The Artificial Wasteland · the National Bridge Inventory, 1992 to 2025

No Field for the Repair

A bridge gets worse with time and better only when somebody works on it, so a record of bridge condition should fall slowly and rise rarely. Across all 34 published snapshots of the National Bridge Inventory, 49,764,872 year-on-year readings of 1,039,109 American bridges, 1,147,581 of them rise: two for every five that fall. 92.9 per cent of the rises have nothing recorded that could explain them, and the reason is not negligence. The inventory has no field for most repairs.

Every highway bridge in the United States is inspected at least every two years, and an inspector writes down three numbers: the condition of the deck, of the superstructure, and of the substructure, each on a scale from 9 down to 0. Those numbers go into the National Bridge Inventory. They decide which bridges count as being in poor condition, and that count is a federal performance measure, a line in the news, and an argument about money.

A bridge is a physical object. It gets worse with time and better only when somebody works on it. So the record of its condition should be a staircase going down, with the occasional step up where a crew came. Here is one bridge's record. Take a few.

Two rise for every five that fall

Joining all 34 published snapshots on the bridge gives 49,764,872 places where one component's rating in one year can be set against the same component's rating in the next. Most of them do not move. Of the ones that do:

The share that is upward is not steady. It runs near three and a half per cent of all transitions in the mid 1990s and settles near one and a half per cent after 2000, with spikes that belong to particular states in particular years rather than to the weather.

Item 106 is not a work log

The inventory does have a field for work: item 106, year reconstructed. Of the 1,147,581 upward movements, 1,065,846, or 92.88 per cent, have nothing in it that falls anywhere near the year the rating rose.

The obvious reading is that somebody failed to write the work down. That reading is wrong, and the Recording and Coding Guide says so in its own words. To count as reconstruction, the work

“must have been eligible for funding under any of the Federal-aid funding categories.”

and the Guide then lists eligible work that is expressly not to be counted:

“Painting of structural steel. … Overlay of bridge deck as part of a larger highway surfacing project … Emergency repair to restore structural integrity to the previous status following an accident. Retrofitting to correct a deficiency which does not substantially alter physical geometry or increase the load-carrying capacity. Work performed to keep a bridge operational while plans for complete rehabilitation or replacement are under preparation (for example, adding a substructure element or extra girder).”

Read that list again. Emergency structural repair after a crash. Adding a girder. Adding a substructure element. Resurfacing a deck. Painting the steel. Every one is physical work that can honestly raise a condition rating, and every one is excluded from the only completed-work field the legacy inventory has. Item 75, the other candidate, is type of work proposed.

The Guide is also explicit that a rating is relative, not absolute:

“Condition ratings are used to describe the existing, in-place bridge as compared to the as-built condition.”

which cuts the other way and is worth saying plainly. Work that restores a component toward how it was built licenses a higher number, honestly, with nothing owed to item 106. An unaccompanied rise is not evidence of anything on its own.

So the finding is not that the work is missing. It is that there is no field for it. The National Bridge Inventory, as it was specified for thirty-four years, cannot distinguish a bridge that was repaired from a bridge that a different inspector looked at. Neither can anybody reading it.

Widening the window does not rescue the number. Counting item 106 as explaining a rise if it names any year within three of the transition, rather than one, moves the unaccompanied share from 92.88 per cent to 92.06 per cent.

The control: this panel can see a repair when there is one

If the method could not detect recorded reconstruction at all, none of the above would mean anything. It can. Lining every upward movement up against the year item 106 names gives a sharp peak one and two years later, exactly where the first inspection after the work falls, and nothing before.

The two populations do not look alike

A reconstruction moves a rating a long way. A disagreement moves it by one. Splitting the upward movements by whether item 106 records anything gives two visibly different distributions: 32.0 per cent of the accompanied rises are a single point, against 77.7 per cent of the unaccompanied ones. Put the other way: 68.0 per cent of rises next to a recorded reconstruction climb two points or more, and only 22.3 per cent of the rest do.

220,613 times a rating went up and came back

The cleanest thing in the record is the excursion that undoes itself. A rating rises, and within three years it is back where it started or lower, with nothing recorded in item 106 anywhere in the window. Steel does not do that.

