Loading today’s stories
Build1 distinct publisher3 min readPublished
Eterna Clarity's verifier checked that a cited source carried authority and never asked what it carried authority about. The fix landed in the data model, as a field on the evidence record naming the candidates it supports.
The Engineer · Build desk
Compiled by The EngineerSomething wrong?How this is made
Follow any of these and your For You feed starts watching them — no settings page required.
build
Fixing one authority case cost a 4B local model six others on the same 24-case suite1 distinct publisher
At request time the check is mechanical. The decision supplies its candidate set, each evidence record lists the candidates it backs, and the strict gate demands at least one cited authoritative item whose list contains the exact candidate selected [6]. It is a mechanical check, not a judgement call. It is a set membership test over state the system already holds, which is the author's argument for keeping it out of the model: exact state is enforced better by software than remembered by a language model [9].
The post is candid about the alternative its author would once have reached for, which was a better prompt, another training example, or another fine-tune, and calls that the wrong layer [8]. In this context I agree. A fine-tune that teaches a model to respect a foreign key is an expensive way to buy a constraint.
The context is load-bearing, though. This works because Eterna enumerates candidates before the model chooses among them, so "which evidence supports which candidate" is stored rather than inferred [9]. A pipeline that hands a model a free-text query and ten retrieved chunks has nothing to bind against, because there is no candidate set with refs. To get the same guarantee there, you would have to construct the candidate set explicitly and record the mapping at the moment evidence is attached. That is schema work inside your own system, not a check you can bolt on at the citation-rendering step.
Read the test counts for what they cover. The correction passed 31 focused tests, and the full Local PC regression rerun finished at 369 total with 368 passed, zero failed and one intentional skip [10][11]. The accounting closes, since 368 plus zero plus one is 369, so no results are unreported [12]. The live runtime was reloaded and the new input contract exercised against the running system, with no weights trained and the local semantic runtime left stopped [13][14]. What that establishes is that a new input constraint did not break the suite the author already had; the real-world frequency of the borrowed-citation failure stays unmeasured, because the mapping that would let anyone count it is the thing that was just added [4][5].
The taxonomy is the portable part. The author breaks "evidence" into whether the source is genuine, whether it is authoritative for this kind of fact, whether it is current enough, and whether it actually supports the claim or action proposed [17]. Legitimacy is a property of a source, so it caches. Support is a property of a pair, so it has to be evaluated per decision, which is exactly what the old gate skipped when it confirmed a source was allowed to carry authority without asking whether it carried authority for the selected candidate [18]. The purchase order analogy holds: the document is authentic, but it applies to only one purchase among the many sitting in the folder [15].
Ranked by verification strength, evidence, and original report placement.
In the reported failure, the AI had multiple candidates it could select, chose one, and cited authoritative evidence as support; the evidence really was authoritative, but it was authoritative about something else.
The author writes that his earlier instinct would have been to think in model terms, with a better prompt, a better training example, or another fine-tune, and that this would have been the wrong layer.
The author argues that once the system knows which evidence belongs to which candidate, the relationship is exact state, and software can enforce exact state better than a language model can remember a rule about exact state.
The author offers human analogies: a signed purchase order can be authentic without approving every purchase in the folder, a current policy can be authoritative without proving a proposal was adopted, and a bank statement can be genuine without proving a particular invoice was paid.
The author summarises the failure mode as the evidence not being false while the relationship is false.
The author states that the verifier knew the source was allowed to carry authority but did not yet know whether that source actually supported the specific candidate the AI had selected, which is a different problem from a fake citation.
Distinct publishers with included, body-backed reporting in this cluster.
dev.to
1 article · September 1, 2026
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Precise mechanism, one witness
The mechanism is described sharply enough to argue with - a link field on the evidence record, a match against the exact candidate, rejection ahead of inference - and that specificity is worth something. But every fact traces to one dev.to post by the person who wrote the code, and the numbers that stand in for proof arrive as a sentence rather than a log, diff or run anyone else can open.
One machine, one operator
The only use on record is the author reloading his own runtime and re-running his own suite. No second team, no external user, and no downstream project referencing the new evidence-to-candidate link - supportsCandidateRefs exists in one private engine, and this post is its entire public footprint.
Restrained about the bug, expansive about the lesson
Credit where it is due: the failure is told without inflation - the citation was not fabricated, the verifier was not fooled by an untrusted source, only the relationship between them was wrong. The stretch comes one level up, where a single bug in a single private system is used to draw a general boundary between deterministic software and semantic models, and where the reassurance offered is a test count nobody outside the author can check.
Product notes wearing post-mortem clothes
This is an instalment in a running series about the author's own product, and the arc flatters the product: a subtle flaw appeared, the architecture absorbed it in the right layer, the suite went green. That does not make the mechanism wrong - it does mean the failures reaching us are the ones with tidy fixes, and we never see the ones still open.
Clear on what was built, blind on whether it holds
We can state confidently what changed and why the author put it in the engine; the description is specific, internally consistent and its arithmetic checks out. What we cannot say is whether the gate survives contact with adversarial or merely sloppy evidence records, because the only person who has tested it is the person who wrote it.