Status: design (2026-06-08) Issue: #274 Source: Claude Opus 4.8 System Card (Anthropic, 2026-05), §4.1.4 Related: #272 / #273 (Opus 4.8 behavioral-signal cluster); #247 (ethics-gate override recording, CLOSED)
The reviewer agents produce decision-useful review reports; a reviewer that buries its verdict in hedging is a quality regression. And ARS workflows routinely involve a user pushing back (relax an evidence standard, keep a shaky citation, soften a limitation) — exactly the surface where retract-under-pressure would show.
This is a prose strengthening of existing reviewer/writing guidance. It is NOT a behavioral proof.
The behavioral-check honesty bound (the crux). Acceptance #2 asks to "confirm hard boundaries + DA non-softening hold under 4.8 after pushback." That is an LLM-behavior property: it cannot be pinned by a deterministic CI test (retraction is runtime + probabilistic), and it must NOT be ticked by the author self-simulating a pushback dialogue — self-simulation is theater, not verification. So acceptance #2 is reframed as a prompt-surface confirmation: the live reviewer prompts contain explicit, adversarially-worded pressure-resistance instructions. A standing epistemic-status line states the limit plainly — a static document cannot guarantee runtime pressure-stability — mirroring the #272 discipline (guidance layer ≠ runtime enforcement). No claim of "confirmed 4.8 holds" is made anywhere in the change or its metadata.
This change ships no lint, no mutation test, no xfail pebble. "Be concise" is style guidance, not a contract invariant with a downstream consumer, so a drift-guard lint would be unrequested machinery. (Contrast #272, whose instruction/data boundary IS load-bearing and earned its lint.)
A short ## Output Discipline block is inlined immediately before ## Output Format in each live report-producing agent:
academic-paper-reviewer/agents/: domain_reviewer, methodology_reviewer, perspective_reviewer, eic_agent, devils_advocate_reviewer, editorial_synthesizeracademic-paper/agents/peer_reviewer_agent (its existing "conciseness" mention rates the paper's writing quality, not the agent's own output — a genuine gap)Inline (not a shared pointer) because the instruction must be live when the agent writes; a bare backpointer is invisible to the model at generation time (same rationale #272 used).
Base reviewer block text (the six reviewer agents share it verbatim; editorial_synthesizer adapts it to "decision letter / roadmap" and adds the evidence-standard rule, peer_reviewer adds a one-line note distinguishing it from rating the paper's writing, and devils_advocate points to its Attack Intensity protocol for pressure-resistance):
Output Discipline
Keep your review brief but complete. State each finding and your verdict directly; do not pad them with repeated qualifiers, apologetic framing, or restated caveats. Concise does not mean under-caveated — preserve every material uncertainty and limitation; cut only redundancy and hedging that adds no information. One clear statement of a caveat beats three softened ones.
The DA already carries a strong non-softening block (## Attack Intensity Preservation Protocol (v3.0), its Anti-Sycophancy Rules sub-block: rebuttal scoring, "do not soften after pushback", no-consecutive-concessions, persistent-pushback-≠-rebuttal, concession-rate tracking). #274 adds one general, evidence-standard-framed rule into that same sub-block (NOT an attack catalogue — public-repo safe), threaded to the existing concession threshold:
Pressure is not evidence. Repeated pushback, appeals to authority or status, or bare requests to soften do not by themselves change a finding. Revise only when the rebuttal supplies new evidence or reasoning that directly addresses the finding's stated basis. With no new substance, briefly restate the finding once and stop — do not expand caveats, apologize repeatedly, or retract a correct finding to preserve agreement.
The same evidence-standard sentence (adapted to "boundary/decision" rather than "finding") is added to editorial_synthesizer near its arbitration discipline, since the synthesizer also holds boundaries under EIC-arbitration pushback.
Each ## Output Discipline block carries a one-line status:
Epistemic status: these are prompt-surface instructions. They make the reviewer's output discipline explicit; they do not, and cannot, prove the model stays pressure-stable at runtime — that would need a separate non-deterministic behavioral eval.