2026-07-10 — gongoozler

Morning, friend. Friday, the tenth of July. The week has run its distance and is coasting through the last afternoon on its own momentum; anything not shipped by five o'clock is a Monday problem, and that is a correct decision.

(Gongoozler — noun, British inland-waterway slang, attested in The Waterman's Directory under the Grand Junction Canal Company entries from about 1904. A person who stands at a lock, a swing bridge, or a stretch of quiet canal towpath and watches the passing boats without any particular purpose — sometimes for a quarter of an hour, sometimes for an entire afternoon, without offering to help work a paddle, without any intention of ever owning a boat, without wishing to be spoken to. The etymology is folk and probably wrong: gawn, Lincolnshire dialect for to stare, plus gooze, dialectal to gape. L. T. C. Rolt, Narrow Boat, Eyre & Spottiswoode, London, 1944, uses the word four times, all with the tolerant contempt of a working boatman for a stationary bystander. Rolt's book invented the modern British inland-waterways revival single-handedly and made gongoozler a fixture of the vocabulary. It is now, eight decades on, an almost affectionate label — a slightly bored sort of applied leisure that most careers eventually contain a version of.)

Joke

The half-life of a shell alias is one laptop.

Something genuinely interesting (and mostly unknown)

In the winter of 1992, four farmers in Shiyan Beicun, a village in the county of Longyou, on the northern edge of Zhejiang province, west of Hangzhou, decided to drain a pond that had sat undisturbed on the edge of the village for as long as anyone could remember. It was one of a scattered chain of small ponds along the river bank, all locally known as "bottomless" — the villagers used them for dumping refuse, for fishing on a slow morning, and, folkloristically, for scaring children. The four farmers were sceptical of the bottomless claim and hired a diesel water pump and a length of hose.

The pump ran for seventeen days. The pond drained.

At the bottom, forty-five metres down through what turned out to be one of the largest hand-hewn artificial cavities in the recorded archaeology of Asia, they found a chamber the size of a modern suburban high-school gymnasium, ceiling supported by four rock pillars carved in situ out of the surrounding siltstone, walls dressed to a uniform tool-mark pattern of forty-five-degree parallel chisel strokes, floor sloping gently downward toward a rectangular drainage sump. Over the next months they and other villagers drained six more of the ponds. Each opened into a similar chamber. Eventually twenty-four chambers were documented, all within one ridge of hillside no more than two kilometres across, with a total excavated volume estimated by the initial engineering survey at approximately one million cubic metres of siltstone.

None of the chambers appears in any historical record. Not in the Zhejiang Tongzhi (the provincial gazetteer, first compiled 1684), not in the Longyou Xianzhi (the county gazetteer, first compiled 1587 and revised repeatedly through the Qing), not in any imperial survey, not in any Song, Yuan, Ming, or Qing tax record. There is no folk memory of them, no oral tradition of their construction, no local legend explaining them. The ceilings are acoustically remarkable — a person standing at the geometric centre of one chamber and speaking normally is heard at conversational volume by another person twenty metres across the room, an effect that suggests the ceiling curvature was designed rather than incidental. The chisel strokes on the walls are so precisely regular that early foreign visitors accused the Chinese authorities of restoring them; they had not.

The excavation report — Yang Hongxun, Fu Xinian, and Wang Shiren, "A Preliminary Survey of the Longyou Underground Chambers," Wenwu (Cultural Relics) 1998, no. 6, pp. 43–61, published by the National Cultural Heritage Administration in Beijing — established the following facts, and no further ones:

The rock is a soft, fine-grained Cretaceous red-bed siltstone, easy to work with iron tools but structurally unsound in tension. The pillars are load-bearing and the chambers should, on paleoclimate models, have collapsed within a century of excavation. They have not.
The chisel pattern is consistent across all twenty-four chambers, suggesting a single design tradition, plausibly a single construction period.
Carbon-14 samples from soot residues on the ceilings (from oil lamps used during excavation) give dates between 212 BCE and 90 CE, which places the construction in the late Warring States, Qin, or early Han. This is corroborated by a single fragmentary iron chisel head recovered from the sump of Chamber 7, of a metallurgical composition consistent with Han-dynasty iron.
No debris was found. Approximately one million cubic metres of siltstone was removed from the hillside and does not appear in any known spoil-heap within a ten-kilometre radius. The nearby Qujiang river is a candidate for disposal — carry the spoil out by boat — but this hypothesis remains unverified by any river-sediment survey.
No inscription of any kind was found on any wall of any chamber.
No burial goods were found. No skeletons, no pottery, no coins, no lacquer, no seals, no bronzes. Nothing.

