The guide · Case study · Record

The complete record of the TicketTracker research run

For checking, not for reading straight through. The story, with what each step teaches, is on the case study page.

This is the full record, kept for anyone who wants to check the story against it. It has every timestamp, and each row opens to show what was asked, what came back, how it was checked, and the working file it comes from. The short labels are the run's own names for its helpers: LIT-A to LIT-D searched for existing research, FEAS checked the data, DQ listed data problems, METH and METH-1 checked methods, PUB looked for journals, and MURDER-1 and MURDER-2 were told to kill an idea (MURDER-2 the marked-car idea, MURDER-1 the street-sweeping idea). "Wave" and "phase" in the file names mean rounds of helpers.

Each brief is shown in its original wording. In the briefs, "principal investigator" means the lead agent, "subagent" means a helper agent, and "Sonnet-class" means a Claude Sonnet model. Where the run's own wording claimed more than its tests showed, a dated note has been added to the row; the original wording is kept.

34 of 34 rows

Before the run: planning the process
What happened
Tyler worked out the multi-step process that drove the run, built around helpers trying to disprove each idea, in a separate ChatGPT conversation and pasted it into the Claude session by hand. No research ran in ChatGPT.
Outcome
The lead agent ran the pasted prompts and adapted them. The later stages of the process (narrowing to one remaining idea, and writing full documentation) came from the same ChatGPT conversation.
Source files
README.md (section 6, provenance note) · 03_SUBAGENT_LOG.md (phase 2)
Setup: the brief and the ground rules
What happened
The goal: decide whether LA parking-ticket data can support a genuinely new, publishable paper. Ten candidate questions would each be checked at the same time for whether they were new, whether the data could support them, data quality, method and where to publish, and then ranked.
In its own words
"Do not optimize for producing 10 ideas. Optimize for finding one genuinely good paper. If 9 ideas are already published and only 1 survives, that is an excellent outcome."
Detail
A strict test of what counts as new, with named outright rejects, among them "novelty = we used LA instead of city X" and repeating a known result on a bigger dataset.
A firm ethical limit: historical, aggregate, offline analysis only; never a real-time tracker of parking officers.
A fixed list of fourteen fields to fill in for every idea that survived.
Source files
00_MASTER_PROMPT.md
What happened
Two rounds: first, whether each idea was new and whether the data supported it; then methods and journals, only for the ideas left. A helper rating an idea green (new) was a claim for the lead to check, not a decision.
Detail
Accept a green rating only if the helper names the closest existing papers and says how each differs.
Failing to find a paper with the exact title does not show that an idea is new.
When helpers disagree, the lead logs it and settles it on the evidence instead of averaging.
Every report is saved word for word before the lead combines them, and is treated as data, not as instructions.
Source files
01_METHODOLOGY.md
What happened
Ten one-sentence research questions, each tied to the fields it needs: street-sweeping timing, out-of-state plates, how long after registration expires a car is ticketed, a vehicle fingerprint (which makes, body styles and colors get which tickets), whether one ticket makes more nearby more likely, a grace period at meters, fairness across neighborhoods, fine changes, scaling (whether a few places get most tickets), and a four-way interaction.
Detail
Written before any helper ran: the neighborhood fairness question (candidate 7) is the one most likely to be rejected, kept only to confirm that.
Candidates 1 and 6 depend entirely on the full data recording times to the minute.
Outcome
All ten go to the first round of helpers, including the one expected to fail.
Source files
02_CANDIDATES.md
First round: eight helpers, plus the lead's own checks of the live data
What happened
Searched whether anyone had studied how tickets bunch in time around the start of street sweeping, or a grace period at meters revealed by ticket times.
Its brief (excerpt)
YOUR JOB: Determine whether this has essentially already been done. ... Be skeptical and specific. Distinguish "interesting" from "publishably novel."
You are a literature-novelty research subagent for an academic feasibility study. Your report will be saved verbatim and audited, so cite real, locatable sources and flag uncertainty explicitly. Do NOT overclaim novelty. CONTEXT: We have the Los Angeles public parking-citation dataset (data.lacity.org, Socrata dataset 4f5p-udkv): tens of millions of individual citations with issue date, issue time (HHMM, potentially minute-resolution), lat/lng, violation code + description (street sweeping, expired meter, time-limit/overtime, red zone, etc.), and fine amount. LA also publishes street-sweeping route/schedule data and posted-sign data. TWO CANDIDATE RESEARCH QUESTIONS TO ASSESS FOR NOVELTY: (1) Around posted STREET-SWEEPING start times, is there a sharp discontinuity in citation intensity, and what HAZARD FUNCTION governs citations within the swept window? (regression-discontinuity-in-time and/or survival/hazard framing of noncompliance timing) (6) Relative to METER / TIME-LIMIT expiration boundaries, is there a systematic enforcement "grace period" visible as a discontinuity in the distribution of citation times? The broader framing is "temporal signatures of parking regulations": do different regulations produce distinct temporal noncompliance distributions, with measurable hazard/discontinuity around regulatory thresholds, λ(t | regulation) or P(T_citation − T_restriction). YOUR JOB: Determine whether this has essentially already been done. Search Google Scholar / the web broadly. Use exact phrasings AND synonyms AND alternative terminology from transportation research, urban planning, economics, behavioral science, and statistics. Specifically search: parking violation temporal patterns; parking enforcement timing; parking citation hazard model; street sweeping compliance; regression discontinuity in time enforcement; parking grace period enforcement; citation timing around parking restrictions; parking behavior regulatory thresholds; survival analysis parking; bunching around notches/thresholds (public economics) applied to parking or traffic. Also find older foundational work (e.g., notch/bunching econometrics, street-sweeping citation studies). DELIVER a markdown report with EXACTLY these sections: ## VERDICT (one per candidate: 🟢 strongly novel / 🟡 adjacent literature exists / 🔴 already done) ## CLOSEST PAPERS (5 to 10; each: title, year, authors, venue, link/DOI, what they did, how our question differs) ## NOVELTY SENTENCE (the specific "To our knowledge, no previous study has..." sentence, ONLY if the search genuinely supports it; otherwise write "NOT SUPPORTED" and say why) ## KEY RISKS TO NOVELTY (what an academic reviewer would cite to say "this is already known") ## CONFIDENCE (low/medium/high + what would change it) Be skeptical and specific. Distinguish "interesting" from "publishably novel."
What it found
Rating: yellow for both. Related research exists, but neither question had been done.
Closest earlier work for candidate 1: CURB (curb.guide), a civic project that described the burst of tickets after sweeping starts in San Francisco.
The method exists in other settings: Hausman and Rapson (2018) on regression discontinuity in time (comparing just before and just after a cutoff); Gonçalves and Mello on officers bunching tickets just below speeding-fine thresholds.
How it was checked
The lead checked the report against its rule: a green or yellow rating is accepted only with the closest papers named and the reason each differs (03_SUBAGENT_LOG, wave 1 results).
What the lead did with it
Accepted at yellow. The possible new angle: a formal before-and-after-the-cutoff analysis of LA's minute-level data, compared across ticket types. Confidence medium, because the helper could not reach Google Scholar or TRID (a transportation research database).
Source files
findings/LIT-A.md · 03_SUBAGENT_LOG.md
What happened
Searched whether self-exciting point process models (Hawkes, ETAS), in which each event raises the short-term chance of more nearby, had been applied to parking enforcement.
Its brief (excerpt)
Distinguish "applying a known method to a new dataset" (weak) from "a genuine methodological or empirical contribution" (strong). Be explicit about which this would be.