This is the point at which honesty costs something, so here is the cost. FHWA's own recommended framework for a bridge inspection quality programme tells a state to define out-of-tolerance for a condition rating, and offers as its worked example

“rating of +/- 1”

and the agency's own reliability study measured that spread directly. Ten inspection tasks at seven test bridges, with 49 inspectors from 25 state agencies:

“Routine Inspections were completed with significant variability, and the Condition Ratings assigned varied over a range of up to five different ratings. It is predicted that only 68 percent of the Condition Ratings will vary within one rating point of the average, and 95 percent will vary within two points.”

So the 197,866 single-point reversals here sit inside a band the agency itself describes as tolerable, and calling them errors would be wrong. The 22,747 larger ones sit at or beyond the outer edge of the measured inspector-to-inspector spread, and they are not repairs, because they went back.

A physical process does not respect state lines

If unrecorded maintenance explained the upward movements, states would differ because their maintenance programmes differ, and there would be no particular reason for a state that repairs a lot to also be a state whose ratings bounce. Across the 48 filing units holding at least 2,000 bridges, the rate of unexplained upward movement runs from 0.75 per cent of transitions in New Hampshire to 5.66 per cent in Oklahoma, a factor of 7.5. And the two quantities travel together almost perfectly.

What it costs: the country ages at half speed

Averaged over every component-year in the panel, a rating falls by 0.0360 points per year, which is one point every 27.8 years. That is the number the record hands to anybody fitting a deterioration curve, and the standard practice in that literature is to delete the upward movements first and rarely to say how many were deleted.

Here is how much they are worth. Treat every upward movement that has nothing in item 106 as though the rating had simply held, leave everything else alone, and the same average becomes one point every 15.4 years.

140,634 bridges left poor condition, 122,617 of them with nothing recorded

A bridge is in poor condition, federally, when the lowest of its applicable ratings is 4 or less. That definition became the sole basis of the federal measure for the 2018 reporting year, under 23 CFR 490; before that the label also keyed off appraisal items. Counting crossings of the 4/5 line in the upward direction, across the whole panel:

• 140,634 upward crossings, out of 1,893,339 bridge-year pairs that began in poor condition. 122,617 of them, 87.2 per cent, have nothing in item 106.
• 202,369 crossings in the other direction.
• The rate barely moved when the federal measure did: 7.62 per cent of at-risk pairs before the 2018 reporting year, 6.54 per cent from it.
• 22,875 of the reversals cross that line going up and cross back.

The claim that died

One of the seven registered predictions was that a large share of rating changes would be unwitnessed: the two snapshots carry the same inspection date, so the record changed its mind about a bridge it had not been back to look at. It measured 6.30 per cent against a registered threshold of 10, so it broke, and it deserved to. An unchanged inspection date across two annual files is the ordinary state of a bridge on a 24-month cycle, not an anomaly.

Rebuilt properly it dies completely, which is worth more than the version that broke. Restrict to bridges whose designated inspection interval is 12 months or less, require that item 90 has not moved and that none of the three critical-feature inspection dates in item 93 has moved either, so that no inspection of any recorded kind happened between the two readings. That leaves 1,122,824 transitions, and 8,169 of them move: 0.728 per cent, against a background rate of 8.102 per cent.

The record does wait for an inspector. That is the answer, it is the opposite of what was predicted, and it makes the rest of this page sharper rather than weaker: the ratings move when somebody looks, and 92.9 per cent of the upward moves have nothing behind them but the looking.

Two things found by asking a question that cannot have the answer it got

One. Asking which filing-unit-years contain no change at all should return none, because a state with 13,432 bridges inspects thousands of them every year. Five came back. Kentucky's 2003 submission carries 34,651 component ratings forward from 2002 with 0 changed, and not one of 13,423 inspection dates advanced. Rhode Island's 2005 submission does the same: 2,117 ratings, 0 changed, 0 of 748 dates advanced. These are not unchanged files. Kentucky's 2003 file differs from its 2002 file in the linear-referencing route fields for essentially every bridge, in the kilopoint, in the federal-lands flag and in the date-last-update. The file was reworked. The inspection half of it was carried over whole. Kentucky repeats the pattern in 2005 and 2007, which is what a biennial submission looks like from the outside.

