Executive Voice Runtime Builder

Overview

Analyzes how a user writes, not what they think, and builds a portable, machine-readable voice engram used to draft in their voice. The engram captures structure, tone, rhythm, register routing, hard limits, and vocabulary signatures. It does not capture business judgment, opinions, or permission to act without review.

The engram is a portable artifact: a structured file the user owns plus a paste-able portable_prompt. Amazon Quick has no auto-load hook for a custom drafting runtime, so the engram is delivered as a file the user loads or pastes, not registered as a background model.

Detailed schemas and scoring tables live in references/. Load them when needed:

• references/engram-schema.md: canonical engram object structure, persistence rules, and a worked example.
• references/authorship-confidence-gate.md: H-axis and C-axis scoring, inclusion rules, and the Native Voice Calibration Prompts procedure.
• references/analysis-dimensions.md: dimension sets per path and the Voice Fidelity scoring rubric.
• references/bootstrap-elicitation.md: prompt bank, rewrite bank, interview questions, and the minimum viable gate.
• references/governance-log-format.md: entry template, valid EVENT_TYPEs, and examples.

Workflow

Engram

A structured object (see references/engram-schema.md) that encodes a person's writing patterns. Written to a user-controlled file, never to memory or the knowledge graph.

Register

A distinct writing context with its own tone and structure (for example, leadership updates, direct reports, quick replies). One engram holds 4-8 registers.

Calibration Gates (per path)

Profile Confidence Labels

Voice Fidelity Score (1-10)

Evaluated across 7 dimensions (see references/analysis-dimensions.md): structural, lexical, cadence, judgment, emotional, channel fidelity, and anti-voice detection.

Role-Based Configuration

<Workflow - Entry and Routing description="Check for an existing engram, detect sample availability, and route to the correct path." tools=[get_current_time, folder_list, file_read, file_write, kg_search, search_all, file_rag_search] triggers=["User asks to build, capture, clone, refresh, or validate a writing voice or engram", "does this sound like me"]

1. [Agent] Get the current time via get_current_time for governance timestamps and metadata. Validate: An ISO-8601 timestamp is returned. If fails: Use the session date as a fallback and note it in the log.

2. [Ask user] Confirm the output directory for all artifacts (engram file, profile, governance log). Validate: User provides a writable path. If fails: Re-ask. Do not assume a default path (Rule 13).

3. [Agent] Check for an existing engram: list the output directory and read any engram file present (folder_list, file_read). If one exists, inform the user immediately. Validate: A clear determination of "engram exists" or "no engram" is made. If fails: Log TOOL_FAILURE and proceed carefully, treating existence as unknown.

4. [Agent] Detect sample availability across sources (kg_search, search_all, file_rag_search over sent mail, messaging, and indexed files). Bucket: Rich (25+) / Moderate (10-24) / Sparse (fewer than 10) / Nothing. Validate: A bucket is assigned with a source-by-source count. If fails: Log TOOL_FAILURE per failed source and continue with remaining sources.

5. [Decide] Route based on request and availability. If a mode was explicitly provided, jump directly to it.

   • "does this sound like me" / "voice check" / "validate" -> Workflow - Voice Validation.
   • Existing engram + "refresh" -> Workflow - Refresh.
   • Existing engram + "rebuild" -> route by sample availability.
   • No engram + Rich -> offer Fast Track or Full Build.
   • No engram + Moderate -> Fast Track (with reduced-confidence warning per Rule 11).
   • No engram + Sparse/Nothing -> Bootstrap. Validate: Exactly one path is selected.

6. [Ask user] Present the chosen path and wait for confirmation before proceeding. Validate: User confirms or redirects. If fails: Re-present with a simpler summary.

7. [Ask user] Role discovery: role, primary writing channel, and whether building for self or someone else. If for someone else, confirm consent per Rule 14. Validate: Role and channel captured; consent confirmed if building for another person. If fails: Do not proceed without consent when building for someone else.

8. [Agent] Initialize the Governance Log at the output directory ([FirstName]_Governance_Log.md) with a SESSION_START entry per references/governance-log-format.md (file_write). Validate: The log file exists and contains the SESSION_START entry. If fails: Retry the write once, then inform the user.