You are a literature-novelty research subagent for an academic feasibility study. Your report will be saved verbatim and audited. Cite real, locatable sources; flag uncertainty; do NOT overclaim novelty. CONTEXT: We have the Los Angeles public parking-citation dataset (data.lacity.org, Socrata 4f5p-udkv): tens of millions of individual citations with timestamp, lat/lng, violation type, fine, and issuing agency. This is ENFORCEMENT output (where/when officers wrote tickets), an offline historical record, NOT real-time. CANDIDATE RESEARCH QUESTION TO ASSESS: (5) Can historical parking citations be modeled as a SELF-EXCITING SPATIO-TEMPORAL POINT PROCESS (e.g., Hawkes / ETAS-type), and does estimated self-excitation / triggering reveal latent officer-route or patrol-cluster STRUCTURE, as a purely offline, aggregate statistical characterization of enforcement? Ethical scope for framing: strictly offline, historical, aggregate. NOT a real-time officer tracker or evasion tool. Frame novelty around statistical characterization of enforcement processes, not prediction of live officer locations. YOUR JOB: Determine whether this has essentially already been done. Crime/policing uses Hawkes processes heavily (Mohler et al. 2011 self-exciting crime; predictive policing literature). Search whether SELF-EXCITING / point-process modeling has been applied specifically to PARKING enforcement or parking citations, and whether "enforcement as a self-exciting process" or "recovering patrol structure from citation point patterns" exists. Search: self-exciting point process crime/policing; Hawkes process enforcement; spatio-temporal point process citations; parking enforcement point process; patrol route inference from enforcement data; ETAS applied to urban events; log-Gaussian Cox process urban enforcement; predictive policing critiques (relevant to framing). Also foundational: Ogata ETAS, Mohler 2011, Reinhart 2018 review of spatio-temporal point processes. DELIVER markdown with EXACTLY: ## VERDICT (🟢/🟡/🔴) ## CLOSEST PAPERS (5 to 10; title, year, authors, venue, link/DOI, what they did, how ours differs) ## NOVELTY SENTENCE ("To our knowledge, no previous study has...", or "NOT SUPPORTED" + why) ## KEY RISKS TO NOVELTY ## CONFIDENCE (low/medium/high + what would change it) Distinguish "applying a known method to a new dataset" (weak) from "a genuine methodological or empirical contribution" (strong). Be explicit about which this would be.
What it found
Rating: green, but modest. These models are standard for crime (Mohler et al. 2011) and traffic accidents, and none had been applied to parking enforcement.
The new part would be using an existing method in a new area.
How it was checked
Lead review. The DQ helper separately raised the same risk of circular reasoning. The lead's second live data check showed a route field exists, so patrol routes are not entirely hidden.
What the lead did with it
Accepted as green but modest, on condition that it be described as a study of enforcement, not of driver behavior. Later moved down to a secondary idea.
Source files
findings/LIT-B.md · 03_SUBAGENT_LOG.md
What happened
Searched four questions about vehicles: out-of-state plates, the lag after registration expires, a fingerprint of make, body style and color, and a four-way interaction.
Its brief (excerpt)
Be skeptical; "we have unusual fields" is only novel if the QUESTION is unanswered, not merely the dataset. Say which candidate is strongest and why.
You are a literature-novelty research subagent for an academic feasibility study. Your report will be saved verbatim and audited. Cite real, locatable sources; flag uncertainty; do NOT overclaim novelty. CONTEXT: The Los Angeles public parking-citation dataset (data.lacity.org, Socrata 4f5p-udkv) is unusual in publishing per-citation VEHICLE ATTRIBUTES: `make` (e.g., TOYT, HOND), `body_style` (passenger, truck, etc.), `color`, plus `rp_state_plate` (registration STATE of the plate, e.g. CA/AZ/NV, not the plate number) and `plate_expiry_date` (registration expiration). Also timestamp, lat/lng, violation type, fine. Most cities do NOT publish these vehicle fields, which is central to the potential novelty. FOUR CANDIDATE QUESTIONS TO ASSESS: (2) Do OUT-OF-STATE-registered vehicles (`rp_state_plate` ≠ CA) show violation-type / spatial / temporal profiles distinct from in-state, net of location exposure? (3) What is the distribution of the LAG between registration expiry (`plate_expiry_date`) and citation date, and does it reveal targeted enforcement of expired-registration vehicles? (4) Do vehicle attributes (make × body_style × color) map to violation types beyond exposure, a "violation fingerprint" of vehicle classes? (10) Does the 4-way interaction vehicle-type × violation-type × time × geography reveal structure invisible to each margin alone (e.g. oversized/commercial vehicles clustering in overnight or oversize violations spatially/temporally)? YOUR JOB: Determine whether these have essentially been done. Search broadly across transportation, urban planning, economics, criminology, behavioral science. Search: vehicle characteristics parking violations; out-of-state plates parking/traffic enforcement; scofflaw registration; expired registration enforcement; vehicle make/color and traffic citations; commercial vehicle parking violations; vehicle type violation heterogeneity; disparate enforcement by vehicle; luxury vehicle parking behavior; tourist/visitor parking violations. Note: there is a body of work on race/demographics and traffic stops (Stanford Open Policing) and on vehicle value and rule-breaking (e.g., high-status-car studies), assess whether these already cover our questions or are distinct. DELIVER markdown with EXACTLY: ## VERDICT (one per candidate 2,3,4,10: 🟢/🟡/🔴) ## CLOSEST PAPERS (5 to 10; title, year, authors, venue, link/DOI, what they did, how ours differs) ## NOVELTY SENTENCE (per surviving candidate, "To our knowledge, no previous study has..." or "NOT SUPPORTED" + why) ## KEY RISKS TO NOVELTY ## CONFIDENCE (low/medium/high + what would change it) Be skeptical; "we have unusual fields" is only novel if the QUESTION is unanswered, not merely the dataset. Say which candidate is strongest and why.
What it found
Ratings: fingerprint (candidate 4) green and the strongest of the four; four-way interaction green; out-of-state plates and registration lag yellow.
Brazil (2020, Urban Affairs Review) used this same LA dataset to study fairness across neighborhoods, never individual vehicles.
Kawamura et al. (2014), a two-way analysis of truck violations in Chicago, is the closest earlier work to the four-way question.
How it was checked
Lead review against its checking rules.
What the lead did with it
Accepted. The registration-lag question was cut back to describing the lag only: there is no record of cars that were not ticketed, so it could not show targeting.
Source files
findings/LIT-C.md · 03_SUBAGENT_LOG.md
What happened
Searched the neighborhood fairness, fine-change and scaling questions, with an instruction to reject well-covered topics firmly.
Its brief (excerpt)
Be aggressive about rejecting saturated topics. For (9), explicitly judge whether a heavy-tailed result would be a genuine contribution or just "interesting/expected."
