savage-synthesis

Savage Synthesis (Resume)

The synthesizer runs in two stages:

• Stage 4 — merge the five specialist scratchpads into draft-v1.md.
• Stage 6 — revise the previous draft against the adversary's kill-list into draft-v{N+1}.md.

This skill defines the rules for both stages.

Inputs available

• ingest-summary.md — the neutral Stage 1 summary
• claims.md — the Stage 2 structured claim list
• scratchpads/market-demand.md
• scratchpads/claim-audit.md
• scratchpads/rejection-modes.md
• scratchpads/relevance.md
• scratchpads/positioning-critique.md
• (Stage 6 only) kill-list.md + draft-v{N}.md

Merge rules

1. The synthesizer is not an editor. It does not soften. It does not hedge. It does not introduce new claims; every assertion must trace to a scratchpad line or a claim line.

2. Contradictions between scratchpads stay visible. If the market-demand-hunter says the stack is in demand and the relevance-checker says the role is eroding under AI in 24 months, both statements appear and the conflict is flagged under Red flags with both sources.

3. Down-weighting is explicit. If an input is T3+, the synthesizer writes [source-tier: T3] and the confidence drops accordingly. It does not paper over the tier. The candidate's own LinkedIn is T3 about the candidate.

4. Every claim from claims.md — quantitative, qualitative, differentiation, skill — gets handled in Claims-vs-reality with one of four verdicts:

   • Supported — claim-auditor found T1/T2 evidence that confirms it
   • Partially Supported — evidence confirms a weaker version of the claim (e.g., "grew revenue 300%" vs. a filing showing ~40% growth for the line the candidate owned)
   • Contradicted — T1/T2 evidence directly contradicts
   • Unverifiable — no T1/T2 evidence either way; marked [UNSUPPORTED] in the downstream report

5. Every assertion in the draft ends with [confidence: High|Med|Low — <one-line reason>] and, where an external source is involved, [source-tier: T<n>].

6. The Verdict paragraph is written last, after all other sections, and may contain no hedging. The verdict is one of:

   • Ship-ready for the stated target role. (rare — requires nearly all claims Supported or Partially Supported, and market-demand + relevance-trajectory both positive)
   • Savable with surgery. (specific fixable defects; list them under What would change the verdict)
   • Dead for the stated target role. (fundamental positioning, market, or claim-integrity issues that can't be fixed with a resume edit)
   • Incoherent target; cannot verdict. (positioning-critic flagged fatal framing; verdict on the nearest well-formed target is supplied instead)

   If the evidence does not support any of the above directly, the verdict paragraph says so as a factual statement, e.g. "Insufficient independent evidence to verdict this resume for the stated target role. Specifically: <X, Y, Z were unverifiable against T1/T2 sources>."

Revision rules (Stage 6)

1. Read kill-list.md top to bottom. For each entry, either:
   • Fix the flagged passage in the draft, or
   • If the entry is factually wrong (the adversary misread the draft or the evidence), write a one-line rebuttal in the corresponding place and keep the passage.
2. No passage left unhandled. Every kill-list entry must be either fixed or rebutted in the revised draft.
3. New evidence is not introduced in revision. If the kill-list demands evidence the scratchpads don't contain, downgrade the claim to [UNSUPPORTED] or Unverifiable — do not fabricate.
4. Never soften a verdict in revision unless the kill-list specifically provides new supporting evidence that was missed. A kill-list entry that argues "be nicer" is out of scope and should be rebutted.

Output

• Stage 4 → draft-v1.md
• Stage 6 → draft-v2.md, or if the self-audit rejects v2, draft-v3.md.