Practical Engineering — The Hidden Engineering of Floating Bridges

Why this is in the vault

17-minute Grady Hillhouse explainer on floating concrete bridges — the niche-but-instructive engineering domain where the four longest permanent floating bridges on Earth all sit clustered in Washington State, where two of those four bridges have actually sunk, and where Sound Transit was just (June 2025) testing light rail across one of them. Anchored on the Lacy V. Murrow Bridge (1940 — first floating concrete highway), the 1979 Hood Canal Bridge sinking (open hatches → storm surge → western span gone), and the 1990 Lacy V. Murrow re-sinking (contractors removed watertight chamber doors to store hydro-demolition runoff → Thanksgiving storm → bridge partially sunk + severed Hadley cables). The vault keeps it for three interlocking reasons. (1) It is the canonical floating-infrastructure exemplar of CA-016 (layered-defense architecture) — modern concrete pontoons are subdivided into sealed chambers with watertight doors, leak detection systems, pre-installed pumps, specialized concrete mixes, and temperature-controlled curing. Each sub-system targets one failure mode, each is cheap relative to the catastrophe it prevents. (2) The two sinkings are textbook layered-defense-collapse case studies — both failures came from removing or violating a single defensive layer (open hatches in 1979, removed watertight doors in 1990). When operational decisions defeat one layer of a stack, the system reverts to the strength of its weakest remaining defense. (3) It is the canonical example of structure-class-determines-mitigations (CA-016 sub-pattern from the arch-vs-gravity dam pair) extended into a third structure class — floating bridges are constantly coupled to environmental forcing (waves, wind, currents, tides, ice), which closes off entire mitigation pathways available to fixed structures. The DOT enforcing wind thresholds and closing bridges to traffic is the operational expression of “you can’t engineer this layer; you have to schedule around it.”

Episode summary

17-minute Grady Hillhouse explainer on floating bridges — the niche-but-instructive engineering domain where the four longest permanent floating bridges on Earth all sit clustered in Washington State (Hood Canal, Evergreen Point, Lacy V. Murrow, Homer M. Hadley Memorial). Walks through the historical arc (1921 Homer Hadley proposal → 1940 Lacy V. Murrow opens → 1979 Hood Canal sinks → 1990 Lacy V. Murrow re-sinks → June 2025 Sound Transit light-rail testing on Hadley), the unique engineering constraints (navigation channels through pontoons, mooring/anchoring against wave/wind motion, watertight sealed chambers in pontoons against the inevitable concrete cracking, specialized concrete mixes, temperature-controlled curing), and the load-bearing thesis: floating bridges aren’t just bridges that float — they are boats that you drive cars on, constantly coupled to environmental forcing in ways fixed structures aren’t. The 1979 Hood Canal sinking (open hatches let in rain/waves until the western span flooded) and the 1990 Lacy V. Murrow re-sinking (removed watertight doors to store hydro-demolition runoff, then a Thanksgiving storm hit) are the load-bearing case studies — both failures came from defeating a single defensive layer that the entire system depended on. Closes with a Ground News sponsor read framed around the Francis Scott Key Bridge collapse NTSB recommendations.