You are a literature-novelty research subagent for an academic feasibility study. Your report will be saved verbatim and audited. Cite real, locatable sources; flag uncertainty; do NOT overclaim novelty. CONTEXT: The Los Angeles public parking-citation dataset (data.lacity.org, Socrata 4f5p-udkv): tens of millions of citations with timestamp, lat/lng, violation type, fine amount, issuing agency, and a location/route string. LA has changed parking fine amounts on specific dates over the years. THREE CANDIDATE QUESTIONS TO ASSESS: (7) Does spatial enforcement intensity track neighbourhood DEMOGRAPHICS net of parking supply (an equity question)? (8) Did discrete LA FINE-AMOUNT CHANGES act as natural experiments altering violation rates (regression discontinuity in time around dated fine increases)? (9) Does per-segment citation count follow a HEAVY-TAILED distribution (Zipf/lognormal/power-law) and do inter-citation intervals show universal SCALING / clustering regularities? YOUR JOB: Determine whether each has essentially been done. Search broadly. - For (7): LA parking equity is likely heavily studied, search Michael Manville, Brian Taylor UCLA parking; parking enforcement equity; racial disparity parking tickets; LA Controller parking analysis; ProPublica/news parking equity; ticketing and income. Assess whether (7) is already saturated. - For (8): search fine amount elasticity parking; deterrence parking fines; penalty increase natural experiment; regression discontinuity fine; price elasticity of violations; do higher fines reduce parking violations. - For (9): search Zipf's law urban events; power law crime/citations; scaling laws cities (Bettencourt, West); inter-event time distributions human activity (Barabasi bursty); heavy-tailed enforcement; is a citation-count power law a known/expected result. DELIVER markdown with EXACTLY: ## VERDICT (per candidate 7,8,9: 🟢/🟡/🔴) ## CLOSEST PAPERS (5 to 10; title, year, authors, venue, link/DOI, what they did, how ours differs) ## NOVELTY SENTENCE (per surviving candidate, or "NOT SUPPORTED" + why) ## KEY RISKS TO NOVELTY ## CONFIDENCE (low/medium/high + what would change it) Be aggressive about rejecting saturated topics. For (9), explicitly judge whether a heavy-tailed result would be a genuine contribution or just "interesting/expected."
What it found
Fairness (candidate 7): red, already well covered. Manville and Pinski (2021), Brazil (2020) and Manville et al. (2024) cover LA and fifteen other cities.
Scaling (candidate 9): red as stated. A few places getting most of the tickets is the expected result (Weisburd on crime concentration, Barabási on bursts, Bettencourt on city scaling).
Fine changes (candidate 8): yellow; it needs a clean source of variation outside the COVID period.
How it was checked
Lead review. Before any helper ran, the lead had named candidate 7 as the one most likely to be already covered.
What the lead did with it
Accepted. Candidates 7 and 9 were dropped at the first cut.
Source files
findings/LIT-D.md · 03_SUBAGENT_LOG.md · 08_AUDIT/killed_hypotheses.md
What happened
Confirmed what the ticket dataset and related public data hold by querying the live city data portal, not by reading its documentation.
Its brief (excerpt)
Verify facts against primary sources; do NOT assume fields exist, confirm them. ... CRITICAL: Determine the precision of `issue_time`. Is it HHMM (minute resolution) or only hourly? ... This single fact gates two research ideas.
You are a dataset-feasibility research subagent. Your report will be saved verbatim and audited. Verify facts against primary sources; do NOT assume fields exist, confirm them. Where you cannot confirm, say so explicitly. TASK: Establish ground truth about the Los Angeles parking-citation dataset and the auxiliary public datasets needed for downstream analysis. Use WebFetch/WebSearch against the actual Socrata portal and its metadata/API. PRIMARY DATASET: data.lacity.org "Parking Citations", Socrata id 4f5p-udkv. - Confirm the dataset still exists and is public; give the human page URL and the SODA JSON endpoint (https://data.lacity.org/resource/4f5p-udkv.json). - Confirm the FULL record count and the date range covered (earliest and latest issue_date). Use the Socrata API: e.g. $select=count(*), and $select=min(issue_date),max(issue_date). Report exact numbers if reachable; if the endpoint is unreachable from your tools, say so and report the best documented figures with source. - Enumerate the ACTUAL columns and their types from the dataset metadata / column API (try https://data.lacity.org/api/views/4f5p-udkv.json or the "API Docs"/column listing). For EACH of these, state present/absent and type: issue_date, issue_time, meter_id, marked_time, rp_state_plate, plate_expiry_date, vin, make, body_style, color, location, route, agency, violation_code, violation_description, fine_amount, latitude, longitude, and any geocoded point field. - CRITICAL: Determine the precision of `issue_time`. Is it HHMM (minute resolution) or only hourly? Look at real sample values (query a few rows: $select=issue_time&$limit=20 or similar) and report the actual observed values. This single fact gates two research ideas. - Determine how latitude/longitude are represented (WGS84 decimal degrees? state-plane easting/northing in lat/long columns? a separate 'location_1'/point column?). Report observed sample values. - Report update frequency and license. - Report how to obtain the FULL dataset (not a 25K sample): Socrata paging with $limit/$offset or $order + cursor, the CSV export endpoint, and any documented row caps or throttling / app-token needs. AUXILIARY DATASETS (confirm existence + access): - LA STREET-SWEEPING schedule/route data (LADOT / GeoHub / lacity GeoHub), posted sweeping days/times by street segment. Give dataset name + URL if found. - LA posted parking-sign / restriction data if any. - LA parking-meter inventory + rate/time-limit schedules (LADOT / LA Express Park). Give URL. - History of LA parking FINE AMOUNTS by violation and effective dates (LADOT fine schedule, city council files). Note how findable. - Any authoritative source that maps LA `violation_code` -> description -> statutory rule (so codes can be grouped). DELIVER markdown with EXACTLY: ## DATASET STATUS (exists?, URLs, license, update freq) ## RECORD COUNT & DATE RANGE (with the exact query used + whether confirmed live or from docs) ## COLUMN INVENTORY (table: field | present? | type | sample value | notes) ## issue_time PRECISION VERDICT (minute vs hour, with observed sample values) <-- most important ## COORDINATE REPRESENTATION (with sample values) ## FULL-DATA ACQUISITION METHOD (concrete steps, endpoints, caveats) ## AUXILIARY DATASETS (table: dataset | url | what it gives | confirmed?) ## UNCERTAINTIES (anything you could not confirm) Prefer primary sources. Quote exact sample values you retrieve.
What it found
25,775,853 rows confirmed live; usable from 2014; CC0 license (public domain); updated daily.
issue_time is recorded to the minute: on one sample day only 71 of 3,070 tickets (2.3%) fell exactly on the hour.
Coordinates are ordinary latitude and longitude (WGS84 degrees); the dataset description that calls them EPSG:3857 is out of date. The vin column is empty.
Related data exists: posted street-sweeping routes (krk7-ayq2, with start and end times and boundaries written as text, no map shapes) and meter time limits (s49e-q6j2).
It reported both a meter_id column and a route column present. The meter column was mostly empty, and where it was filled it held stray symbols such as "|" and "!" that could not be used as meter numbers.
How it was checked
Every figure came from a live query quoted in the report; the lead confirmed the main counts itself in its first live data check.
What the lead did with it
Accepted. Minute-level times, which two candidates depended on, were confirmed, and the related data made matching tickets to the sweeping schedule possible in principle.
Source files
findings/FEAS.md · 03_SUBAGENT_LOG.md
What happened
Listed the problems with this data that could sink any paper built on it.
Its brief (excerpt)
YOUR JOB: For this dataset, catalogue the inference threats that could sink a paper, and note which candidate analyses each threat most endangers. ... Be specific and skeptical; the goal is to kill fragile ideas early.
