What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
A language model’s earlier answer should not become evidence about what a user wants or what is true. In a 30 September 2026 interview with The AI Journal, founding AI engineer Baber Javed described how college recommendations kept steering a student toward computer science after the student had shifted to cognitive science. His team’s response, as he tells it, was to track where profile information came from and give the student’s own statements priority over model-generated content.
How a recommendation loop mistook the model’s output for a student’s preference
Javed recounts a student who initially explored computer science and later wanted to focus on cognitive science. Recommendations continued to include computer science because earlier conversations and recommendations remained in context. The system, in his words, was treating its own previous output as evidence about the user’s interest.
The problem was not simply that the model failed to understand the student’s newer messages. Rather, model-generated references to computer science were carried forward alongside the student’s statements, making an earlier recommendation look like supporting evidence. Javed says the team addressed this by recording the provenance of profile information and prioritizing direct student statements over generated output. This is his account of a particular incident and fix, not an independently audited case study.
Where the model stops in Javed’s college-matching system
Javed describes the model as an interpreter and assistant, not the authority on college facts. It turns a conversation into a typed Search Spec; structured lookups and database filters then apply requirements, and the model may help rank options from the filtered results.
#1 Best Overall
From conversation to a structured search
The Search Spec can contain subjects, hard requirements such as location, and softer preferences such as budget. Subject names are mapped to federal program codes using a lookup table. Hard requirements become database filters, while softer criteria guide sorting. The resulting shortlist is based on structured program data rather than facts generated from conversation alone.
What the model may still do
After the database has narrowed the options, the model can help select or rank programs. That leaves room for language-based assistance while keeping requirements and program facts outside the model’s authority. Javed says names, admission rates, and tuition fees need to come from structured data.
Why retrieval alone did not make recommendations reliable
Javed says an earlier approach represented college programs and student interests as vectors and compared them by similarity. The matches could sound plausible but still be unsuitable—for example, returning History programs to a student interested only in archaeology. In his account, retrieval and edge-case handling mattered more than switching to a supposedly smarter model.
- Filter during retrieval. Apply relevant constraints before producing a candidate list instead of relying on semantic similarity alone.
- Simplify ranking and diversity logic. Javed reports reducing complexity in how results were ranked and diversified.
- Use relevance thresholds and fallbacks. Set a minimum relevance bar, and broaden searches when restrictive filters return too few options rather than presenting weak matches as good ones.
- Ground extracted preferences in the student’s words. Require a preference to be supported by what the student said before normalizing it and looking it up.
These are pipeline changes Javed describes, not a universal recipe. The underlying point is to separate retrieval from inference: a plausible semantic match is not proof that an option meets the student’s stated needs.
Keep source-of-truth facts and calculations deterministic
Javed draws a distinction between tasks with verifiable answers and those involving subjective fit. For payroll, he says calculations must follow deterministic rules; a model can help shortlist applicable rules or explain the reasoning, but it should not supply the authoritative calculation. A payroll error can directly underpay someone.
College recommendations also depend on reliable facts, such as fees and program details, but deciding whether a program is a good fit is more subjective. Javed says that makes recommendation errors harder to detect than a wrong payroll result. The practical boundary is to keep factual inputs and enforceable constraints in structured systems, while using the model for interpretation or assistance that can be checked.
Evaluate conversation quality separately from matching
Javed says his team treats conversation quality as subjective and matching against known data as testable. For conversation quality, he describes an LLM judge scoring repetition, closure, and hallucinations on a 1-to-10 scale, with human-labeled conversations used to evaluate the judge. These are his team’s reported practices, not an industry standard or independently verified benchmark.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →In the interview, Javed reports comparing at least 30 conversations against human labels, with at least five conversations flagged by humans as hallucinations in the evaluation set. His team’s stated hallucination-detection criteria were zero false negatives and no more than two false positives in the judge pass. The interview does not give the underlying dataset or the score bounds for repetition and closure, so those details cannot be inferred from the figures.
Best Value
He also says the labeled examples are run as regression tests after prompt changes or model switches. If a critical metric regresses, the team investigates and does not ship until it understands why. A useful evaluation setup therefore needs both a defined test set and a decision process for what happens when results worsen; a headline score alone cannot establish that a system is safe or accurate.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Plan agent permissions, recovery, and observability before launch
Javed’s pre-launch advice is to decide what an agent may produce and do, what requires human approval, and what access it needs. He also recommends choosing recovery behavior for misunderstandings and invalid tool calls in advance: retry, ask for clarification, fall back to a deterministic rule, or escalate as appropriate.
For observability, he says a team should be able to reconstruct the request, context, tools called, tool results, decision, and resulting action. That record helps explain how an outcome occurred; it does not make the model’s reasoning a source of truth.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →What Javed’s lesson means for AI system design
“Never let a model turn its own previous outputs into facts,” Javed says. In practice, that means preserving provenance, distinguishing user-provided information from generated content, and routing consequential facts or calculations through systems that can enforce and verify them. Models can still interpret language, summarize options, and help rank candidates—but their earlier answers should not silently harden into evidence.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