Two. Item 8, the structure number, is required by FHWA's own data checks to be unique within a filing unit within a year; a duplicate is listed there as a fatal error. Across the 1,782 files, after dropping the route records that legitimately share a number, 2,858 structure numbers appear twice anyway, 1,496 of them as byte-identical fifteen-character fields and the rest differing only in padding. They concentrate in 1995. Every bridge-year involved is excluded from every count on this page, because which of the two rows survives would otherwise be decided by file order.

Nothing in the federal pipeline would ever notice

FHWA publishes the complete list of checks a state's submission is run through. All 170 of them sit inside a single record inside a single submission: the deck rating must be 0 to 9 or N, item 106 if present must be greater than item 27, and so on. Not one compares a field to the same field in the previous year; the word “prior” appears in that document only in the phrase “prior to submittal”, which is a deadline. So a rating that rises two points and falls back the next year passes every test the federal government applies to the record, because no test looks at last year.

One qualification, found by going to read the rest of the page rather than stopping at the quote above. FHWA's recommended QC/QA framework does tell a state that its own sampling for quality assurance review should consider

“Bridges with unusual changes in condition ratings (e.g. more than 1 appraisal rating change from previous inspection).”

That is a recommendation to a state about which bridges to go and look at again. It is not a check on the submitted record, it produces no field, and it produces no number anybody can count. Which is why, until this page, nobody had counted.

And FHWA has, in effect, agreed with the finding. The Specification for the National Bridge Inventory, which states first submit against in March 2026, replaces item 106 with a work data set: B.W.02, year work performed, and B.W.03, work performed, with around thirty codes covering replacement, major and minor rehabilitation, overlay, sealing and coating. The field is being built. This page measures the thirty-four years before it existed.

The pre-registration

The exclusions and seven predictions were written and committed before the analysis program was. Five held and two broke.

P6 broke narrowly and in the informative direction: the registered claim wanted at least 80 per cent of unexplained rises to be a single point and got 77.7, while the other half of the same prediction was right by a wide margin, 32.0 against a ceiling of 60. P5 is the one discussed above.

Honest edges

• An unaccompanied rise is not proof that no work was done. It is the absence of a field. This is the whole reason the argument leans on the reversals, the size distribution and the agreement between states rather than on the raw share.
• A single point is inside FHWA's own example tolerance. 89.7 per cent of the reversals here are single-point. The claim is about the 22,747 that are not, and about how many there are in total, which nobody had counted.
• The filing units are not constant. 52 in 1992: the fifty states, the District of Columbia and Puerto Rico. Guam and the US Virgin Islands enter the delimited archive in 2018, and Guam did not submit for 2025. They hold tens of bridges between them and change no rate.
• 2025 is a soft terminal year. Several states were already capturing inspections under the new specification and mapping them back to the legacy items for the March 2025 file. The series is shown whole; nothing here turns on its last point.
• A state may change how it rates without telling the file. Since the mid 1990s a state may inspect at element level and run FHWA's translator to produce the NBI ratings, and no field records which method produced a rating. NY in 1998, where 38.8 per cent of all transitions were unexplained rises, is what that looks like from outside, and it is reported as a detection rather than as an accusation.
• Reversals are truncated at the end. The window is three years, so an excursion beginning in 2023 or later cannot complete inside the panel.
• Structure numbers are not a guaranteed key. States renumber. A transition is dropped if item 27 differs between the two snapshots, which removed 89,670 bridge-year pairs. To check whether the anomalies ride on join failure instead, the rate at which bridges vanish between snapshots was correlated with the unexplained-rise rate across 1,671 filing-unit-years: Spearman 0.082. It does not.
• Item 62 is a different scale. Culverts have their own verbal descriptions in the Coding Guide, so a step of one there does not mean what a step of one on a deck means. Every table reports the four components separately as well as pooled.
• The closest prior work. Din and Tang (2016, Procedia Engineering 145) name this exact construct, “temporal outliers (i.e., bridge condition improving without reported improvement works)”, on Texas and Oklahoma, and report it qualitatively. A separate line of work reads the same signal in the opposite direction, inferring unrecorded maintenance from rating increases. Nobody, as far as a search could establish, has counted them nationally.

curl -O https://artwaste.land/checks/verify-no-field-for-the-repair.mjs
node verify-no-field-for-the-repair.mjs