You are a data-quality / inference-threat audit subagent for an academic feasibility study. Your report will be saved verbatim and audited. Be brutally honest and concrete. CONTEXT: A research team wants to publish original findings using the Los Angeles public parking-citation dataset (data.lacity.org, Socrata 4f5p-udkv): individual citations with issue date, issue time (HHMM), lat/lng, violation code/description, fine, issuing agency, and vehicle fields (make, body_style, color, registration state, plate expiry). Note: the app "TicketTracker" that surfaces this data uses only a BIASED 10K/25K/50K recency-stratified SAMPLE, the research must instead use the FULL underlying dataset. Candidate analyses include: hazard/RD-in-time around street-sweeping and meter thresholds; out-of-state-plate and expired-registration profiles; vehicle-attribute violation fingerprints; self-exciting spatio-temporal point process of enforcement; fine-change natural experiments; heavy-tailed/scaling regularities. YOUR JOB: For this dataset, catalogue the inference threats that could sink a paper, and note which candidate analyses each threat most endangers. Use web search for documented data-quality issues with the LA parking dataset and general known pitfalls, plus first-principles reasoning. Address concretely: - SAMPLING/CENSUS: the fundamental "citations are enforcement output, not violations" problem, we never observe vehicles that parked illegally and were NOT ticketed, and we never observe legal parking (no denominator, no exposure). Which candidates survive this and which don't? - SELECTION/DETECTION bias: citations reflect where/when officers patrolled, not where violations occurred. Implications for the point-process "recover officer routes" idea (is it circular?) and for equity claims. - MISSING DATA: missing/zero coordinates, missing times, "future-dated" and pre-dataset junk dates, placeholder values, missing vehicle fields. Rates if documented. - TEMPORAL precision & rounding: if issue_time is officer-entered, is it rounded to :00/:05/:15? Heaping would wreck a minute-level hazard analysis, flag this as a make-or-break check. - CHANGES OVER TIME: violation code changes, fine schedule changes, enforcement policy changes (e.g., COVID-19 2020 enforcement suspensions in LA), agency/reorg changes, geocoding method changes, all threaten pooled/longitudinal analyses and natural experiments. - GEOCODING: accuracy, snapping to block centroids, duplicated coordinates at fixed points. - DUPLICATES / data-entry error. - CONFOUNDING specific to street-sweeping RD (holidays, sweeping cancellations, the posted-time vs actual-sweep-time gap) and meter grace-period (does the data even record meter expiry time per citation? likely NOT, flag). - REGISTRATION-STATE / plate-expiry interpretation pitfalls (rental fleets, dealers, expiry-date parsing, out-of-state ≠ tourist). DELIVER markdown with EXACTLY: ## TOP INFERENCE THREATS (ranked; each: threat, why it matters, which candidates it endangers, severity) ## MAKE-OR-BREAK CHECKS (specific empirical checks to run BEFORE committing, e.g., time-heaping histogram; missingness rates; whether meter-expiry time exists in the schema) ## PER-CANDIDATE SURVIVABILITY (brief line for candidates: sweeping-hazard, meter-grace, out-of-state, expired-reg, vehicle-fingerprint, point-process, fine-RD, scaling, does it plausibly survive the data quality issues? yes/marginal/no + why) ## DOCUMENTED ISSUES (any sourced, with links) ## RECOMMENDATION (which 2 to 3 candidates are most defensible on data-quality grounds alone) Be specific and skeptical; the goal is to kill fragile ideas early.
What it found
Problem 1, which applies to every candidate: tickets record enforcement, not violations. Nobody counts the cars that were not ticketed, so rates and claims about targeting cannot be supported.
Problem 2: working out officer routes from clusters of tickets is close to circular, since the routes cause the clusters.
It also reported that the data had no meter_id and no route field.
How it was checked
The lead queried the live data (its second live data check) and found both fields: route filled in on 17.3% of a 3.76 million row slice, about one ticket in six, and meter_id on 15.3%, though its codes could not be used as meter numbers. FEAS, the helper told to check the columns directly, had already reported both present.
What the lead did with it
Partly accepted. The problems of missing counts and of selection were applied to every candidate; the claim about missing fields was overruled by the live check. DQ's underlying point, that no meter expiry time is recorded, stood.
Source files
findings/DQ.md · findings/LEAD-PROBE-2.md · 03_SUBAGENT_LOG.md · AUDIT_TRAIL.md
What happened
The lead queried the live data itself instead of relying only on the FEAS helper.
Detail
25,775,853 rows; the real dates end 2026-09-11; some junk future dates, confirmed and easy to filter out.
issue_time is to the minute, with slight rounding: 24.5% of minutes are multiples of 5, against 20% expected.
Out-of-state plates about 4.7%, roughly 1.2 million tickets.
A marked_time field exists that no helper had mentioned: the time the officer chalked the tire.
How it was checked
The exact queries are saved word for word in this check's file.
Outcome
Both timing questions pass the check on time precision. marked_time is noted for follow-up, with no claim made yet.
Source files
findings/LEAD-PROBE.md
What happened
Follow-up queries on 3,763,909 tickets from 2022 to 2023.
Detail
marked_time is filled in on 79,833 rows, about 2.1%, nearly all "PARKED OVER TIME LIMIT".
Time from chalk mark to ticket, in a first sample of 4,000: median 139 minutes; 10% under 22, 10% over 277. (A later rerun on all 48,827 usable 2022 tickets gives a median of 138; see data check 5.)
route is filled in on 17.3% of rows, about one in six. meter_id is filled in on 15.3%, but its codes cannot be used as meter numbers.
How it was checked
Counts taken live from the data portal.
Outcome
Opens the mark-to-ticket question (later one of the two main ideas), and corrects DQ's claim about missing fields.
Source files
findings/LEAD-PROBE-2.md
What happened
DQ had reported no meter_id and no route field; FEAS had reported both present, and the live check found both. One of the checking rules the lead wrote before any helper ran says that when helpers disagree, the lead settles it on the evidence rather than averaging.
Outcome
The missing-fields claim was overruled; DQ's points about missing counts and selection were kept for every candidate. The override was written down with its reason.
Source files
03_SUBAGENT_LOG.md · AUDIT_TRAIL.md · 00_MASTER/agent_registry.md
What happened
Tested, from timestamps alone, whether LA tickets show the burst after sweeping starts that CURB had shown for San Francisco. Sample: 40,000 street-cleaning tickets from 2023.
Detail
A sharp rise from 242 tickets in the 7:00 hour to 11,644 in the 8:00 hour; the 8:00 hour holds 29.1% of all sweeping tickets.
Peaks at 8, 10 and 12 match two-hour windows starting on those hours.
Within the first fifteen minutes the count ramps from 375 to 826 rather than spiking at :00.
How it was checked
Live counts by hour and by minute within the hour.
Outcome
The effect is real in LA and can be measured from timestamps, so candidate 1 is safer to pursue.
Source files
findings/LEAD-PROBE-3.md
What happened
The mix of body styles within each ticket type, 2023.
Detail
Oversize tickets are 35.7% trucks against under 1% for meter tickets, but that contrast is true by definition: oversize violations need oversize vehicles.
The less obvious contrasts are small (passenger cars are 95.7% of meter tickets against 92.8% of sweeping) and mixed up with how much each kind of vehicle is parked where.
Registration-related tickets: 17,244 in 2023, enough for the registration-lag question.
How it was checked
Live counts by ticket type.
Outcome
The fingerprint idea has a real but modest signal, and its claims to be new are toned down to match.
Source files
findings/LEAD-PROBE-4.md
Second round: methods and journals for the four ideas left
What happened
Recommended methods for each of the four remaining ideas, and the limit of what each could claim.
Its brief (excerpt)
Recommend rigorous, APPROPRIATE methods, not sophistication for appearance. Be concrete about estimators, identification, and robustness checks. Flag where an approach cannot support a causal claim.
