times quantised to the half minute.
10. The geometry rests on a fitted axis and digitised anchors. The MTA memos give locations as descriptions, not coordinates. The offsets are good to tens of metres, not to metres.
11. One handset, two days, one observer, one weather condition. Nothing here establishes anything about a different season, time of day, wet rail, or a different phone.
12. The largest error in v1.0 was not arithmetic. Every number in the withdrawn claim was computed correctly. What was wrong was joining two things that were never joined in the field. That failure mode is invisible to every check this directory runs — the tests verify that numbers are computed correctly and that pages render, not that an inference between two datasets was licensed by how they were collected. The only defence is asking the person who captured the material what each capture was for, and that only happened here after publication.
13. The operator notes came from the person who made the recording, after the fact, with knowledge of what the recording was for. They are rated 2/5 for that reason. The between-clip test does not depend on them; it would have flagged the clip had they never been offered. The statements of intent in section 0 are a different kind of thing and are rated 5/5 — nobody else can say what a capture was for.
14. The evening-presence count is one frame, and it is a floor. More than a hundred people in one photograph at one instant on one Monday cannot establish a rate, a dwell time, a cohort split, or a typical evening. It is existence evidence and a lower bound on one frame. The two frames were shot at different focal lengths — 23 mm and 69 mm equivalent — so they cover different depths of street and cannot be compared to each other at all, in either direction. If any of this is read as a density, a rate or a corridor total, it is over-read.

10. What this changes, and what it does not

It does not change any published number about the bridge. No decibel here, no exposure figure, no revision to any MTA-derived quantity.

What it changes:

• The echo chamber is photographed. Until now the noise canyon page drew the corridor from 76 surveyed footprints and OpenStreetMap ways, with nobody having stood in it. These frames are the first photographic record here of the geometry that page models. They do not measure it; they show that the modelled canyon and the built canyon are the same place. This was the purpose of the session and it is what the session delivered.
• The receptor is photographed. The lawn frame is the only image in this repository showing people, in the corridor, under the deck, including a family group with a pushchair. It is not a count. It is the population the cohort model has only ever inferred from turnstile arithmetic.
• Evening weekday presence is directly observed for the first time. Two stills from 3 August 2026 (a Monday) at 19:00:55 and 19:01:52, 57 seconds and 4.3 m apart on Washington Street, show the roadway occupied kerb to kerb. The camera stood 159–162 m from the fitted track axis; the people in the frame lie between it and the bridge, so their distances run from about that figure down toward zero. The operator states these were captured to show evening pedestrian presence on a weekday night. They are read for that purpose, not for the echo-chamber purpose the video serves. What they establish and what they do not is set out in section 10.1.
• The captures are placed against the measured points. Two stand 26 m and 27 m from the fitted track axis, closer than three of the four points anyone has ever measured. That is geometry from embedded GPS and is unaffected by everything withdrawn above.
• The observed rate agrees with the timetable at this location, weakly, in the first check of the schedule against anything observed on the ground. It is the one stopwatch quantity that does not depend on the pairing.
• One recording is known to be defective, and was known before disclosure. That is a finding about method, not about DUMBO, and it is the only thing the by-product audio establishes.

What it explicitly does not change, after the withdrawal in section 5:

• There is no measured duty cycle for this corridor. The 44% figure is a property of three files and constrains nothing.
• There is no empirical anchor on event duration. v1.0 offered one by placing the stopwatch median beside the longest audio excursion and calling the pair a range. Those are two unrelated samples an hour apart and that range is withdrawn. Section 1.7 of Document 1 remains unchecked against anything observed.

10.1 The evening stills as pedestrian-cohort evidence

Added in v1.2. The operator states:

The first 2 photos that are described as 'under the bridge, evening' were not taken with the same purpose. Those 2 photos are useful for showing tourists in the evening on a weekday night. The metadata on the photo shows the exact location and time. This should go into the study on the pedestrian cohort as evidence.

Rated 5/5, OPERATOR STATEMENT OF INTENT. This reassigns the two evening stills from the echo-chamber material to the pedestrian presence study.