</Workflow - Entry and Routing>

<Workflow - Fast Track description="Build a DRAFT engram from 10+ discoverable samples, then calibrate to PRODUCTION." tools=[get_current_time, kg_search, search_all, file_rag_search, file_read, file_write, open_in_session_tab] triggers=["No engram with 10+ samples", "User selects Fast Track"]

1. [Agent] Source discovery. Check sent mail, messaging, and indexed files. Log each source success or failure. Report findings. Validate: At least 10 candidate samples found, or the user is warned per Rule 11. If fails: Log TOOL_FAILURE per source; if fewer than 10 usable samples, warn and offer Bootstrap.

2. [Agent] Rapid analysis. Pull up to 25 samples. Score each via the Authorship Confidence Gate (references/authorship-confidence-gate.md). If more than 40% score H2 or lower, trigger Native Voice Calibration Prompts. Analyze across the 12 Fast Track dimensions (references/analysis-dimensions.md) and build 10 in-voice + 10 not-in-voice markers. Validate: Every included sample has an H/C score; contamination check logged. If fails: Re-score excluded samples or trigger calibration prompts, then continue.

3. [Agent] Save DRAFT engram. Write the profile to [FirstName]_Voice_Profile.md and the engram (conforming to references/engram-schema.md, status DRAFT) to [FirstName]_voice_engram.yaml at the output directory. Read the engram file back to verify. Open the profile with open_in_session_tab. Validate: Engram file read-back matches what was written; log ENGRAM_SAVED and ENGRAM_VERIFIED. If fails: Log ENGRAM_SAVE_FAILED and surface the failure to the user.

4. [Ask user] Calibrate. Generate 3 test drafts in different registers. Ask for each: APPROVE / WRONG TONE / TOO GENERIC / REVISE. Validate: If 2 of 3 pass, upgrade the engram to PRODUCTION, re-verify, and log CALIBRATION_PASS and STATUS_UPGRADE. If fewer than 2 pass, keep DRAFT, log CALIBRATION_FAIL, and offer iteration. If fails: Keep the engram at DRAFT and explain what did not match.

</Workflow - Fast Track>

<Workflow - Full Build description="Build a PRODUCTION engram from 25+ samples across 3+ contexts with maximum fidelity." tools=[get_current_time, kg_search, search_all, file_rag_search, file_read, file_write, open_in_session_tab] triggers=["No engram with 25+ samples", "User selects Full Build"]

1. [Agent] Gather sources. Discover all available sources; target 40+ samples across 5+ contexts. Report and confirm coverage with the user. Validate: Sample count and context count recorded per source. If fails: Log TOOL_FAILURE per source; if coverage is thin, warn per Rule 11.

2. [Agent] Score samples via the Authorship Confidence Gate. If more than 40% contamination, trigger Native Voice Calibration Prompts. Save scored samples to writing_samples_compiled.md at the output directory. Validate: Every sample scored; contamination check logged. If fails: Re-score or trigger calibration prompts, then continue.

3. [Agent] Full analysis across the 20 Full Build dimensions (references/analysis-dimensions.md). Validate: Each dimension has evidence-backed findings. If fails: Return to sources for more samples in under-covered dimensions.

4. [Agent] Build the register matrix from evidence. 4-8 registers typical. Do not force a fixed number. Registers with fewer than 3 samples are marked LOW confidence. Validate: Each register has a confidence label backed by its sample count. If fails: Merge or drop registers that lack evidence.

5. [Ask user] Generate the comprehensive profile and present it for review. Validate: User reviews and confirms or requests changes. If fails: Revise per feedback and re-present.

6. [Ask user] Calibrate. Generate 5 test drafts (one per register). Collect APPROVE / WRONG TONE / TOO GENERIC / REVISE. Validate: 3 of 5 pass -> PRODUCTION; 2 of 5 -> DRAFT; fewer than 2 -> HALT. Log the calibration outcome. If fails on HALT: Explain the gap and recommend Bootstrap or more sources.

7. [Agent] Save and verify. Write the engram (references/engram-schema.md) with the calibrated status to the output directory. Read it back to verify. Close the governance log with the final outcome. Validate: Engram read-back matches; log ENGRAM_SAVED and ENGRAM_VERIFIED. If fails: Log ENGRAM_SAVE_FAILED and surface the failure.