You are a statistical-methodology research subagent. Your report is saved and audited. Recommend rigorous, APPROPRIATE methods, not sophistication for appearance. Be concrete about estimators, identification, and robustness checks. Flag where an approach cannot support a causal claim. VERIFIED DATA (Los Angeles parking citations, Socrata 4f5p-udkv; treat as fact): - 25.8M citations, real coverage 2014-present, daily; CC0. - issue_time is MINUTE resolution (mild heaping: ~24% of minutes divisible by 5 vs 20% uniform). - Coordinates loc_lat/loc_long are WGS84 decimal degrees. - Fields include: violation_code/description, fine_amount, agency, rp_state_plate (registration state; ~95% CA, ~4.7% out-of-state ≈ 1.2M rows), plate_expiry_date (YYYYMM), make, body_style(+desc), color(+desc), location (free-text address), route (~17% populated), meter_id (present but junk values, unusable). - marked_time: officer chalk time, populated for ~2% of citations and ~exclusively for "PARKED OVER TIME LIMIT" violations; issue_time − marked_time (enforced dwell) has median 139 min, p10-p90 = 22 to 277 min (2022 to 23). Selection caveat: only vehicles still present AND cited at the officer's return pass are observed (survivorship). - Auxiliary: "Posted Street Sweeping Routes" (krk7-ayq2: route_no, time_start, time_end, prose boundaries, no geometry); "LADOT Metered Parking Inventory" (s49e-q6j2: per-location TIME LIMIT e.g. "2HR" + lat/lng). - CORE LIMITATION (binding): citations are enforcement OUTPUT, not violation incidence, no denominator (we never see unticketed illegal parking, legal parking, or vehicles present). So rate/probability/"risk" and causal-targeting claims are generally unsupported; conditional-on-citation distributions are supported. COVID enforcement suspension Mar-Oct 2020 is a structural break to exclude/model. FOUR SURVIVING RESEARCH IDEAS, give methodology for each: S1. VEHICLE-ATTRIBUTE VIOLATION FINGERPRINT: do (make × body_style × color) classes show distinct violation-TYPE distributions, net of spatial/temporal EXPOSURE? (associational, not causal). How to define/estimate a "fingerprint" and correct for exposure (where/when a vehicle class tends to be parked) using only citation data, e.g., conditional composition models, multinomial/Dirichlet, log-linear/correspondence analysis, permutation nulls, effect sizes vs mere significance at N=millions. S2. STREET-SWEEPING CITATION-TIMING HAZARD / RD-IN-TIME: characterize the hazard of citation relative to posted sweeping start; test for a discontinuity at the start time; compare temporal signatures ACROSS violation types (sweeping vs meter/time-limit vs red zone). Address: RD-in-time (Hausman-Rapson) validity, minute-heaping robustness, holidays/sweeping cancellations, posted-vs-actual start gap, autocorrelation. S3. REVEALED ENFORCEMENT GRACE / OFFICER RE-INSPECTION CADENCE (time-limit parking via marked_time): model the distribution of chalk→citation intervals and, joining posted time limits (s49e-q6j2), the implied grace = dwell − limit. Handle survivorship/selection, mixed posted limits (1HR/2HR), and what can/can't be claimed (enforcement behavior vs driver behavior). S4 (secondary). SELF-EXCITING SPATIO-TEMPORAL POINT PROCESS OF ENFORCEMENT: Hawkes/ETAS on citation space-time, framed as characterizing ENFORCEMENT clustering (not violation behavior); use of the route field; the circularity problem (self-excitation ≈ patrol route by construction) and how/whether it can be made a real contribution (e.g., estimating triggering kernels, comparing across agencies/times, separability tests). DELIVER markdown with, for EACH of S1-S4: ## <Sx name> - Primary method + estimator (specific) - Identification / what claim it licenses (and what it does NOT) - Key robustness checks / falsification tests - Minimum viable analysis vs full version Then a final ## METHOD RANKING (which idea has the cleanest method-to-claim fit, which is riskiest) and ## CROSS-CUTTING PITFALLS. Prefer well-justified standard methods over exotic ones.
What it found
Best match of method to claim: the mark-to-ticket gap, described by its distribution and percentiles.
Street-sweeping timing: the jump at the start time is partly guaranteed by the rule, so any contribution would be in the shape after it and the comparison across ticket types.
Riskiest: the model of one ticket making more nearby more likely, which is close to circular because patrol routes create the pattern it would find.
How it was checked
Lead review; consistent with DQ's list of problems and the lead's fourth live data check.
What the lead did with it
Accepted, and used in the ranking at the first cut.
Source files
findings/METH.md · 03_SUBAGENT_LOG.md
What happened
Matched each of the four remaining ideas to legitimate journals, with realistic chances of acceptance.
Its brief (excerpt)
Recommend only LEGITIMATE venues (no predatory / pay-to-publish). Be realistic about contribution level required and acceptance odds.
You are a publication-venue assessment subagent. Your report is saved and audited. Recommend only LEGITIMATE venues (no predatory / pay-to-publish). Be realistic about contribution level required and acceptance odds. CONTEXT: A researcher has ~25.8M Los Angeles parking citations (public, CC0; minute-level timestamps; vehicle attributes; registration state/expiry; lat/lng) plus auxiliary LA datasets (posted street-sweeping times; metered time limits). Constraint: citations are enforcement output with NO violation denominator, so most findings are descriptive/associational or about ENFORCEMENT behavior, with a few narrow quasi-experimental angles. Work must stay historical/aggregate (no real-time officer tracking). FOUR CANDIDATE PAPERS to place: S1. Vehicle-attribute violation "fingerprints" (make×body_style×color → violation-type distributions, exposure-corrected, associational). Contribution type: empirical/urban-informatics, behavioral-adjacent. S2. Temporal signatures of parking-regulation enforcement: minute-resolution citation hazard / regression-discontinuity-in-time around posted street-sweeping start times (and cross-violation-type comparison). Contribution type: quantitative methods + transportation/urban. S3. Revealed enforcement grace period & officer re-inspection cadence in mark-and-return time-limit enforcement (using the officer chalk-time field). Contribution type: enforcement operations / behavioral / transportation. S4 (secondary). Self-exciting spatio-temporal point-process characterization of parking ENFORCEMENT activity (Hawkes/ETAS), offline. Contribution type: applied statistics / spatial data science (with predictive-policing ethics sensitivity). For EACH, deliver: ## <Sx name> - 2 to 3 realistic JOURNALS (name; why it fits; typical contribution bar; rough selectivity) - 2 to 3 legitimate CONFERENCES (name; fit; whether remote/virtual presentation is known to be possible; whether proceedings are archival) - Journal vs conference vs both? - Realistic acceptance read (Low/Moderate/High) + the single biggest reviewer objection to preempt Consider venues across: transportation (Transportation Research Part A/C, Transportation Research Record/TRB, Journal of Transport Geography, Case Studies on Transport Policy), urban science/informatics (Computers Environment & Urban Systems, Urban Informatics, EPB: Urban Analytics & City Science, npj Urban Sustainability), computational social science / data science (EPJ Data Science, Journal of Computational Social Science, Scientific Reports, PLOS ONE), applied statistics (Annals of Applied Statistics, JRSS-C, Journal of Quantitative Criminology for enforcement), and relevant CS conferences (ACM SIGSPATIAL, ACM COMPASS, EPIA, IEEE Big Data / Smart Cities). Note which are reputable-but-accessible vs top-tier. End with ## OVERALL VENUE STRATEGY: which single idea has the best venue-fit-to-effort ratio, and a realistic target for a strong-but-not-famous research team aiming for one solid publication.
