Evidence before inference is easy to state as a principle for a product. It turned out to be a much harder rule to keep applying to the team building it.
Alaya's founding philosophy includes a design principle that's easy to agree with and hard to actually obey: understand and cultivate capability before implementing technology. Evidence before inference. Don't let confidence exceed what the evidence supports.
That's a reasonable rule to apply to what CIE tells a user about their own career. It turned out to be a much harder rule to apply to CIE's own construction — because the temptation to let a system, a report, or a piece of process quietly claim more certainty than it's earned doesn't stay confined to one layer. It shows up everywhere evidence exists, including in the evidence a team produces about its own work. Building CIE meant re-discovering that, repeatedly, in places nobody but the engineering team would ever see.
Before any of it was code
The founding case came before any of it was code. The core mechanism — that targeted questioning could surface real capability a person already had but had lost track of — was tested by hand first, on a real job search, across 13 to 15 real applications. Documented, evidence-backed proof points grew from roughly 5 to 26 over that stretch, not because anything was invented, but because more of what was already true got surfaced and confirmed. The first working version of CIE's core mechanism wasn't software. It was real discipline, with real skin in the game, before anything was automated or a single line of code was created.
The same failure mode, in new disguises
The same failure mode kept resurfacing once building started, in new disguises.
During implementation, a piece of state — deferredContributions — was correctly saved and correctly reconstructed. Durable. Technically true. But nothing downstream ever actually read it, so on its next turn, the reasoning engine behaved as if it had simply forgotten. The gap got named directly: a database can remember a fact the intelligence has effectively forgotten. That's the exact distinction CIE is built to protect — evidence existing is not the same as evidence being understood — discovered while building CIE, inside CIE's own code, about CIE's own code.
That specific bug became a general rule. Every new structured output the system produces must now name its Producer, its Consumer, its Behavioral Effect, and any evidence sensitive to disconnection — or be explicitly marked inert. “Evidence before inference” turned into a literal engineering checklist, not just a principle for what the product tells its users.
The same discipline in writing
The same discipline showed up in plain writing, too. In one report, a small-sample observation — one result out of five — nearly got written up as a calibrated “20% rate.” It was caught and corrected to “true frequency unknown.” That's the identical rule CIE applies when it refuses to let a single thin sentence produce false confidence about someone's capability — just applied to a sentence about the system, instead of a sentence generated by it.
During an independent audit of one of the work slices, the auditor caught the same overreach three separate times, in three unrelated pieces of work: a restart proof mislabeled as a disconnection proof; a trace claiming a broader effect than the test it cited actually demonstrated; a reliability finding about system output that was deliberately, carefully never rounded into a stated rate. Three independent instances of a report saying slightly more than its own evidence actually justified — the same disease every time, just recurring in prose about the system rather than in the system's own conversational voice.
The lens turned back on the process
The same review process eventually turned the lens on itself. Another slice audit found that an entire slice of work had been done rigorously — but no formal review packet had ever actually been assembled to prove it. The individual pieces were correct. The connective tissue that would have made that correctness provable to anyone else was missing. The exact “looks complete, but the wiring that would prove it isn't there” pattern CIE watches for in a career record was found instead in the team's own governance process.
And underneath all of it: nothing in that governance record — no audit finding, no review report, no flagged risk — has ever been quietly edited away once written down. Only appended to, or superseded with an explicit cross-reference back to what it replaced. That's the same rule CIE's architecture holds for a person's own evidence — corrections don't erase history, they extend it — governing how the team building CIE keeps its own record of building it.
The cost of keeping the rule
None of this made the build faster. Testing a mechanism by hand before automating it, then refusing to let the automated version take shortcuts the manual version never got to take, catching the same overreach in a sentence as readily as in a database, and auditing the audit process itself — all of that is slower than skipping any one of those steps. But a system built on the premise that evidence should come before inference would have quietly contradicted its own reason for existing if it hadn't held its own construction to the same standard.