The Medium-Tier Phrase Blocklist

#What this Playbook gives you

A three-tier phrase blocklist that catches predictable LLM drift patterns before they reach the Sonnet 4.6 verifier. The blocklist is the cheap defense; the verifier is the expensive one. Together they're better than either alone.

perea runs the Medium tier in production. Strict and Loose are documented as the tiers above and below — pick the tier that matches your risk tolerance and audience expectations.

#The three tiers

Strict. Blocks urgency manipulation, forward-looking predictions, engagement-bait, and common AI-tells. Suited to highly-trusted brands where any sign of LLM authorship hurts. Phrases like "delve," "in this article," "it's worth noting," and AI-tic transitions are blocklisted alongside the substantive drift patterns.

Medium (perea's tier). Blocks the three substantive drift categories: urgency manipulation, forward-looking confidence, and engagement-bait. Does not block AI-tells — perea is comfortable with the model's natural voice as long as the content stays honest.

Loose. Blocks only engagement-bait. Suited to channels where forward-looking framing is part of the brand voice (e.g., a prediction-market account where forward-looking is the product).

#What each category catches

#Urgency manipulation

Phrases that manufacture urgency the source paper doesn't actually contain:

• "right now"
• "today only"
• "before it's too late"
• "the window is closing"
• "act fast"
• "limited time"

The category is about false urgency. If a regulation has an actual effective date of 2026-12-31, naming the date in a post is fine — the deadline exists in the source. The blocklist catches the pattern where the LLM invents a deadline that isn't in the paper.

#Forward-looking confidence

Phrases that turn hedged claims into confident predictions:

• "by 2027, X will ..."
• "this will reshape ..."
• "the future of X is ..."
• "X is going to ..."

The Hybrid Contract already locks numerics and named entities. The forward-looking blocklist is the second line of defense: even if the LLM keeps numerics correct, it sometimes drops the hedge. "Could become" is the source; "will become" is the drift. The blocklist catches the substitution.

#Engagement-bait

Phrases that prioritize click/like/share over information:

• "you won't believe ..."
• "the one thing X doesn't want you to know"
• "thread"
• "🧵"
• "subscribe for more"
• "follow for daily takes"
• "did this surprise you?"

perea's standing instruction for the close is: link to the source, nothing else. No CTA, no "thread," no engagement question. The blocklist enforces that.

#Why the blocklist is the cheap defense

The Sonnet 4.6 verifier costs real inference per draft. The blocklist is a regex match — milliseconds. Catching a draft at the blocklist saves the verifier round-trip on that draft and gets the optimizer back into the retry loop sooner.

Three-tier framing helps:

• Blocklist hits in production are concentrated in 2–3 patterns. Track the rate by pattern.
• When a new drift pattern appears, the blocklist update is one line. The verifier update is a prompt change with all the testing that entails.
• The blocklist is operator-editable in production. The verifier prompt is not.

That said, the blocklist is not a substitute for the verifier. The verifier catches semantic drift (a number that's wrong, an entity that's renamed); the blocklist catches lexical patterns (a specific phrase the LLM picked up from its training distribution). Both layers are needed.

#Production results

perea seeded the Medium-tier blocklist on day 1 of tick-X-post-from-research running. The seed list lives at state/blocklist.json in the skill's state directory and is editable by the operator.^[1]

Hit rate. In the first 14 days, the most-hit category was engagement-bait — the optimizer occasionally tries to add a 🧵 emoji or a "thread" hint at the start. The Medium tier rejects this on regex, the optimizer retries, the second draft drops it.^[2]

Tier escalation policy. The operator may temporarily escalate to Strict (e.g., during a launch when brand voice matters most) by setting blocklist.json to the Strict tier. The escalation is reversible by editing the same file. No code change.

Recurring drift → contract update. When a phrase recurs across retries (not just one draft), the operator's standing instruction is to update the Hybrid Contract section, not just the blocklist. The contract is the primary defense; the blocklist is the secondary. A recurring pattern indicates the contract is silent on something it should speak to.^[3]

#Failure modes

• Over-blocking the model's voice. If you start blocklisting common transition words ("interestingly," "notably," "for example"), the drafts become stilted. The Medium tier deliberately leaves these alone.
• Under-specified patterns. A simple word match on "thread" will also reject the legitimate use of the word ("the thread of the argument"). perea uses contextual patterns — ^thread: at the start of a line, 🧵 as a standalone marker — not bare-word matches.
• Tier drift. Over weeks, the blocklist accumulates patterns. Without occasional pruning, the list grows until the rejection rate hurts throughput. perea reviews the blocklist quarterly.
• The retry loop loops. If the optimizer is blocked on the same phrase 3 times, the retry loop logs the slug to unpostable.json. The operator reviews. Usually the fix is a contract update; sometimes it is a blocklist correction.

#How to apply in your stack

1. Pick a tier based on brand. Most pipelines should start at Medium.
2. Seed the blocklist as a JSON file with one entry per category. Each entry has a regex + a one-line reason + a category tag.
3. Run the blocklist before the verifier in your pipeline. If the blocklist rejects, surface the matched phrase + reason to the optimizer for the retry.
4. Log every hit. The hit log is your signal for whether a phrase belongs at a higher tier (rare hit → keep at Medium; constant hit → escalate or update the optimizer prompt).
5. Schedule a quarterly review. Prune unused patterns. Add patterns that recurred across drafts.

The blocklist itself is small — perea's seed is ~30 patterns total across the three categories. Most of the value is in the framing: tier picks, hit-rate tracking, and the contract-update standing instruction.

#Quotable Findings

The blocklist is the cheap defense; the verifier is the expensive one — together they catch more than either alone.^[1]

The Medium tier blocks urgency manipulation, forward-looking confidence, and engagement-bait — three substantive drift categories — without blocking the model's natural voice.^[1]

When a phrase recurs across retries the operator updates the Hybrid Contract, not just the blocklist — the contract is the primary defense, the blocklist is the secondary.^[3]

The most-hit category in early production was engagement-bait — the optimizer occasionally adds a thread emoji or a "thread" hint at the start; the regex rejects on first pass.^[2]

A simple word match on "thread" rejects legitimate use ("the thread of the argument") — perea uses contextual patterns, not bare-word matches.^[1]

#References