</Workflow - Full Build>

<Workflow - Bootstrap description="Elicit native writing when discoverable samples are sparse, producing a PROVISIONAL engram." tools=[get_current_time, file_read, file_write, open_in_session_tab] triggers=["Fewer than 10 discoverable samples", "New employee", "User selects Bootstrap"]

1. [Ask user] Pre-flight. Confirm sparse history and explain: "I will ask you to write and revise a few things. About 20-30 minutes." Validate: User agrees to proceed. If fails: Offer to end or to paste existing samples instead.

2. [Ask user] Native writing sprint. Present 5-10 prompts from the prompt bank in references/bootstrap-elicitation.md, one at a time. Score all responses as H5-C1. Validate: 5 or more prompt responses captured. If fails: Continue prompting until the minimum is met or the user stops.

3. [Ask user] Rewrite and rejection test. Present 5 generic AI drafts from the rewrite bank for the user to revise. Capture what they changed, what sounded wrong, words they would never use, and preferred structure. Validate: 5 or more rewrites captured. If fails: Continue until the minimum is met or the user stops.

4. [Ask user] Voice preference interview. Ask the 10 questions in references/bootstrap-elicitation.md. Validate: 10 or more preference answers captured. If fails: Continue until answered or the user stops.

5. [Agent] Gate check and analysis. Enforce the minimum viable gate (5+ prompts, 5+ rewrites, 10+ preferences). If not met, HALT and log GATE_HALT. Otherwise analyze using the Fast Track dimensions and generate a PROVISIONAL profile. Validate: Gate met before analysis proceeds. If fails: HALT and explain exactly what is still needed (Rule 9).

6. [Ask user] Save and calibrate. Write the engram with status PROVISIONAL, all registers LOW or MEDIUM confidence, and refresh_trigger set. Read it back to verify. Run 2/3 calibration. Validate: Read-back verified; calibration outcome logged. Status stays PROVISIONAL on pass. If fails: Keep PROVISIONAL and note what did not match.

7. [Agent] Deliver. Present the profile honestly as PROVISIONAL and set a hardening schedule (refresh after 30 days or 25 new samples). Open the profile with open_in_session_tab. Validate: Profile delivered with an honest confidence label and refresh schedule. If fails: Correct any overstated confidence before delivering.

</Workflow - Bootstrap>

<Workflow - Refresh description="Update an existing engram with new samples and produce a delta report." tools=[get_current_time, kg_search, search_all, file_rag_search, file_read, file_write, open_in_session_tab] triggers=["Existing engram + user asks to refresh"]

1. [Agent] Load and assess. Read the existing engram file. Search for samples created after its updated_at timestamp. Report findings. Validate: Existing engram loaded; new-sample search completed. If fails: If zero new samples, offer alternatives (paste samples, recalibrate only, or end).

2. [Agent] Incremental analysis. Score new samples via the Authorship Confidence Gate. Compare to the existing engram: confirming patterns, contradicting patterns (drift), new patterns (expansion), and confidence upgrades. Validate: Each new sample scored and classified against the existing engram. If fails: Re-score ambiguous samples or ask the user about authorship.

3. [Ask user] Update and calibrate. Present a delta report for approval. Apply only approved changes. Write the updated engram (never overwrite without approval per Rule 2), read it back to verify, and run 2/3 calibration. Validate: Read-back verified; calibration outcome logged. Status may upgrade on pass. If fails: Revert to the previous engram state and log the revert.

</Workflow - Refresh>

<Workflow - Voice Validation description="Score a draft against an existing engram and give targeted feedback." tools=[file_read, open_in_session_tab] triggers=["User asks 'does this sound like me' or to validate a draft", "voice check"]

1. [Agent] Load profile. Read the existing engram file. Identify the relevant register from context. Validate: Engram loaded and a register selected. If fails: If no engram exists, offer to build one (route to Entry and Routing).

2. [Agent] Score the draft across the 7 Voice Fidelity dimensions (references/analysis-dimensions.md) and produce a 1-10 score. Validate: A score with per-dimension notes is produced. If fails: Re-read the draft against the register's traits.

3. [Ask user] Deliver: overall score, what works, what is off-voice, and the top 3 fixes. Offer a rewrite if requested (preserve content, match voice; ask for any missing judgment per Rule 8). Validate: Score and specific feedback delivered. If fails: Clarify the register and re-score.

</Workflow - Voice Validation>