0023 — Deal-breaker tech demotes a posting; it no longer deletes it¶

Status: accepted (2026-07-28). Extends ADR-0015 (soft desired signals) to the negative case; supersedes the hard filter described in ADR-0001 and filtering.

Context¶

Since V1, a posting whose text word-matched any deal_breaker_tech seed term (Kafka, Spark, Flink, Scala, Hadoop) was removed from silver_jobs outright. It never reached gold, so it never reached the digest — regardless of title, location or how many desired technologies it named.

Three properties of that rule turned out to be wrong in combination:

1. It matched the whole description. clean_text is the entire posting with HTML stripped — requirements, nice-to-haves, the company's stack blurb, boilerplate. There is no section awareness.
2. One mention was enough. No threshold, no weighting.
3. It ran before every other signal. A perfect title with six desired technologies died on a single "Kafka a plus".

Measured on a full 167-board run (2026-07-28): 123 postings passed the location gate and were then deleted by this rule.

Casualties included Senior Software Engineer, AI Productivity (4 desired techs, one Kafka mention), Senior Data Platform Engineer, and several Data Engineer roles carrying 4–6 desired technologies — all of them roles the rule was never meant to exclude.

The rule was trying to express "I don't want a streaming/big-data engineering job". What it actually expressed was "I don't want any posting that mentions these words anywhere", and since no deleted posting named a deal-breaker in its title, the signal it keyed on was never the one that identified the job.

Decision¶

Deal-breaker tech becomes a negative signal, not a filter. silver_jobs gains:

• deal_breaker_hits — how many distinct terms matched (0 = clean);
• deal_breaker_terms — which ones, sorted and comma-separated, null when none.

Both flow to gold. The digest orders by deal_breaker_hits ascending first, so flagged postings appear below clean ones, and prints the matched terms in the posting's signal line ("mentions Flink, Kafka, Spark"). Nothing is removed.

Location is now the only hard filter in silver.

This makes the negative case consistent with ADR-0015 (desired tech/title annotate, never drop) and with ADR-0020 (the V2 fit score orders delivery, never filters it). The pipeline now has one rule about rules: V1 keyword signals rank; only geography excludes.

Consequences¶

• Gold grew from 1,179 to 1,302 postings on the measured run. 46 of the 123 recovered have no desired technology at all and will sort to the bottom; 30 carry three or more.
• "Spark" alone accounts for 57 of the 123, and single-term matches for 88 — the shape you would expect from nice-to-have lists rather than from streaming roles.
• The digest gets longer. That is the accepted cost: a posting you can dismiss from its one-line summary is cheaper than one you never learn exists.
• Naming the terms matters more than counting them. "Kafka" reads very differently from "Flink, Hadoop, Kafka, Scala, Spark", and the reader can tell them apart at a glance where the pipeline cannot.
• This is a stopgap for a judgment V1 cannot make. Distinguishing "Kafka required" from "Kafka a plus" — and an analytics role that happens to mention Spark from a streaming engineering role — is reading comprehension, i.e. V2's scorer (ADR-0020). When that lands, deal_breaker_terms becomes an input to the score rather than a sort key.

Alternatives rejected¶

• Threshold at ≥2 distinct terms — would have recovered 88 of 123 and kept the genuinely streaming-heavy roles out. Rejected as arbitrary: it encodes a guess about how many times a technology must appear to matter, which is exactly the judgment being deferred to V2.
• Scope the match to the title — would have recovered all 123, since none name a deal-breaker in the title. Rejected because it makes the rule almost inert: it would only ever catch a title like "Senior Spark Engineer", which the desired-title signal already fails to match anyway.
• Weighting the tech seeds (strong BI tools scoring above generic SQL/Python/AWS) was considered alongside this and rejected outright: an analytics role with no BI tooling and a Tableau-heavy pure-frontend role are the same keyword footprint with opposite verdicts. No weighting of term presence can separate them.