tech-claim-rubric

name: tech-claim-rubric description: Use when evaluating technical claims in high-depth content section units (예: 문제 해결 / 상세 프로젝트 / 경력 기술서). Defines the 5-axis framework (A1 Technical Credibility, A2 Causal Honesty, A3 Outcome Presence & Clarity, A4 Ownership & Scope, A5 Scanability) plus 2 critical authenticity rules (R-Phys, R-Cross) used by tech-claim-examiner agent. Verb-scope inflation (previously a separate rule) is now caught by A4 integrity_suspected sub-flag (see a4-ownership-scope.md).

Overview

This document is the authoritative rubric definition used by the tech-claim-examiner agent to evaluate technical claims in high-depth content section units (예: 문제 해결 / 상세 프로젝트 / 경력 기술서). It does NOT perform evaluation itself — it defines the evaluation contract that the examiner follows.

The rubric consists of:

• Five evaluation axes (A1–A5): The core evaluation framework. Each axis produces one of three verdicts (PASS / FAIL / P1). All five axes use an absolute standard. A5 is additionally structure-agnostic.
• Two critical authenticity rules (R-Phys, R-Cross): Integrity gates evaluated separately from the axes. Both can trigger automatic REQUEST_CHANGES regardless of A1–A4 verdicts and structural_verdict. Verb-scope inflation (previously a separate rule, retired in v4) is now caught by A4 integrity_suspected sub-flag (see a4-ownership-scope.md).

Evaluation structure:

Downstream consumers: resume-forge and review-resume skills consume the examiner's output schema (see output-schema.md).

Why 5 Axes (Rationale)

The previous 11-axis system (v1) was retired by agent-council unanimous vote. The core problems with v1:

1. Over-engineered: Sub-axes (E3a, E3b, sub-dimensions with weighted formulas) added evaluation overhead without improving signal quality. Evaluators spent more time computing scores than assessing claims.
2. Backend-biased: The v1 depth criteria implicitly favored distributed systems and infrastructure work. Frontend, data, and product-engineering bullets systematically underscored due to structural mismatch.
3. Structurally rigid: Phase A/B/C routing imposed a fixed evaluation order that created decision-tree overhead. Simple bullets required the same ceremonial path as complex ones.
4. Catch-22 default-FAIL: The "FAIL unless proven" stance combined with the prohibition on inference created a rubric where legitimately strong bullets failed due to formatting choices rather than substance gaps.

The 5-axis redesign compresses the evaluation into five clear questions:

These five questions cover the full evaluation surface without sub-weighting formulas or cascading phase routing.

Anti-pattern → Axis Traceability Matrix

Common resume bullet anti-patterns, the axis that catches them, and the verdict tier produced.

Five Axes — Quick Reference

Axis Verdicts

Each axis (A1–A5) produces one of three verdicts:

All five axes (A1–A5) use an Absolute standard: the passing bar does not move with experience level. candidate_context.years is referenced only by A4 for scope evaluation, not by A1.

A5 is structure-agnostic: it does not require a specific format (e.g., "action verb + metric + outcome"). It asks only whether the core message is extractable by a scanning reader within 6–30 seconds, regardless of how the bullet is structured.

A1 Evaluation Criteria

Question: Does this bullet reveal technical judgment? (이 bullet에 기술적 판단이 드러나는가?)

PASS — Bullet body explicitly shows all 5 of the following signals:

1. Constraint awareness — technical constraint to solve (throughput bottleneck, race condition, consistency gap, legacy coupling, cost ceiling, etc.)
2. Technology selection — a specific system/algorithm/pattern deliberately chosen
3. Mechanism — how the chosen technology works (partitioning key, memoization strategy, eviction policy, cutover path, etc.)
4. Trade-off / risk — cost/risk accepted, or rejection-reason for a rejected alternative
5. Rationale — context-based "why X over Y"

FAIL:

• Tool/library name drop only (no rationale)
• Outcome metrics only, no selection or mechanism grounding
• Generic verbs ("도입", "구축", "개선", "활용") with no what/how/why
• Feature noun-phrases ("seamless multi-step flow") masquerading as mechanism

P1 (canonical — a1-technical-credibility.md defers to this): Exactly 4 of 5 signals present (one signal absent), OR all 5 of 5 signals present but at least one signal is at name-level only (no mechanism or rationale behind it). Either condition fails the 5/5 strict PASS bar. Not vacuous enough to FAIL — examiner returns improvement hint targeting the shallowest signal.

Section-wide signal mapping: real-world에서 signal이 sub-bullets로 분산 가능 — section 전체에서 매핑 허용. 단일 bullet line에 모든 signal이 집중되지 않아도 section 전체에서 5 signals 모두 확인되면 PASS 판정 가능 — 단일 bullet line에 집중되지 않아도 허용.

Years are not referenced for A1. Ownership signals belong to A4, not A1.

Authenticity as Critical Rules

Authenticity is NOT a scored axis. It is a set of integrity gates evaluated as critical rules by the examiner. Rules are checked after all five axes are scored, and their triggered status feeds directly into final_verdict.