What the metadata establishes

The two positions are 4.3 m apart — the same location to within phone GPS accuracy in a street canyon. The time separation is 57 seconds. They are two frames from one spot at one moment, not two samples from two places.

They are not, however, two views of the same scene. That is the single most important fact about reading them. The EXIF shows the first frame at a 35 mm-equivalent focal length of 23 mm and the second at 69 mm with a digital zoom ratio of 3.63 — the wide camera and then the telephoto. The horizontal angular field differs by a factor of about 3.5, and because Washington Street runs straight toward the bridge, the narrower frame does not show a smaller scene. It shows a much longer one: the same angular resolution reaches perhaps three times further up the street before individuals stop being separable.

Two consequences follow, and both cut against the obvious reading:

• The two counts are not comparable and neither is a density. More people in the telephoto frame does not mean a denser crowd; it means a longer sightline through the same crowd. Any comparison between the frames measures the lens.
• The single offset in the table belongs to the camera, not to the people. The figures visible in the telephoto frame lie between the camera and the bridge, spread along a few hundred metres of street. Their distances from the fitted track axis therefore run from about 162 m down toward zero — most of them are closer to the track than the camera is, and the ones furthest up the street are close to directly beneath it. That widens what the frame evidences and narrows what a single distance figure can be attached to.

What this establishes for the cohort model

The cohort model (build_cohort_model.py) infers weekday 19:00 presence at approximately 20,614 total in the corridor, of which 2,761 non-residents. Until now, no direct observation of evening presence existed anywhere in this corpus. Every figure was derived from turnstile arithmetic and a fitted survival function.

These two photographs are the first direct observation of presence in the DUMBO corridor at any hour. They establish:

1. Existence. People were present on Washington Street at 19:00–19:02 on a Monday evening in August. That is trivially obvious to anyone who has been there and it was never in doubt — but it had never been observed in this corpus. Every prior presence figure was inferred.
2. A floor on the count at one instant, and not an estimate. The telephoto frame (portrait, looking north up Washington Street toward the Brooklyn tower) shows the roadway occupied kerb to kerb. A deliberate count of separable figures in the resolvable band alone passes one hundred, and the crowd continues beyond that band into a mass in which individuals cannot be told apart at any published resolution. The correct output is therefore a floor and not a range: more than 100 people were within this one frame, and the true number is larger by an amount the photograph cannot supply. The wide frame, covering a much shorter stretch of the same street, holds several dozen. Neither number is a density and the two are not comparable, for the reason given above. This is one frame at one instant, not a rate.

What this does not establish

3. Not a rate. Two frames 57 seconds apart cannot measure a flow. Nobody entered or left the frame in a way that can be counted across both images.
4. Not a dwell time. A photograph records who is present, not how long they have been there or how long they will stay.
5. Not a cohort split. The operator's word "tourists" is an interpretation, rated 2/5 as operator recollection. A photograph carries no job titles, no home addresses and no itineraries. Whether the people visible are workers, visitors, transients or residents cannot be determined from appearance, and attempting to distinguish them from clothing or behaviour would be both methodologically unsound and ethically poor.
6. Not a typical evening. One Monday in August is one Monday in August. It cannot establish what any other Monday evening looks like.
7. Not a count for the whole corridor. The wide frame covers roughly one block of Washington Street and the telephoto frame a longer strip of the same street. The corridor extends over a kilometre and includes streets these frames never see. Extrapolation from one vantage point to the corridor population is not licensed by these data — and the lens difference means the frames cannot even be used to estimate the coverage factor that such an extrapolation would need.

The "tourists" interpretation

The operator describes the subjects as "tourists." That word is a statement about who the operator believed the people to be. It is rated 2/5, OPERATOR RECOLLECTION — not 5/5, because this is a recollection about the world, not a statement of intent. The photographs themselves cannot distinguish a tourist from a resident from a worker leaving late. The presence is high- confidence; the characterisation is not.