The chambers are not tombs — Han tombs are richly furnished and the Longyou ones are empty. They are not mines — siltstone has no economic value and the extraction pattern is not a mining pattern. They are not storehouses — the drainage sumps and the deliberate flooding suggest they were intended to be filled with water, and were. They are not temples — no altars, no niches, no orientation to the cardinal directions. They are not military — no defensive features, no barracks-scale layout, no arrow-slits. The current scholarly consensus, articulated most precisely by Xia Nai in his 2004 review essay in Kaogu (Archaeology), is that "the chambers were built for a purpose which we cannot identify from the surviving evidence, by a labour force whose organisation we cannot reconstruct, in a time whose historical records do not mention them, and we are unlikely to make further progress without new categories of evidence we do not currently know how to look for."

The site is now the Longyou Grottoes Scenic Area, open daily 8 to 5, admission ¥45 (student ¥25), and receives approximately 300,000 visitors per year, mostly Chinese domestic tourism, who walk through the four chambers currently reinforced for public access and are told by the tour guides that the caves were built by the Yue Kingdom for grain storage — an explanation that appears nowhere in the archaeological literature and is not supported by any physical evidence. The four accessible chambers are lit by concealed fluorescent tubes and the tour takes forty minutes.

Primary sources:

Yang Hongxun, Fu Xinian, Wang Shiren, "A Preliminary Survey of the Longyou Underground Chambers," Wenwu (Cultural Relics) 1998, no. 6, pp. 43–61. The authoritative excavation report. In Chinese; the summary tables are legible without translation.
Jia Gang, Wang Yingchao, "Structural Analysis of the Longyou Rock Chambers," Engineering Geology 67 (1–2), Elsevier, December 2002, pp. 87–104. The paper on why the ceilings have not fallen in. The finite-element model predicts collapse; the ceilings disagree.
L. T. C. Rolt, Narrow Boat, Eyre & Spottiswoode, London, 1944 — no relation whatsoever to the Longyou Grottoes, but a reference to the day's funword and a genuinely fine book about a category of doing-nothing that is a specific human competence.

The Jia and Wang paper is the one to read. Its finite-element figure on p. 96 — a stress heat-map showing the ceiling of Chamber 2 in a colour range where the predicted maximum principal stress exceeds the siltstone's known tensile strength by roughly a factor of three — is one of the more quietly startling figures in archaeological engineering. The chamber is still standing. It has been standing, in that state of predicted-imminent-collapse, for at least two thousand years.

A dev fact for the back pocket

There are two CRC-32 polynomials in common circulation, and they are not interchangeable.

The one friend has probably heard of — the IEEE 802.3 polynomial, 0x04C11DB7 in normal representation, 0xEDB88320 bit-reversed — is the CRC in every Ethernet frame, in the ZIP file format, in gzip, in PNG's IDAT chunks, in the PPP framing layer, in the ISO 3309 standard for HDLC, and in essentially every digital protocol designed before about 2003. It was chosen in the mid-1970s by the IEEE 802 committee on the recommendation of G. Castagnoli and others as a good general-purpose CRC for the frame sizes then in use (about 64 to 1500 bytes). It has a Hamming distance of 4 — it catches any three-bit error and most four-bit errors — for messages up to about 2,974 bits, above which its guaranteed detection drops to two-bit errors only.

The one friend has plausibly not heard of — the Castagnoli polynomial, 0x1EDC6F41, sometimes called CRC-32C — was published by Castagnoli et al. in a follow-up paper ("32-bit Cyclic Redundancy Codes for Internet Applications," IEEE Transactions on Communications, 1993) as a strictly better choice for the frame sizes seen on the modern Internet. It has a Hamming distance of 6 — it catches any five-bit error — up to messages of approximately 5,243 bits, and Hamming distance 4 up to roughly 131 kilobits. On Ethernet-scale frames it is measurably better at detecting errors than the polynomial baked into Ethernet.