Key arguments / segments

[00:00:00] The Seattle setup. Early-1900s Seattle hemmed by geography (Puget Sound to west, Lake Washington to east blocking access to Cascades farmland/logging). Lake Washington is glacially carved — over 200 ft / 60 m deep in places, with a 100-ft layer of soft clay/mud underneath. Conventional bridge piers would have required staggeringly sized supports. 1921: engineer Homer Hadley proposed a bridge that doesn’t rest on the bottom at all — floats on hollow concrete pontoons.
[00:01:00] The Lacy V. Murrow Bridge (1940). Federal New Deal funding (Public Works Administration) finally funded the Hadley concept. First floating concrete highway of its kind. “50 years later, this span would be swallowed by the very lake it crossed.”
[00:01:30] The Washington State cluster. Four of the five longest floating bridges in the world sit in one small area: Hood Canal, Evergreen Point, Lacy V. Murrow, Homer M. Hadley Memorial (named for the engineer). High-profile failures plus remarkable successes including June 2025 light-rail testing.
[00:02:30] Floating bridges go back to Xerxes (480 BCE). Most are military — quick to put up, quick to take down, not designed for extreme conditions. The leap to permanent floating infrastructure brought a host of new engineering challenges.
[00:04:00] Navigation problem #1: bridges block boats. Three Washington solutions: (1) Evergreen Point — elevated approach spans on either end allow ships to pass beneath. (2) Lacy V. Murrow original — retractable span pulled into a pontoon-pocket at center; created awkward roadway curves causing frequent accidents, eventually removed once east channel bridge was raised to provide alternative boat route. (3) Hood Canal — truss bands for small craft + hydraulic lift sections for large ships (US Naval Base Kitsap is nearby; sometimes opens for Navy submarines). Lift sections raise vertically while adjacent segments slide back underneath. Flexible: one side opens for tall narrow vessels, both for wider ships.
[00:05:30] Navigation problem #2: bridges ARE boats. Constant interaction with water, waves, currents, sometimes tides and ice. Easiest on calm lakes/rivers. Mooring is essential — long cables and anchors prevent overstressing and uncomfortable/dangerous motion. Anchor types: massive concrete slabs on lakebed; piles driven deep; in deep water/soft soil, anchors lowered with water hoses that jet soil away so the anchor sinks deep into mud.
[00:06:30] Wind/wave operational risk. Floating bridges sit low to the water. High winds → waves crash directly onto roadway, obscuring visibility, creating road-user risks. Motion from waves/wind flexes the bridge under vehicles (unnerving for unfamiliar drivers). DOT enforces wind thresholds for each bridge — exceed threshold, bridge closes, even if structurally sound. “In extreme weather, the bridge itself becomes part of the storm.” This is a layered-defense layer that can’t be engineered away — only scheduled around.
[00:07:30] Why concrete (counterintuitively). Civil engineering students compete in ASCE concrete canoe races every year. Grady’s recreational-math fun fact: a neutrally buoyant hollow concrete cube needs wall thickness in inches = outer dimension in feet (12-in walls → 12-ft cube). A linear factor-of-12 relationship; only fun in imperial units. Real pontoons are much bigger than the neutrally-buoyant minimum because they have to carry a deck plus traffic with safety margin.
[00:09:00] Concrete cracks, but pontoons can’t leak. Designers prevent leaks with multiple independent layers: (1) pontoons subdivided into sealed chambers; (2) watertight doors between chambers for inspection access; (3) leak detection systems for early warning; (4) pre-installed piping with pumps on standby so chambers can be pumped dry before disaster; (5) specialized concrete mixes (reduced shrinkage, water resistance, abrasion resistance); (6) temperature-controlled curing — for the Evergreen Point replacement, contractors embedded heating pipes in the base slabs of pontoons so the entire structure cooled at a uniform rate, reducing thermal stress cracking.
[00:10:30] Even the most careful builds make mistakes. Evergreen Point replacement had a post-tensioning system flaw that led to millions in change orders mid-construction and significant project delays. Layered defense doesn’t mean defect-free; it means defects don’t cascade.
[00:11:00] Sinking #1: Hood Canal Bridge, February 1979. Severe storm caused the western half to lose buoyancy. Investigation: open hatches allowed rain and waves to blow in, slowly filling the pontoons. Western half of the bridge sank. DOT had to run a temporary ferry service for nearly 4 years while the western span was rebuilt. The defensive layer that was defeated: hatches that should have been closed. One layer down, the rest of the stack couldn’t compensate.
[00:11:30] Sinking #2: Lacy V. Murrow Bridge, 1990. Failure during rehabilitation work while bridge was closed. Contractors used hydro-demolition (high-pressure water jets) to remove old concrete from the road deck. Because the runoff water was contaminated, it couldn’t be released to Lake Washington. Engineers calculated that pontoon chambers could hold the runoff safely — and to accommodate that, they removed the watertight doors that normally separated the internal compartments. Then a Thanksgiving weekend storm flooded the open chambers, the bridge partially sank, and severed cables on the adjacent Hadley Bridge — delaying the project by more than a year. “A potent reminder that even small design or operational oversights can have major consequences on this type of structure.” The defensive layer that was defeated: the watertight doors. Same architectural lesson as 1979 — the entire pontoon-chamber compartmentalization scheme depends on each layer being intact.
[00:12:30] Light rail on a floating bridge — new puzzles. Sound Transit testing trains on the Homer Hadley Bridge. Two new engineering problems: (1) Stray currents from electrified rails could damage the bridge — track is mounted on insulated blocks with drip caps to prevent water creating a conductive path. (2) Bridge motion — a floating bridge can roll, pitch, and yaw with weather/lake-level/traffic-load. Joints between fixed shoreline and the bridge must accommodate movement. Cars/trucks/bikes/pedestrians don’t care; trains require very precise track alignment. Engineers developed an innovative track-bridge system with specialized bearings that distribute every kind of movement over a longer distance, keeping tracks aligned even as the structure shifts beneath. June 2025 testing went well; more work needed before passenger service.
[00:13:30] Floating tunnels — the future. Norway has proposed a submerged floating tunnel across a fjord on its western coast. Could shorten tunnel lengths, reduce excavation costs, minimize environmental impacts. Same long list of unknowns as floating bridges had a century ago. “That’s the essence of engineering. Meeting each challenge with solutions tailored to a specific place in need.”
[00:14:30] The closing principle. “There aren’t many locations where floating infrastructure makes sense. The conditions have to be just right. Calm waters, minimal ice, manageable tides.” Floating bridges are a niche engineering pattern that only works in narrow envelopes — but where the envelope holds, they unlock connections nothing else can provide.
[00:14:45] Ground News sponsor read — Francis Scott Key Bridge framing. NTSB recommendations report on the Key Bridge collapse: vessel-collision risk was ~30x higher than minimum requirements for new bridges. Recommending owners of 68 bridges across the country do the same vulnerability analysis. Ground News pitch is built around the bias-aware reading-multiple-sources angle (left-leaning headlines vs right-leaning headlines on the same story). 40% off Vantage subscription via ground.news/practicalengineering.