What it found
Best return on effort: street-sweeping timing, the only remaining idea with a design close to an experiment, and with clear transportation journals.
First choice Transportation Research Record, then Journal of Transport Geography.
How it was checked
Lead review; the facts about each venue it relied on (in person or virtual, archival) were quoted with their sources in the report.
What the lead did with it
Accepted. Its warning that the fingerprint idea can show only associations matched what the lead's fourth live data check found.
Source files
findings/PUB.md · 03_SUBAGENT_LOG.md
First cut: ten questions down to four
What happened
The lead ranked all ten questions using the eight helper reports and its own live data checks.
Detail
Dropped: neighborhood fairness (already well covered) and scaling (an expected result).
Kept: the vehicle fingerprint, street-sweeping timing, the mark-to-ticket gap, and, as a secondary idea, the model of one ticket making more nearby more likely.
Set aside as weaker: out-of-state plates, registration lag, fine changes, the four-way interaction.
A caveat on every call about whether an idea is new: medium confidence, because Scopus, Web of Science and tracing the later papers that cite key work were not available.
Outcome
Go ahead with the four kept ideas.
Source files
04_SYNTHESIS.md · 08_AUDIT/decision_log.md · AUDIT_TRAIL.md
Narrowing to the parking-timer question
What happened
Tyler pasted in a critique from the ChatGPT conversation. It agreed with the ranking and changed what to do next.
Detail
Test the mark-to-ticket question on its own; do not combine it with sweeping yet.
Sharpen it to G, the gap minus the posted time limit, which would show how much slack officers give.
The median of 139 minutes mixes the time limit with how often officers come back; test whether the gap has several peaks, with spikes at regular revisit intervals.
First run a literature search aimed at disproving the idea.
Outcome
The lead agreed and re-ranked: the mark-to-ticket question had the best chance of being new, and street sweeping was the safest.
Source files
03_SUBAGENT_LOG.md (phase 2) · 08_AUDIT/decision_log.md
What happened
Tried to rule out the mark-to-ticket idea by finding earlier work that already does it.
Its brief (excerpt)
You are a HOSTILE novelty auditor. Your explicit goal is to KILL a research idea by finding prior work that already does it. ... Do not declare novelty just because an exact phrase is absent. ... Be adversarial. It is a SUCCESS to kill this idea if it deserves killing.
You are a HOSTILE novelty auditor. Your explicit goal is to KILL a research idea by finding prior work that already does it. Your report is saved and audited. Cite real, locatable sources. If you cannot access a database, say so explicitly. Do not declare novelty just because an exact phrase is absent. THE SPECIFIC CLAIM TO TRY TO KILL (be claim-specific, not topic-general): "Using administrative parking-citation microdata that contains BOTH an officer observation/chalk timestamp (a 'marked_time') AND the later citation timestamp, we estimate the distribution of the mark-to-citation interval Δt = T_citation − T_marked (and, net of the posted time limit, the revealed enforcement slack G = Δt − T_allowed), and we test whether that interval distribution is MULTIMODAL, i.e., whether it reveals discrete officer RE-INSPECTION / revisit intervals, for time-limit parking enforcement." DO NOT just ask "has anyone studied parking enforcement" (obviously yes). Ask the sharp questions: 1. Has anyone used a mark/observation timestamp + citation timestamp in citation data to measure enforcement revisit/re-inspection intervals? 2. Has anyone measured a parking-enforcement "grace period" or officer revisit cadence empirically from administrative data? 3. Has the LA parking dataset's marked_time field been used in any published/preprint/thesis work? 4. Would a reviewer call this merely an incremental application of survival analysis / point processes to a new dataset? SEARCH TERMS (try many combinations, and with "Los Angeles", "parking citation", "time limit", "enforcement", "survival"/"hazard"/"point process"): parking chalking; chalk time; chalking time; marked time; wheel marking; tire marking; parking enforcement marking/revisit/re-inspection; enforcement return interval; reinspection interval; parking grace period; parking time-limit enforcement; mark-to-ticket interval; observation-to-citation interval; parking officer patrol interval; parking enforcement patrol cycle; vehicle chalking; parking duration enforcement. SOURCES to attempt (report which you could/couldn't reach): Google Scholar, Crossref, Semantic Scholar, OpenAlex, TRID (trid.trb.org), arXiv, SSRN, dissertations, transportation/urban/econ/OR journals. Do forward/backward citation chasing on anything close. For EACH close paper record: citation (authors/year/venue/link-or-DOI); dataset; geography; years; sample size; variables; method; research question; principal finding; EXACT overlap with our claim; EXACT difference. DELIVER markdown: ## KILL VERDICT: KILLED / WOUNDED / SURVIVES (KILLED = essentially done before; WOUNDED = close adjacent work forces major reframing; SURVIVES = no prior work does this specific claim) ## DATABASES REACHED vs BLOCKED ## CLOSEST PRIOR WORK (with the per-paper template above; 5 to 10 items or as many as exist) ## STRONGEST SINGLE THREAT (the one paper/most likely reason this isn't novel) ## IF IT SURVIVES: the exact defensible novelty sentence, and the specific reframing needed to survive a reviewer who knows the adjacent literature ## CONFIDENCE + what would change it Be adversarial. It is a SUCCESS to kill this idea if it deserves killing.
What it found
Rating: WOUNDED, meaning weakened but not ruled out. No paper analyses LA's marked_time field, but the idea of working out patrol revisit intervals from ticket timestamps exists as US Patent 8,229,658 (Dabell, issued 2012).
A commercial product based on it, Conduent CitySight, is reported to be in use in Los Angeles.
Semantic Scholar limited the helper's requests and ScienceDirect full text was blocked, so the helper could not confirm that no closer paper exists.
How it was checked
The report lists which databases it reached and which were blocked; the lead recorded the rating with that caveat attached.
What the lead did with it
Accepted as WOUNDED. Together with the lead's fifth live data check, it set the mark-to-ticket idea at CONDITIONAL, leaning marginal.
Source files
05_VERIFICATION/novelty_murder_marked_time.md · 08_AUDIT/decision_log.md · 08_AUDIT_TRAIL/agent_outputs/MURDER-2.a72ee9213262c7556.jsonl
What happened
Measured the mark-to-ticket gap on every ticket with a mark time, and on a 2022 slice.
Detail
Tickets with a chalk time, across all history: 729,430.
2022, all 48,827 usable tickets: clearly two humps, near 70 and 135 minutes, median 138.
No spikes at fixed multiples of 15, 30 or 60 minutes: under 1% of tickets near each.
Every ticket in both humps carries one code, 80.69C, "PARKED OVER TIME LIMIT"; the posted limit is not recorded.
How it was checked
Live queries; the analysis can be rerun with the scripts in 03_PIPELINE.
Outcome
The data weakened the most interesting version of the idea: the two humps match one-hour and two-hour posted limits, not officers returning on a regular schedule. Note added 2026-10-22: the check looked only for spikes at multiples of 15, 30 and 60 minutes. Humps near 70 and 135 minutes would also come from a patrol loop of about 65 to 70 minutes, so the limits reading stays an inference until each ticket is matched to its posted limit, as the run's own claim register says.
Source files
06_RESULTS/findings.md · 08_AUDIT/decision_log.md
What happened
The idea that officers come back at fixed intervals (the "discrete re-inspection comb") was disproven on the 2022 slice. What remains, slack beyond the posted limit, needs each ticket matched to its posted limit, which public data may not support.
