V4.1.4 applies three classes of fixes derived from Review Squad analysis of a live debate transcript.
The core finding: voice profiles describe tics but the model ignores them without enforcement.
Descriptive voice prose was converted to hard RULES with explicit must-appear-once-per-turn
constraints for every agent. Structurally, a third round was added — each agent posts a single
falsifiable claim before Nando delivers the final verdict — and opening prompts now require a steelman
of the strongest counterargument. Two systemic prompt patterns that made every agent sound identical
were explicitly banned.
-
Banned “[Name] didn’t just [verb]” opener
Every Round 1 opening used the same structural frame — five agents, five identical openers. The opening prompt now explicitly states:
Do NOT open with "[Name] didn't just [verb]" — find a different entry point. The instruction targets the pattern, not a specific name, so it applies regardless of topic or assignment.
-
Rebuttals now target agents before Nando
Every Round 2 rebuttal opened with “Nando called it X — but…”, meaning agents were rebutting Nando’s recap rather than each other. The rebuttal prompt was rewritten: agents must first name the strongest argument made by one other agent and explain why it doesn’t beat theirs, then address Nando’s critique. The
Do NOT open with "Nando called it X — but" instruction is now explicit.
-
Emily — four enforced constraints
"Look" must appear at least once per turn to grab attention before a key point. At least one sentence per turn must be under six words. Must never end on an abstraction — end on a human being or a concrete dollar figure. Must self-interrupt once with a dash: actually — better example —. The "look" to redirect attention voice description was insufficient; the model needs a must-appear rule to actually produce it.
-
FC — two enforced constraints
No sentence exceeds 25 words. Each turn must end with a single-sentence closer that doesn’t build — it just lands. Period, not ellipsis. The existing “short declarative sentences” voice description produced sentences that were short on average but not capped — the 25-word hard limit enforces it.
-
Jared — three enforced constraints
"The problem is" must appear at least once per turn. One sardonic aside per turn in parentheses. Open cold — first sentence IS the argument, no framing or build-up. The sardonic aside was listed as “occasional” in the voice profile; without a per-turn mandate it never appeared in practice.
-
PM Cory — three enforced constraints
"Sure, [opposing point] — but" must appear at least once per turn. "Here's the thing" must appear at least once per turn. Must self-interrupt once with a dash to sharpen a word mid-sentence. The voice profile mentioned both phrases as signals; moving them to must-appear rules ensures they appear every time rather than occasionally.
-
Nando — two enforced constraints
When naming an agent’s argument or flaw, always use their first name directly and conversationally:
Jared — that's the sharpest rebuttal of the round. Every agent whose argument is assessed gets their name said out loud. The previous voice profile said “call people by first name” but Nando consistently used generic phrasing like “one agent argued…” rather than naming agents directly.
-
Round 3 — falsifiable claims
A third round was added after rebuttals. Each agent posts a single sentence: a falsifiable claim that, if proven wrong, would undermine their entire position. This forces agents to commit to a testable prediction rather than retreating to untestable assertions.
generateFalsifiable() is a new function with a distinct prompt; the final verdict prompt was updated to reference “falsifiable claims” in the transcript description.
-
Opening statements include a steelman
The opening prompt now requires agents to acknowledge the strongest counterargument to their position in one sentence, then show why their position still wins. This prevents agents from presenting only upside and gives Nando more signal on who actually understands the tradeoffs.
-
Announce message states judging criteria upfront
Nando’s opening announcement now reads: “Judged on: clarity of position, strength of evidence, quality of rebuttals. Best argument wins.” Previously agents debated without explicit criteria, which gave Nando more room for arbitrary verdicts. Stating criteria upfront orients all agents toward the same evaluation standard.
-
Flow header updated — 6 steps
The file header comment was updated to reflect the new flow: opening statements → Nando interim verdict → rebuttals → Round 3 falsifiable claims → Nando final verdict. Step numbering renumbered from 5 to 6 steps accordingly.