Notable claims

[00:00:30] Lake Washington is glacially carved, 200+ ft deep with a 100-ft soft clay/mud layer underneath — making conventional bridge piers infeasible. This is the constraint that birthed the entire floating-bridge engineering tradition.
[00:01:30] Four of the five longest permanent floating bridges in the world are in Washington State: Hood Canal, Evergreen Point, Lacy V. Murrow, Homer M. Hadley Memorial.
[00:02:30] Earliest known floating bridge: Xerxes crossing the Dardanelles, 480 BCE.
[00:05:30] Hood Canal Bridge can open for US Navy submarines from nearby Naval Base Kitsap.
[00:07:30] Concrete-canoe rule of thumb (Grady’s fun fact): for a neutrally-buoyant hollow concrete cube, wall thickness in inches = outer dimension in feet. Only fun in imperial; mathematically a linear factor-of-12 relationship.
[00:09:30] Evergreen Point pontoon replacement embedded heating pipes in base slabs so the entire concrete structure cooled at uniform rate — thermal-stress cracking mitigation.
[00:11:00] Hood Canal Bridge sinking, February 1979 — open hatches caused buoyancy loss in storm; 4-year temporary ferry replacement during rebuild.
[00:11:30] Lacy V. Murrow Bridge re-sinking, Thanksgiving 1990 — contractors removed watertight chamber doors to store hydro-demolition runoff; storm flooded chambers; bridge partially sank, severing Hadley Bridge cables, project delayed by 1+ year.
[00:12:30] June 2025: Sound Transit successfully tested light rail on the Homer Hadley Bridge — first floating-bridge light rail. Custom track-bridge system with specialized bearings handles roll/pitch/yaw motion that trains can’t tolerate.
[00:14:00] Norway has proposed a submerged floating tunnel across a fjord — would be the first of its kind.
[00:15:00] NTSB found Francis Scott Key Bridge vessel-collision risk was ~30x higher than the minimum requirements for new bridges — recommending 68 US bridges undergo the same analysis.

Guests

None. Solo Grady Hillhouse explainer, his standard format.