Outcome
Verdict on the mark-to-ticket idea: CONDITIONAL, leaning marginal. The evidence was not enough to go ahead and did not rule it out; a defined set of checks could move it either way. Note added 2026-10-22: "disproven" here means no spikes at 15, 30 or 60 minute multiples. A revisit loop of about 65 to 70 minutes would produce humps near 70 and 135 and was not tested.
Source files
08_AUDIT/decision_log.md · 08_AUDIT/killed_hypotheses.md · STATUS.md
Late night: separate helpers check each step
What happened
Tyler pasted in the later stages of the process: a numbered research folder, a list of every helper agent, a register of each claim's status, and a rule that two to four separate helpers check each major step before the lead moves on.
In its own words
"YOU MUST SPAWN A SMALL INDEPENDENT AUDIT TEAM BEFORE MOVING FORWARD. For EVERY major stage, spawn approximately 2-4 independent Sonnet-class subagents."
Outcome
Two helpers sent to check the street-sweeping idea: one to try to disprove it as new, and one to check what its design could actually prove.
Source files
00_MASTER/ (registry, charter, hypotheses, master timeline) · AUDIT_TRAIL.md · the protocol text as supplied in the session, quoted on the summary timeline
What happened
If enforcement follows the posted start, sweeping tickets should crowd into the start of the hour, while ticket types with no on-the-hour start, used as a placebo (a comparison where no effect is expected), should spread evenly. 2023, 20,000 tickets per type.
Detail
Red zone 33.4% / 33.6% / 33.0% across the three twenty-minute thirds of the hour; disabled placard 33.5% / 32.9% / 33.6%; meter expiry close to even.
Street sweeping only 34.8% in the first third, against 33.3% for an even spread: weak crowding once all sweeping windows are pooled.
No-stopping tickets are not a valid placebo: they have on-the-hour windows of their own.
How it was checked
Live counts; METH-1 later reread the result (see its row below).
Outcome
The spike by hour does not show up as a sharp effect within the window unless each ticket is matched to its own sweeping start time.
Source files
06_RESULTS/placebo_minute_of_hour.md
What happened
Both remaining ideas need each ticket matched to the posted rule on its street: a sweeping start time for street sweeping, a time limit for the mark-to-ticket gap.
Detail
Matching on the route field fails. The schedule's route numbers (e.g. 10P136 Th) and the tickets' enforcement beat codes (e.g. 0350A) are different code systems.
Matching by location is required, which means turning 864 written boundary descriptions into map shapes. Not attempted here; it needs a proper mapping (GIS) build.
Outcome
Matching by route code recorded as ruled out. Matching tickets to posted rules becomes the main obstacle for both ideas.
Source files
03_PIPELINE/joins/regulatory_join_feasibility.md · 08_AUDIT/claim_status.md
What happened
Tried to rule out the street-sweeping timing idea by finding earlier work that already does it.
Its brief (excerpt)
You are a HOSTILE novelty auditor. Goal: KILL a research idea by finding prior work that already does it. ... do NOT declare novelty just because an exact title is absent. ... Killing it if it deserves killing is a success.
You are a HOSTILE novelty auditor. Goal: KILL a research idea by finding prior work that already does it. Report is saved and audited. Cite real, locatable sources; state which databases you could/couldn't reach; do NOT declare novelty just because an exact title is absent. CLAIM TO TRY TO KILL (be claim-specific): "Using minute-resolution Los Angeles parking-citation microdata (~25.8M records, 2014-present), we estimate the CITATION HAZARD as a function of time since the posted STREET-SWEEPING start time, characterize the post-onset hazard SHAPE (rise/peak/decay, not merely the mechanical jump at t=0), and COMPARE these temporal signatures ACROSS violation types (street-sweeping vs meter/time-limit vs red-zone as placebo) within a regression-discontinuity-in-time / event-time-hazard framework." Sharp questions (answer each): 1. Has anyone estimated a hazard/event-time model of parking-citation issuance relative to a posted regulatory start time (street sweeping especially)? 2. Has anyone used regression-discontinuity-IN-TIME on parking/street-sweeping enforcement? 3. Has anyone compared temporal citation signatures across violation types as a placebo/identification design? 4. Does the CURB app (curb.guide) or any civic/industry project have a preprint/methods writeup that would count as prior art for the street-sweeping timing result? 5. Would a reviewer call this an incremental application of RDiT/hazard methods to a new dataset? SEARCH (many combinations + synonyms): street sweeping enforcement/citation timing; street cleaning parking enforcement; parking citation hazard/survival; regression discontinuity in time parking/enforcement/street sweeping; event-time analysis citations; bunching at enforcement thresholds; citation timing around restriction start; parking enforcement temporal discontinuity; officer arrival time street sweeping. SOURCES to attempt (report reached vs blocked): Google Scholar, TRID (trid.trb.org), Crossref, OpenAlex, Semantic Scholar, arXiv, SSRN, Google Patents, transportation journals (TRR, TR Part A/C, J Transport Geography, Transport Policy). Forward/backward citation-chase Hausman & Rapson 2018 (RDiT), Gonçalves & Mello (enforcement bunching), and anything on street-sweeping citations. For each close paper: citation (authors/year/venue/link/DOI); dataset; geography; years; N; variables; method; question; finding; EXACT overlap; EXACT difference. DELIVER markdown: ## KILL VERDICT: KILLED / WOUNDED / SURVIVES ## DATABASES REACHED vs BLOCKED ## CLOSEST PRIOR WORK (5 to 10, template above) ## STRONGEST SINGLE THREAT ## IF IT SURVIVES: defensible novelty sentence + reframing needed to survive a knowledgeable reviewer ## CONFIDENCE + what would change it Be adversarial. Killing it if it deserves killing is a success.
What it found
Rating: WOUNDED, with 60% confidence that no single paper does the full combination.
CURB (San Francisco) and an L.A. Material data-journalism analysis of the same LA dataset already show the effect publicly; a 2026 paper gives a formal model of ticket timing; Pyle et al. ran an event study (a before-and-after comparison around set times) on Chicago tickets.
Checking the roughly 379 works that cite Hausman and Rapson could not be completed.
How it was checked
Databases reached and blocked are listed in the report; the lead recorded the unfinished check of citing works as an open condition rather than treating finding nothing as evidence the idea is new.
What the lead did with it
Accepted. The idea could survive only as a better method than the public work, not as a new finding.
Source files
05_VERIFICATION/novelty/novelty_murder_H1.md · 08_AUDIT/decision_log.md · 08_AUDIT_TRAIL/agent_outputs/MURDER-1.a10792190ef6157c4.jsonl
What happened
Checked whether the street-sweeping design shows anything beyond the rule itself.
Its brief (excerpt)
You are an independent econometrics / statistical-identification auditor. Be adversarial about the identification strategy. ... Is RDiT valid here, or is the τ=0 "jump" MECHANICAL ... Do not rubber-stamp RDiT.
