1 · LLM extract
Proposes value form, claim, evidence flags, and alignment confidence from raw text. Conservative — flags only when text supports them.
Work → Artifact → Outcome → Confirmation. Story & Trace Assembly drafts the 30-second read and nine-layer Trace in that grammar. Grades are derived from value-form-specific evidence thresholds, not from narrative confidence.
Proposes value form, claim, evidence flags, and alignment confidence from raw text. Conservative — flags only when text supports them.
Evidence coverage (0–100) measures progress toward Established threshold. Confidence blends coverage + LLM alignment.
deriveGrade sets the legal grade. LLM never inflates above what JUDGMENT_RULES allow. Witness attestation can cap or contest.
A candidate enters private review only at ≥90% confidence with exact AI and user quotes verified against captured turns. It renders on a shared Trace only after explicit user approval.
Passive capture scans automatically. Directed mining parses the user's time window, topic, and value form. Both enter shared-discovery-v3 and produce the same pre-built draft Trace shape.
Passive trigger
Captured conversations arrive automatically.
Directed trigger
“Mine my Q1 chats for vendor negotiation.”
Template-driven assembly (story-trace-assembly-v2) drafts situation → key decisions → outcome and the full Trace document. Every field is editable at confirm. Voice stays plain and specific — never AI marketing copy. Heavy edits train later drafts as style corrections.
Story
Situation · key decisions · outcome. Tools fold into the sentences.
Nine layers
Claim through contribution, including Work → Artifact → Outcome → Confirmation.
Voice + corrections
Banned hype phrases · ≥30% edits become drafting style signals.
Work
Judgment and decisions before the outcome moved
Artifact
The thing that shipped or changed
Outcome
Measured delta or value form claim
Confirmation
Witness, audit, or external system
Work → Artifact → Outcome → Confirmation. Story & Trace Assembly drafts the 30-second read and nine-layer Trace in that grammar. Grades are derived from value-form-specific evidence thresholds, not from narrative confidence.
An LLM pass detects explicit correction, rejection, and divergence patterns in captured conversations. Candidates remain private until the user confirms the exact quoted context.
Override
Human explicitly rejects or replaces the AI recommendation.
Error caught
Human identifies a factual, logical, technical, or source error.
Divergence
Human chooses a materially different approach and explains why.
Prompt in, pre-built draft Traces out — then Story & Trace Assembly fills the 30-second read and nine-layer Trace. Mining changes the filter—not the shared clustering, drafting, assembly, grading, or review pipeline.
Prompt intent
Parse time window, topic, and value form.
Shared discovery
Same filter → cluster → draft path as passive capture.
Draft Traces
Never a blank form: Work → Artifact → Outcome → Confirmation.
Learn + measure
Confirm, split, or toss; track mining share of confirmed Traces.
3 similar tosses suppress that cluster type
Health: mining-confirmed ÷ all confirmed Traces
Guardrail: mining is the power tool, never the front door
Structured assertion — limits visible, minimum evidence not met.
A structured professional assertion with its limits visible.
Correct minimum evidence for this value form.
The correct minimum evidence for this value form is present.
Full evidence threshold for this value form.
Full evidence threshold for this value form is met.
Amber = Supported minimum · Green = Established full threshold
S = Supported · E = Established · OR rules in cards above
| Form | Temporal | Artifact | Counterfactual | Contribution | Independent | Adoption | Reinvestment |
|---|---|---|---|---|---|---|---|
| Revenue | · | E | E | E | E | · | · |
| Cost avoided | · | E | E | E | · | · | · |
| Time saved | · | E | E | E | · | · | E |
| Quality | · | E | E | E | · | · | · |
| Risk | · | E | · | E | E | · | · |
| Durable asset | · | E | · | E | · | E | · |
Markdown reference: docs/grading-model.md · Sync rule: .cursor/rules/grading-model-sync.mdc