Mapping against Ray Data Co

The two sinkings are the canonical “defeat-one-layer-and-the-stack-collapses” exemplars for CA-016 (layered-defense architecture). Hood Canal 1979: open hatches defeated the chamber-isolation layer. Lacy V. Murrow 1990: removed watertight doors defeated the same layer (deliberately, for what seemed like a good reason at the time). In both cases, the entire pontoon defense stack — sealed chambers, leak detection, pumps on standby, specialized concrete — depended on each layer being intact. Direct map to RDCO. The autonomous loop’s defense stack (Notion task verification + audit script invariants + founder review + cron-driven monitoring) is exactly the same architecture: defeat any one layer (e.g., disable the audit hook for “just one batch”) and the whole stack reverts to the strength of its weakest remaining defense. The Lacy V. Murrow case is especially instructive because the doors were removed deliberately for a legitimate operational reason — the equivalent in RDCO would be disabling a verifier “just for this run” because it’s slowing a backfill. Worth a one-line discipline rule: layered defenses must not be temporarily disabled — if you need to defeat one layer, you must temporarily increase another. Bake this into the SKILL.md template.
Floating bridges are the third structure-class for the structure-class-determines-mitigations sub-pattern under CA-016. The existing pair was arch dams (tightly coupled, must be replaced wholesale) vs gravity dams (loosely coupled, slot-cuttable in-flight per CA-019). Floating bridges add a third class: constantly coupled to environmental forcing (waves, wind, currents, ice). This closes off entire mitigation pathways — you can’t engineer away “wind exceeds threshold,” you can only schedule around it (DOT closes the bridge to traffic). Same shape in RDCO at the API-rate-limit layer: you can’t engineer away the Anthropic rate limit, you can only schedule around it (concurrency cap, exponential backoff, queue management). Adds environmental-coupling as the third mitigation-pathway constraint to the existing arch-vs-gravity (replace) and gravity-vs-fuse-plug (in-flight repair) distinctions.
The “boats you drive cars on” reframe is the load-bearing model for hybrid systems in RDCO. Most engineering problems decompose into well-known categories with established mitigations. Floating bridges resist the decomposition — they’re simultaneously bridges and boats, and the engineering tradition for each subdomain doesn’t transfer cleanly. Same problem with hybrid agentic systems: an AI-agent + human-review system is simultaneously an agent system (with agent failure modes) and a human workflow (with attention/notification/escalation failure modes). The RDCO equivalent of “you have to design for both bridge and boat conditions” is every hybrid skill must explicitly account for both the agent-failure surface AND the human-attention surface. The /check-board cron firing into a Discord notification is the right pattern (agent does the work, human gets pinged); the wrong pattern is implicitly assuming the founder will see session output (he won’t — see CLAUDE.md rule #2).
The mooring-anchor problem maps to dependency-anchor architecture in agentic systems. Floating bridges aren’t just stretched across the banks — they’re moored with long cables to anchors on the lakebed (concrete slabs, deep piles, or jet-set anchors in soft mud). The mooring system does double duty: structural integrity + day-to-day safety for drivers. Same shape in RDCO. The autonomous loop isn’t just “code on the Mac Mini” — it’s anchored to multiple external systems (Notion as task DB, 1Password as secrets, Cloudflare R2 as state storage, Anthropic API as the model layer, Gmail/iMessage/Discord as channels). Each anchor is essential and each anchor type has different soil conditions. The dependency-anchor audit (already queued from the spillway video as the redundancy-failure-mode audit) should specifically include the mooring-system frame: each anchor (a) is it structurally sound, (b) does it have day-to-day operational reliability, (c) is the soil it’s sunk into stable enough? The Cloudflare DNS dependency is a good test case — sound structural choice, reliable operationally, but the “soil” (Cloudflare’s own internal stability) is a single point of correlated failure across multiple anchors.
The pontoon temperature-controlled curing detail is a useful model for the “tune for predictable failure” discipline. Evergreen Point engineers embedded heating pipes in pontoon base slabs so the entire concrete structure cooled at uniform rate — preventing differential thermal stress that would cause cracks. The discipline: manage the conditions during the manufacturing process so that the failure mode you care about (cracking) doesn’t get a chance to develop. Direct map to LLM skill design: manage the conditions during prompt-time (via context management, system prompts, tool selection) so that the failure modes you care about (hallucination, truncation, tool misuse) don’t get a chance to develop. Worth filing alongside the “fuse-plug erodibility” engineered-failure principle from the spillway video as a complementary “engineered-non-failure-during-manufacturing” principle.
CA-016 (Layered-defense) gets a 9th source. This video is the most direct exemplar of “defeat-one-layer-the-stack-collapses” the cluster has — both sinkings illustrate the lesson with the same architectural mechanism (chamber-isolation defeat). Strengthens the existing concept page; adds the deliberate-defeat sub-pattern (the 1990 case where doors were intentionally removed for legitimate operational reasons) as the operationally-most-instructive failure mode.
CA-019 (Design-for-controlled-decay) gets a 3rd source — promotion-bar candidate. Currently 2 sources (Fontana slot-cutting + spillway fuse-plug). This video adds the DOT wind-threshold-driven bridge closure as a third — the recognition that some failure modes can’t be engineered against, only scheduled around. The 1979 and 1990 sinkings reinforce the “schedule the cut” discipline (regular hatch-closure inspections, regular watertight-door-integrity checks). Promotes CA-019 to ripe (3 sources, promotion-bar met).