You are an independent econometrics / statistical-identification auditor. Be adversarial about the identification strategy. Report is saved and audited. DESIGN UNDER AUDIT: LA parking citations, minute-resolution issue_time (mild ~24% heaping at multiples of 5). For street-sweeping citations, define relative time τ = T_citation − T_sweep_start, where sweeping windows begin on the hour (commonly 8:00/10:00/12:00). Proposed analyses: (a) regression-discontinuity-in-time (RDiT) jump in citation intensity at τ=0; (b) event-time HAZARD of citation after window opens; (c) cross-violation-type comparison (sweeping vs meter/time-limit vs red-zone placebo). KNOWN FACTS: citations are enforcement OUTPUT (no denominator: unticketed illegal parking + legal parking + vehicles-present are unobserved). Observed: street-sweeping citations spike sharply at the window-start hour and RAMP UP over the first ~15 minutes (not a single spike at :00). No per-space occupancy data. Assess, adversarially and concretely: 1. Is RDiT valid here, or is the τ=0 "jump" MECHANICAL (citation is near-impossible before the legal window opens, so the discontinuity is definitional, not behavioral)? If mechanical, what, if anything, is the real estimand of scientific interest? 2. Is this better framed as an event-time HAZARD / interrupted-time-series / point-process intensity than as RD? What does each license? 3. What exactly can be claimed? (enforcement-process timing? officer arrival/traversal? NOT driver behavior or deterrence, no denominator.) State the boundary crisply. 4. Threats: running-variable manipulation/sorting; serial correlation / clustering (route-day, officer); minute-heaping effect on bandwidth and on any density/McCrary test; posted-vs-actual sweep start-time gap → fuzzy running variable; holidays/cancellations; window-length heterogeneity (1hr vs 2hr sweep blocks) and multiple start times aliasing in minute-of-hour space; selection (only cars still parked past start are cited). 5. Are meter/red-zone citations VALID placebos/negative controls, or is that a false-placebo (different mechanisms, different locations/times)? What is the correct negative control? 6. What is the minimum defensible analysis that yields a real (non-mechanical) finding, and what would make it merely descriptive? DELIVER markdown: ## IDENTIFICATION VERDICT (is there a valid non-mechanical estimand? RD vs hazard vs ITS vs descriptive) ## THE REAL ESTIMAND (what can honestly be claimed) ## TOP THREATS (ranked; each: threat, why, fix/mitigation, or fatal) ## PLACEBO/CONTROL VALIDITY (are the proposed controls real; what's the correct one) ## MINIMUM DEFENSIBLE SPECIFICATION (concrete) ## VERDICT: is the scientific content beyond the mechanical spike SUBSTANTIVE, MARGINAL, or ABSENT? Prefer standard, defensible methods. Do not rubber-stamp RDiT.
What it found
The jump at the start time follows from the rule: a sweeping ticket cannot exist before the window opens. Reporting it as an effect would only confirm that officers follow the rule.
The only defensible thing to measure is the rise in tickets after the start, described as the timing of enforcement, not as driver behavior.
That rise may come from pooling street segments with different actual start times; only matching each segment and repeating the analysis can tell.
Meter and red-zone tickets are not valid placebos for this question: they come from different processes.
How it was checked
Its conclusion about the jump matched the earlier METH helper's note on street sweeping, reached separately.
What the lead did with it
Rating MARGINAL, accepted. The design was changed from a before-and-after-the-cutoff comparison to models of ticket timing over time (point-process and interrupted time series), and the placebo result was relabeled.
Source files
05_VERIFICATION/methodology/identification_audit_H1.md · 08_AUDIT/decision_log.md · 08_AUDIT_TRAIL/agent_outputs/METH-1.af2b0805804b876b0.jsonl
What happened
Red-zone and placard tickets have no posted window, so their even spread was expected and says nothing about sweeping.
Outcome
Relabeled from evidence about enforcement timing to a check on rounding: it confirms the timestamps have no artificial on-the-hour pattern, and nothing more. The right comparisons are within the sweeping tickets themselves and need the street-by-street match. Note added 2026-10-22: the check splits the hour into thirds, so it can show only that ticket types with no start time do not bunch into the first third. It says nothing about minute-level rounding, which does exist: 24.5% of ticket minutes are multiples of 5, against 20% expected (data check 1).
Source files
06_RESULTS/placebo_minute_of_hour.md (correction section) · 08_AUDIT/claim_status.md
What happened
Every claim has a status and its evidence: verified, partial, inference, needs external, or disproven.
Detail
Verified: 25,775,853 rows; times to the minute; WGS84 coordinates; 729,430 tickets with a chalk time; the gap has two humps.
Inference: the two humps are one-hour and two-hour limits (not yet matched to the posted rules).
Needs outside confirmation: slack beyond the posted limit; matching sweeping tickets to the posted rules; whether the rise after the start is real.
Disproven: fixed revisit intervals; matching by route code; the jump at the start time as the thing to measure; the placebo as a test of enforcement timing.
Note added 2026-10-22: "fixed revisit intervals" was tested only at 15, 30 and 60 minute multiples; a loop of about 65 to 70 minutes would fit the two humps and was not tested.
Source files
08_AUDIT/claim_status.md
What happened
Each change is written in the decision log with its reason.
Detail
Street sweeping: changed from a before-and-after-the-cutoff design to point-process and interrupted time series models (METH-1: the jump is guaranteed by the rule).
Mark-to-ticket gap: changed from grace periods and revisit schedules to a mix of posted limits (from the project's own data).
Both ideas set to CONDITIONAL, contribution score 2.
Source files
AUDIT_TRAIL.md (change control) · 08_AUDIT/decision_log.md
What happened
The process called for two to four helpers at each of 25 stages and a panel of five reviewer agents at the end.
Outcome
Judged to repeat the eleven checks already run; the strongest reviewer objections were drawn instead from the existing reports by helpers told to disprove ideas. The decision was recorded and pointed out to Tyler. No reply from him, approving or objecting, is recorded.
Source files
AUDIT_TRAIL.md · FINAL_RESEARCH_READINESS.md (reviewer simulation)
The verdict
What happened
Both ideas were rated weakened on being new and marginal on what the design can prove, and both depend on matching tickets to street locations, which public LA data does not cleanly support.
Detail
Contribution score 2 out of 5, against the bar of 3 the process set for going ahead.
The lead's own estimates, not calibrated probabilities. Confidence the central contribution is novel: about 30 to 40%. That it is substantive rather than marginal: about 20 to 25%. That the evidence is ready for peer review now: about 15 to 20%, or 45 to 55% at a mid-tier transportation venue if all three conditions below clear.
Could a competent academic start writing a legitimate paper from this tomorrow? The recorded answer is no.
Outcome
Three conditions for going ahead, all required, any one able to rule the idea out: finish checking the works that cite Hausman and Rapson and search TRID directly; build and test the street-by-street match to posted rules; show the rise after the sweeping start holds for single street segments, with uncertainty ranges from resampling whole segments (cluster bootstrap).
Source files
FINAL_RESEARCH_READINESS.md · 08_AUDIT/decision_log.md
After the verdict
What happened
An index to read first, listing every document, prompt, audit file and script, with steps for someone new to reproduce the main analysis.
Detail
The exact prompts can be recovered as the first message of each raw transcript.
Thirteen transcript files exist; two labelled UNMAPPED are "stray session tasks, not part of the 11 research subagents."
Outcome
The run ends just over 23 hours after the opening brief was written.
Source files
README.md
What happened
The run's audit file and its readiness file said 13 helpers, while the role-by-role list in the same files names 11. The review checked the list role by role to settle which was right.
Outcome
The correction was added below the original text: 11 named research helpers. The 13 counted the transcript folder, which also holds the two stray UNMAPPED files.
Source files
AUDIT_TRAIL.md (correction, 2026-09-13)

← Back to the case study Analysis and drafting assisted by Claude (Anthropic). Each page says how its results were checked (recomputation, published values or expert review) and labels anything not yet checked.