Relationship to the cohort model

The cohort model's weekday 19:00 estimate of ~2,761 non-residents is a corridor-wide figure. Counting more than a hundred people in a single frame does not confirm or refute it — the two are not commensurable without a spatial coverage factor and a separation of residents from non-residents, neither of which these photographs supply. It is worth noting only that the frame is a small fraction of the corridor and the count in it is a floor, so nothing here suggests the model is too high; that is the one direction the observation can weakly speak to, and it speaks to it weakly.

What the photographs do supply is the qualitative confirmation that the corridor is populated at this hour — which the model predicted and which is now observed rather than inferred. That is a genuine contribution, and it is a small one. It does not substitute for Method 28 (dwell-time measurement).

Privacy and ethics

The two evening frames are taken on a public street from a normal standing height and show a crowd at moderate distance. At the derivative resolution published in web/ (1600 px on the long edge), the largest face in either frame spans roughly 50–60 pixels — measured on the published file, not assumed — and every other face is smaller, turned away, or softened by motion and low evening light. That is at or below the size at which automated recognition is generally considered reliable. The claim made here is the narrow one it can support: no individual is identifiable from what is published. It is not a claim about what any particular system could do. The masters (4000 × 3000) are not committed to this repository and are excluded by .gitignore; at master resolution some foreground faces do approach identifiability, which is the reason they are withheld rather than an incidental consequence of file size.

No crop, enlargement or annotation that would make any individual identifiable is published. Counts are reported rather than faces. This is consistent with the standing ethical position in section 12 below and with the standing condition on Method 24 in Document 6: these are residents and visitors, not subjects, and this programme takes the burden of privacy voluntarily.

11. The questions this opens

Q56. Which way does automatic gain control bias a duty cycle, and by how much?

This question was framed in v1.0 to defend a number that is now withdrawn. It is kept, and its value has changed rather than disappeared: it is no longer about rescuing one figure, it is about whether any duty cycle can be computed from consumer capture — including from the shielded, metered capture arriving this week, which will still run through a compressor unless the capture path is explicitly changed.

A duty cycle of this kind is computed by counting the fraction of a clip sitting above a threshold relative to that clip's own rolling median.

That estimator and automatic gain control interact, and the interaction has a sign. A compressor pulls loud passages down toward the median and lets quiet stretches back up. Both movements shrink the measured excursion above the median. So a duty cycle computed from AGC-processed audio is biased low, and by an unknown amount.

The direction is arguable from first principles. The magnitude is not. It could be 1% or it could be 20%, and nothing in this repository distinguishes those.

This is answerable on a bench, in an afternoon, for nothing. Play a synthetic signal with a known duty cycle through a speaker, record it on the same handset through the same default capture path, and compute the same statistic. The known input gives the answer directly. Repeat under AudioSource.UNPROCESSED — which Document 5 establishes is queryable on Android — and the difference between the two runs isolates the compressor from everything else in the chain.

Three outcomes, all useful:

• The bias is small, and duty cycles from consumer capture are usable with a stated correction.
• The bias is large, and no duty figure from a phone means anything unless the capture path is changed first — which would apply to the new material as much as to the old.
• The bias depends on the signal, in which case no duty cycle of this kind is computable from consumer capture at all, and Document 5's claim that relative quantities survive an uncalibrated chain needs narrowing.

There is no way to find out this is wrong except by running it, and it should be run before the shielded capture rather than after, because it determines whether that capture needs UNPROCESSED enabled to be worth taking.

Q57. What can a timed photographic survey at two or three vantage points establish about corridor presence that turnstile arithmetic cannot?

Two frames 57 seconds apart proved existence and gave an approximate count at one instant. A deliberate survey — timed frames at known intervals from fixed positions — could establish a flow rate, which is the quantity Method 28 asks for and which no existing dataset in this programme provides. The question is whether the information gain justifies the privacy cost: every frame of a deliberate pedestrian survey is a frame taken because people are in it, which is a different ethical posture from incidental capture. The standing condition on Method 24 applies.

Q58. Does the direction of travel change the noise?