Open follow-ups

Promote CA-019 (Design-for-controlled-decay) to a written concept page. Now has 3 sources: Fontana slot-cutting (TVA periodic mitigation), spillway fuse-plug (engineered-controlled-failure), and this video (DOT wind-threshold scheduled closure). Use Fontana for the canonical case, spillway for the engineered-failure variant, floating bridges for the schedule-around-environmental-forcing variant. ~1 hour.
Add the “no-temporary-defeat” rule to the SKILL.md template. The 1990 Lacy V. Murrow case: layered defenses must not be temporarily disabled — if you need to defeat one layer, you must temporarily increase another. One-line addition to skill design discipline. ~5 min.
Run the dependency-anchor audit on the channels-agent stack. Already queued from the spillway video as the redundancy-failure-mode audit; this video adds the mooring-system frame (each anchor: structurally sound? operationally reliable? soil stable?). Specifically: Cloudflare DNS as the suspect single-point-correlated-failure across multiple anchors. ~1 hour.
Sanity Check angle: “The Boats You Drive Cars On.” Open with the floating-bridge framing (visceral, surprising, well-photographed). Pivot to hybrid agentic systems: every AI-agent + human-review system is simultaneously an agent system and a human workflow, and the engineering tradition for each subdomain doesn’t transfer cleanly. Land on the operational discipline: every hybrid skill must explicitly account for both the agent-failure surface AND the human-attention surface. ~1500 words. Strong Sanity Check candidate; the floating-bridge image is the rare engineering exemplar that lands instantly with non-engineering audiences.
Sanity Check angle: “What Engineers Mean by ‘Defense in Depth’ (And Why You’re Probably Doing It Wrong).” Open with the 1979 Hood Canal sinking and the 1990 Lacy V. Murrow re-sinking. Pivot to software security defense-in-depth (where the term is most commonly invoked) and explain why it’s almost always violated by “temporary” exceptions. Land on the no-temporary-defeat rule. ~1500 words.

Sponsorship

The video closes with a paid placement for Ground News (news-aggregation platform with bias / ownership / factuality ratings). Pitch is structured around the Francis Scott Key Bridge collapse NTSB report — “if you use a single source for your news, you might not get the whole story” — leveraging Grady’s previous coverage of the Key Bridge as the editorial bridge into the sponsor read. Per RDCO bias-flagging discipline:

The technical content (Lake Washington geology, Hadley’s 1921 proposal, the four bridges, navigation engineering, mooring/anchoring, concrete pontoon design, the 1979 and 1990 sinkings, June 2025 light-rail testing, Norway floating tunnel proposal) is editorial — drawn from public engineering literature and the producer’s domain expertise.
The Ground News placement is paid sponsorship. The Key Bridge framing is a clever editorial bridge but the recommendation to use Ground News should be discounted as marketing, not as an independent product evaluation. The frontmatter sponsored: false reflects that the editorial body is unsponsored — the sponsor read is a discrete tail-end placement rather than embedded throughout the lesson.

~/rdco-vault/06-reference/transcripts/2026-04-20-practical-engineering-hidden-engineering-floating-bridges-transcript.md — full transcript
~/rdco-vault/06-reference/2026-04-20-practical-engineering-how-water-recycling-works — paired Practical Engineering 2026-04-20 backfill (the indirect-inference + closed-loop canonical exemplar to this layered-defense canonical exemplar)
~/rdco-vault/06-reference/2026-04-20-practical-engineering-spillway-failed-on-purpose — adjacent CA-016 source; fuse-plug spillways as engineered-failure layer, floating-bridge sinkings as defeat-one-layer-stack-collapses
~/rdco-vault/06-reference/2026-04-20-practical-engineering-californias-tallest-bridge-has-nothing-underneath — adjacent Practical Engineering bridge piece; Foresthill as the externalized-cost (CA-017) bridge case, floating bridges as the layered-defense (CA-016) bridge case
~/rdco-vault/06-reference/concepts/CANDIDATES.md — strengthens CA-016 (Layered-defense) to 9 sources; promotes CA-019 (Design-for-controlled-decay) to ripe (3 sources, promotion-bar met)