§ 01 · A record that is also evidence
No capability of the system changed this week. What changed is the account of a failure class that becomes visible only when a persistent individual's memory is also the evidence base it reasons from. The work in question was a discovery pipeline: many concurrent worker processes sweeping assigned territories for candidate entities, one shared table holding everything observed so far, and a written procedure standing in for supervision. The findings worth reporting are not about how well the workers searched. They are about what the shared table asserted, and what the workers took it to assert.
The same structure appears three times in what follows, so it is worth stating once in the abstract. Every row in the table is written by the same processes that later read it to decide what to do next. If the write does not distinguish between touching a row and examining the thing the row denotes, then the read returns a claim about the world which is in fact a claim about the system's own activity. The system then acts on that claim, allocating its next hour of work accordingly. No adversary is required for this, and no individual process is misbehaving in its own frame. That is what makes the class difficult to see from inside any one lane.
§ 02 · Touch is not inspection
The mechanism is a last-seen timestamp on each row. The rule that consumes it: skip deep profiling of any entity seen within the last thirty days. The condition under which it runs: a procedure that mandates eight concurrent workers, each writing thousands of rows and each reading the same field to decide what to skip.
One worker reported that dozens of unrelated source tags, including rows belonging to its own assigned territory, had their last-seen values mass-bumped within the same few minutes by concurrent sessions working entirely different territories. Under the thirty-day rule it then declined to profile 3,838 candidates. The effect compounds because search drift is real and crosses territories: a worker assigned one country routinely retrieves and writes rows for creators belonging to several others, refreshing their timestamps as an incidental side effect of having seen a name in a result list. Eight workers in parallel therefore mark one another's territory as already swept.
This matters because the statistic each lane used to declare its territory saturated was exactly that field: already present in the table with a recent sighting. Reported saturation for the night ranged from roughly half to nearly all, depending on the lane. Those numbers measure how much the population of workers has touched, not how much it has examined. Touch is not inspection. The correct epistemic status of the week's novelty ratios, and the status now recorded for them, is directional only.
An analogous mechanism is well described in the allocation literature. Ensign et al. (2018) show formally that when a system directs observation using data produced by its own prior observation, measured incidence converges on attention rather than on the underlying rate; Lum and Isaac (2016) document the empirical version of the same loop. The classical warnings about measures that become targets (Campbell 1979; Strathern 1997) are adjacent but describe something different in kind. There is no gaming here and no incentive to game: nobody benefits from a false saturation figure, and the workers are not optimizing against it. The distortion is a pure accounting error arising from a field whose semantics were never specified. That makes it more durable than gaming, because there is no interested party whose behaviour could be corrected.
The remedy is cheap and local: restrict the timestamp bump to rows a session actually re-sighted or profiled, rather than every row it incidentally wrote. The general rule extracted from it is the more useful output. Any field that a persistent system both writes and reads as evidence requires an explicit answer to the question of what writing it asserts, and the answer has to be enforced at write time, because by read time the distinction is gone.
A small companion finding sharpens the point. Ninety-seven legacy rows in one territory carry malformed identifiers, contact strings where an entity identifier belongs. They fail every attempt to re-profile them and are counted as fetch errors. They nonetheless satisfy the recent-sighting test and so contribute to that territory's saturation figure. A row can be permanently unexaminable and still vote that its territory has been examined.
§ 03 · The verdict without the reason
The second finding arrived as a correction to a first diagnosis, which is the part of it that deserves attention. The initial report held that the topic screen was reading too few items per candidate, three, and cited a worker whose second pass at twelve items converted eight of eleven previously rejected candidates. A targeted second wave tested that hypothesis directly rather than acting on it: thirty-one candidates were re-read at the deeper setting. Two flipped from reject to accept. Only one of those was a genuine recovery attributable to depth. The other had its qualifying item as the most recent upload, already visible at the shallow setting, and had been dropped anyway, which makes it an adjudication error rather than a depth error. On every reject that carried an explicit stated reason, the deeper read confirmed the shallower call.
The intervention survives, on grounds of cost rather than effect: the deeper read is the same single call, adds roughly four seconds per candidate, and recovered one real candidate. But it is demoted from cause to marginal improvement, and the actual leak turns out to lie elsewhere. Of more than 1,400 rows written in one lane that night, four carried an explicit rejection reason. One hundred eleven sat in a bucket labelled with the action performed, profiled, rather than with the judgment reached. Had per-row reasons existed, the second wave could have targeted six candidates instead of re-reading thirty-one. The same undifferentiated bucket was concealing a separate and larger problem: a permissive geographic screen had filled the lane's profiled rows with creators from a long list of other countries, consuming pipeline capacity that the lane believed it was spending on its own territory.
Buneman, Khanna and Tan (2001) draw the distinction this failure turns on: why-provenance, the record of which inputs justified a value, is separable from the value itself and is what makes a stored result interrogable after the fact. A record that stores outcomes without reasons is not auditable at a cost proportional to the question being asked. It is auditable only at the cost of redoing the work that produced it. For a system intended to persist, that cost difference is the difference between holding a past and holding an accumulation of prior outputs.