Every level in this corpus is direction-blind: the MTA memos report a receptor, not a heading, and the schedule feed shows the two directions very nearly balanced in every period (weekday daytime 372 northbound / 367 southbound), so no existing instrument can distinguish a Manhattan-bound crossing from a Brooklyn-bound one. The operator states directly that direction appears to matter and should be captured next session — rated 2/5, OPERATOR RECOLLECTION of an event, not a finding.

Settled by Method 39: two timecode-synchronised recorders separated 50–80 m along the fitted track bearing of 336.93°, at matched perpendicular offset. The order of the two level peaks gives direction; the interval gives speed: v = d / (t₂ − t₁). It needs no calibration, because it uses no level — only the sign of a time difference. A perpendicular-offset mismatch of 20 m contributes about 0.06 s against a signal of several seconds, so the geometry is forgiving. The session that establishes it should also log direction by eye, because that is what validates the array against ground truth and licenses its use unattended afterwards. See FIELD-KIT.md for the operational specification.

Q59. How long does a crossing remain audible, and what fraction of the day does that cover?

This is the operator's "noise time on clock versus floor time." It has never been measurable here, and the reason is specific: automatic gain control actively destroys decay tails — it pushes gain up as the sound fades, so a recorded tail reads both longer and shallower than it is. Any duty cycle computed from such a record is a statement about Samsung's compressor rather than about the railway.

During the stopwatch session (section 4 above) the operator was listening for the first train noise in either direction, so the stopwatch marked onsets, not durations. That is consistent with the finding that the 63.8 s cycle is pairing-independent — the tail was never measured by any instrument in this corpus. The threshold sweep in section 5 quantifies this gap: the same audio yields a duty figure anywhere from 44% down to zero depending only on where the threshold is placed.

Settled by Method 40. A fixed-gain linear recording makes time-above-(peak − 10 dB), time-above-(background + 3 dB) and a fitted decay constant all computable, and all of them are gain-independent. The measurement is background-limited: a quiet-hour session will measure a materially longer tail than a midday one, and the difference between the two is itself informative — it answers how much of the day the tail actually covers. This converts the v1.1 threshold sweep from a caution into a deliverable. See FIELD-KIT.md.

Q60. What happens to the level when two trains cross the bridge at the same time?

The published schedule cannot answer this and never could. Every scheduled departure in the MTA feed falls on an exact :00 or :30 second — 1,073 weekday traversals return only {0: 528, 30: 545} — so the entire 0–30 s window in which two crossings merge acoustically is empty by construction. That is what forced the withdrawal of the merged-pair table in this repository's own phase 9 work, and it means observation is the only available instrument.

Two incoherent equal sources sum to +3.01 dB; a materially larger rise would indicate that simple energy summation is not what happens on a four-track suspended deck, which would be a new result. Because it is a difference between two events in the same recording, unknown gain cancels entirely — so this is answerable from the same uncalibrated consumer capture that cannot carry an absolute level.

Settled by Method 41, which is pure observation during any field session and costs nothing beyond being present when it happens. At a weekday daytime rate of 61.6 crossings per hour, the expected waiting time for a simultaneous pair is short. See FIELD-KIT.md.

12. Ethics

People appear in the lawn frame at a distance, in a public park, incidentally and unidentifiably. The two evening stills on Washington Street show a crowd at moderate distance; at the published derivative size no face is individually identifiable, and the masters are not committed. No face is resolvable at the published size, and the published audio carries no intelligible speech.

The v1.1 blanket claim is withdrawn. It read "no frame was selected for a person in it", and that no longer holds for the whole set. The two evening stills were taken for the people in them — that is the operator's stated intent, it is why they were reassigned to the cohort study, and pretending otherwise would be claiming a weaker ethical exposure than the material actually carries. It remains true of every other frame here. What replaces it is a narrower and checkable commitment: counts are reported rather than faces, and no crop, enlargement or annotation that would make any individual identifiable is published anywhere in this repository. That is the constraint a deliberate pedestrian survey would also have to meet, which is why Q57 asks whether the information gain justifies the privacy cost before any such survey is designed.

That is the position Document 5 set out and it is carried voluntarily: this programme argues for the interests of the people it would incidentally record, so it takes the burden rather than waiting to be made to.

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