R-Phys — Physical Impossibility

Trigger condition: A numeric claim is physically or mathematically impossible given the described context.

Examples of R-Phys violations:

• "Reduced latency by 50,000%" (percentage improvement cannot exceed 100% of the baseline for latency reduction)
• "Increased throughput from 10 RPS to 10,000,000 RPS with a config change" (plausible magnitude would require hardware, not config)

Effect: Automatic REQUEST_CHANGES regardless of A1–A4 verdicts and structural_verdict. The examiner must name the specific number and explain why it is physically incoherent.

R-Cross — Cross-Entry Contradiction

Trigger condition: The claim directly contradicts another entry on the same resume, and both cannot be simultaneously true.

Example: Entry A claims "Designed and implemented the entire payment microservice from scratch." Entry B claims "Contributed to payment microservice API design alongside a team of 8 engineers." These contradict each other on ownership scope.

Applicability: R-Cross는 항상 emit된다. cross-entry context가 제공되지 않은 경우 triggered: false로 기록하고, reasoning에 "cross-entry context not provided"와 같이 absence를 명시한다.

Effect: Automatic REQUEST_CHANGES regardless of A1–A4 verdicts and structural_verdict. The examiner must cite both entries and identify the specific contradiction.

Verb-Scope Inflation — Retired in v4

The verb-scope inflation check (previously a standalone critical rule) is retired as a separate rule in v4. Detection is now handled by the A4 integrity_suspected sub-flag (see a4-ownership-scope.md). Detailed worked examples for R-Phys and R-Cross are in agents/tech-claim-examiner.md.

Evaluation Protocol

The examiner evaluates axes sequentially: A1 → A2 → A3 → A4 → A5. After all five axes, critical rules are checked.

Reasoning-Before-Verdict

For each axis, the examiner MUST:

1. Write technical reasoning first — what evidence is present, what is absent, what questions the claim raises
2. Provide an evidence_quote — a direct excerpt from the bullet text supporting the verdict
3. Assign the verdict last — PASS / FAIL / P1 derived from the reasoning, not assumed upfront

If the reasoning does not support the verdict, the verdict is wrong. Verdict-first reasoning is prohibited.

Critical Rule Invariant

After A1–A4 are scored and A5 emits structural_verdict, the following invariant is applied before generating final_verdict:

IF r_phys.triggered == true OR r_cross.triggered == true
THEN final_verdict = REQUEST_CHANGES
(regardless of A1–A4 verdicts and structural_verdict)

Final Verdict Derivation (when invariant does not apply)

Note: A5 verdict는 structural_verdict로 노출된다. structural_verdict == FAIL AND A1-A4 모두 PASS/P1 AND count(P1 across A1-A4) < 3인 경우 final_verdict = REQUEST_CHANGES를 트리거하되, consumer routing은 source-extraction이 아닌 readability-fix lane으로 처리된다. See output-schema.md §A5 Co-failure Disambiguation and a5-scanability.md.

P1 verdicts do not block APPROVE but are surfaced in interview_hints as improvement recommendations.

Output schema details (field names, types, required/optional) are in output-schema.md.

Output Format

The examiner's full output schema is defined in output-schema.md. Key fields:

INTERNAL fields (reasoning trace, not shown to candidates):

PUBLIC fields (returned to downstream caller):

Canonical definition: output-schema.md §Output Schema. 이 테이블은 downstream consumer를 위한 요약이며 schema 변경 시 반드시 동기 수정.

interview_hints Rules

• Written in the same language as the source bullet (Korean bullet → Korean hints)
• Do NOT include axis identifiers (A1, A2, etc.) in the hint text — hints are candidate-facing
• Each hint is actionable and specific, not generic ("add more technical detail" is prohibited)
• P1 verdicts from any axis generate a hint even when final_verdict is APPROVE

Migration from v1 (11-axis)

All v1 axis tokens (E1–E6, R1–R5, E3b, Constraint Cascade, CASCADING, LISTED, FLAT, Narrative Necessity, Layer 2, Phase A/B/C) are retired. They must not appear in examiner output or downstream skill prompts outside of this migration reference.

Completion Checklist

The examiner verifies all of the following before delivering output:

• A1 Technical Credibility: reasoning written, evidence_quote included, verdict assigned
• A2 Causal Honesty: reasoning written, evidence_quote included, verdict assigned
• A3 Outcome Presence & Clarity: reasoning written, evidence_quote included, verdict assigned
• A4 Ownership & Scope: reasoning written, evidence_quote included, verdict assigned
• A5 Scanability: reasoning written, evidence_quote included, verdict assigned
• R-Phys: triggered status explicitly stated (true / false)
• R-Cross: triggered status explicitly stated (true / false). cross-entry context not provided인 경우 false로 기록하고 reasoning에 absence 명시
• Critical rule invariant applied: if r_phys or r_cross triggered, final_verdict is REQUEST_CHANGES
• interview_hints written in source bullet language, no axis identifiers in hint text
• final_verdict determined and recorded