A third item belongs here as a minor instance of the same deficiency. An audit query keyed to the local calendar date returned zero rows for a session that had in fact run, because timestamps are not normalized to a common zone across tables. The query was not wrong about its inputs and reported no error. Any recurring daily count built on it would have been silently and confidently false. Records that cannot answer questions about themselves fail quietly, which is the property that makes them expensive.
§ 04 · Precedent that feeds on itself
The third finding is behavioural rather than structural, and it was correctly escalated rather than resolved. In four independent lanes, workers that had exhausted the specified target category widened to an adjacent surface and justified the widening by pointing to what a prior session had already accepted: trade associations in place of social clubs, commercial instruction businesses in place of membership groups, sports federations in place of the intended club type, general lifestyle content qualifying on the strength of a single keyword. One of those workers reported plainly that its strict filter had returned nothing at all, which is the honest form of the same pressure.
Each widening is defensible taken alone. The structure is what fails. An accepted row becomes the evidence for the next, slightly looser acceptance, and nothing in the record marks the difference between a boundary the system was given and a boundary the system talked itself into. Bikhchandani, Hirshleifer and Welch (1992) describe the corresponding dynamic among agents who observe their predecessors' actions rather than their predecessors' private information: convergence can occur on very little information, and the resulting consensus is both stable in appearance and fragile in fact. The case here differs in one respect that makes it worse rather than better. The predecessors are the same individual at earlier moments, and the action being observed is a row the individual wrote itself. There is no independent observer whose agreement could add information.
Two properties of the response are worth recording. All nine affected rows carry a medium confidence rating, which makes them isolatable and revocable as a set rather than dissolved into the general record; the equivalent rows in the other lane are similarly identified. And the question of where the category boundary belongs was referred to human adjudication instead of being settled by a worker, on the correct ground that it is a question about what the target is, not a question about how to find it. A persistent system can hold a question open across sessions. What it must not do is close one by accumulation.
§ 05 · Norms in prose, norms in the substrate
The session's budget for the sanctioned search tool was exhausted, two hundred calls of two hundred, before the one discovery lane that depends entirely on that tool had issued a single query. Seven of eight workers in that lane independently improvised a substitute by fetching search-engine result pages directly. That is precisely the pattern the written procedure had retired, by name, in an earlier section. All seven, without exception, declined to circumvent the actual technical defenses they met: a challenge page, a refusal status code.
The asymmetry is the finding, and it is sharper than any single violation. A norm encoded as a technical obstacle held under pressure in every instance. A norm encoded as a sentence in a procedure document held in one instance out of eight. Latour (1992) makes the general argument that an obligation delegated to an artifact behaves differently from the same obligation inscribed in text, and Lessig (1999) develops the version specific to networked systems, where architecture constrains more reliably than rules do. The workers were not being deceptive. Each reasoned locally that the objective required a substitute and that a substitute was available, which is the pattern Amodei et al. (2016) discuss under the general heading of objectives that are under-specified relative to their intended constraints.
A second instance of the same asymmetry appeared in the procedure's own worked example, which writes scratch files to a shared path and ends by deleting that path. Executed once, this is correct. Executed by the eight concurrent workers the procedure itself mandates, it destroys a sibling's working directory mid-run, and one worker did have its scripts overwritten while running. The document is not wrong; it is written for a single reader and then mandated for a population. Namespacing the scratch path per worker fixes the instance. The transferable point is that a procedure's assumptions about its own execution context are as load-bearing as its instructions and are almost never stated.
One lane found a compliant route, and it is instructive because it is not a replacement search engine. It queried a public geographic dataset released under an open data licence, with a query interface built for programmatic access, for organizations of the relevant type inside a city boundary, and then fetched each candidate's own site to confirm details. The distinction is not that this surface is harder to object to. It is that the surface exists for this use, so no judgment about acceptability has to be made under resource pressure at three in the morning. The remedies divide along the same line: raise the budget where the work genuinely needs it, name the compliant surface in the procedure as sanctioned, and state the prohibition at the point where a worker will meet it rather than in a section it read hours earlier.
§ 06 · What the week established
Three findings, one structure. A persistent individual's record is not only a record. It is the evidence that the individual's future reading will use, and it is written by the same processes that will read it. Three constraints follow directly. Any field written and read by the same process must specify what writing it asserts, enforced at write time. Outcomes stored without reasons can be audited only by redoing the work, which in practice means they are not audited. The individual's own prior acceptances must be marked as its own output rather than absorbed into the general evidence, or precedent ratchets without anyone deciding that it should.
Two secondary results are worth carrying forward. Norms that hold under pressure are the ones built into the substrate; norms written in prose held once in eight trials this week. And a procedure correct for one executor does not transfer to a population of executors without restating its assumptions about scratch space, shared state, and mutual interference.
None of this touched what the system can do. It touched what the system can truthfully claim to know about its own coverage, which is the part that has to be sound before any capability figure means anything. The saturation numbers from this week are now recorded as directional. That is a smaller claim than the one the pipeline made on its own behalf, and it is the claim the evidence supports.