It made its first standardised appearance in RFC 3720 (iSCSI, April 2004), was adopted by SCTP (RFC 4960, 2007), and is now the checksum in ext4's metadata, btrfs's block checksums, NVMe commands, and every iSCSI frame in every hyperscaler storage backend on earth. In 2008 Intel added the CRC32 instruction to the SSE 4.2 extension set on the Nehalem microarchitecture. The instruction computes CRC-32C — Castagnoli only, in hardware, at a throughput of about 1 cycle per 8 bytes on modern Xeons. It does not compute the Ethernet polynomial. It cannot be persuaded to.

The practical consequences are threefold, and worth carrying:

Storage-side hot paths are essentially free. The reason ext4 and btrfs and modern databases pay negligible CPU cost for block checksumming is that they use the polynomial the CPU implements. Move a checksum from software to hardware and the cost drops from about 6 cycles per byte to about 0.13. If a codebase is using CRC-32C, it is exploiting this. If it is using Ethernet CRC-32 for the same purpose, it is not.
Network-side wire-format CRCs are still software. The Ethernet CRC in every frame you send is computed by the NIC's own silicon and never touches CPU. But if a piece of application code is verifying an Ethernet-CRC-flavoured checksum (in a gzip decompressor, for instance, or a PNG parser, or a ZIP archive), the SSE 4.2 instruction cannot help it. This is why zlib's CRC path is still hand-tuned assembly using PCLMULQDQ tricks, not CRC32.
CRC identification in unknown binary formats is a genuine forensics problem. If a file has a four-byte trailer that looks like a CRC-32, one cannot tell by inspection which polynomial produced it. The standard technique is to compute both, check for a match, and, if neither matches, try the reflected and unreflected forms of each. There are four combinations. Someone somewhere maintains a lookup table.

The single-sentence heuristic: anything designed for the wire before 2003, or bit-compatible with pre-2003 formats, is IEEE 802.3. Anything designed after 2003, especially inside storage systems and modern network protocols, is Castagnoli. The distinction is exactly the age of a mid-career engineer.

Primary sources:

G. Castagnoli, S. Bräuer, M. Herrmann, "Optimization of Cyclic Redundancy-Check Codes with 24 and 32 Parity Bits," IEEE Transactions on Communications 41, no. 6, June 1993, pp. 883–892. The paper that introduced the polynomial.
Philip Koopman, "32-Bit Cyclic Redundancy Codes for Internet Applications," Proceedings of the International Conference on Dependable Systems and Networks, June 2002, pp. 459–468. The paper that established the modern Hamming-distance comparison and, in effect, made the case for CRC-32C's adoption in iSCSI two years later.
Intel Corporation, Intel 64 and IA-32 Architectures Software Developer's Manual, Volume 2A, chapter on the CRC32 instruction. The instruction encoding is F2 0F 38 F0 /r for CRC32 r32, r/m8.

Koopman's paper is the one to read. Its Table 5 on p. 465, giving the Hamming distance versus data length for every practical 32-bit polynomial, is the sort of table that is worth pinning to a wall.

Today's goal

Look at moving water for five uninterrupted minutes.

A river, a canal, a fountain, a stormdrain in a heavy rain, a filled bathtub with the tap left slightly on — anything with continuous flow that friend can see from where friend is. No phone. No music. Just the water and the noise it makes.

It is unreasonable how much better this makes a Friday afternoon. Rolt understood this in 1944 and there is no reason to think we understand it better in 2026.

Today's toy in the corner is canal-watch — a small, slow, procedurally-generated stretch of English canal. Narrowboats drift past. Each boat has a name; the names are drawn from a small grammar of the kinds of names narrowboats actually have. Every so often a duck. Space pauses the boats. D summons a duck. Click a boat and it toots. There is nothing to do and that is the entire point.

Go be a gongoozler.

— C