p4vc-core · architecture · compounding engine
Signal → Evidence → Assessment → Memory → Δ → Intervention → Outcome → Signal
The Compounding Engine is the named loop that runs across Foundation Core and the seven canonical agents. It is not an eighth agent. It is not a new data layer. It is the architectural commitment that every founder interaction makes the next one smarter — because each turn deposits a layer, no turn erases the prior turn, and the radius grows.
Something happens. A founder speaks, ships, stalls, or is mentioned.
the seven stations
Each station has an owner, a write target, and an evidence rule. The loop will not advance past a station whose evidence rule has not been satisfied.
station 01
liveSomething happens. A founder speaks, ships, stalls, or is mentioned.
owner · Any of the seven agents · ingested through /api/public/signal
station 02
·The signal is recorded as a citable observation, not yet a claim.
owner · Leaf node that emitted the signal · sometimes promoted by operator
station 03
·Compass scores the founder against the readiness axes using the new evidence.
owner · Compass · reads core.diagnostic_metric, writes provisional scores
station 04
·Compass proposes a memory write. Decision Engine gates it.
owner · Compass proposes · Decision Engine decides · operator may co-sign
station 05
·Mirror Loop diffs the new memory against the prior version. Names what changed.
owner · Mirror Loop · reads memory version chains
station 06
·Breakthrough or Compass proposes the next-best action — as an explicit bet with Expected Δ.
owner · Breakthrough Command / Compass · gated like every other memory write
station 07
·Observed Δ is compared to Expected Δ. The residual is the second-order signal — did we predict correctly?
owner · Whichever node observes the outcome · calibration scored by Compass (owns the dimensions)
governance · the rules the loop obeys
rule 01
Memory is versioned and superseded forward; both the old and the new are readable. The history of a founder's interpretation is part of the founder's interpretation.
rule 02
Every memory write carries a human-readable rationale and at least one evidence_ref. The Decision Engine rejects writes that don't.
rule 03
Known fact, inferred pattern, and speculation are distinguished. progression.confidence_note records the level at the time of writing.
rule 04
When two sources disagree, progression.contradiction_flagged makes the tension first-class. The operator resolves it on the record.
rule 05
Readiness scores are navigational. Narrative judgment remains primary. Founders are never reduced to a number.
rule 05b
No intervention without Expected Δ: predicted dimension, magnitude, confidence, evaluation window. The Decision Engine rejects intervention writes that don't predict. Without a prediction, no outcome can be wrong — and a system that cannot be wrong cannot learn.
rule 06
The operator surface always exposes what Core does NOT yet know about a founder — empty axes, stale signals, unresolved interventions.
the second-order loop · expected Δ
Every intervention is a bet with a stated prediction. Every outcome is a comparison, not just a measurement. The residual — observed minus expected — is how the system grades itself over time.
intervention · the bet
expected_delta: [
{ dimension: "narrative_clarity", magnitude: +0.8, confidence_pct: 74, window_days: 14 },
{ dimension: "investor_readiness", magnitude: +0.3, confidence_pct: 62, window_days: 14 }
]Decision Engine rejects intervention proposals without Expected Δ. No prediction → no write.
outcome · the grade
observed_delta: [{narrative: +0.1}, {confidence: +1.3}, {gtm: 0}]
expected_delta: [{narrative: +0.8}, {investor_readiness: +0.3}]
residual: [{narrative: -0.7}, {investor_readiness: -0.3}, {confidence: +1.3 unpredicted}]
calibration_score: 0.41 // Brier, lower is betterCompass owns the dimensions, therefore Compass scores the bet. Breakthrough proposes; Compass grades. The ±15 calibration cap stays honest in both directions.
what this unlocks
design boundaries · what this isn't
boundary
The canonical agent suite is locked at seven: Polaris, Compass, Mirror Loop, Breakthrough Command, Astra Outreach, Partner Engine, Oracle. Compounding is the loop they run; promoting it to an agent would duplicate what already exists.
boundary
No new core.* tables. Identity, evidence, memory, decisions, audit — all already exist. Compounding adds one memory axis (progression), seven topics, envelope additions for Expected Δ, one read view, and three admin RPCs.
boundary
Cross-founder pattern recognition belongs to Oracle. Trajectory diffing belongs to Mirror Loop. Intervention proposal belongs to Compass/Breakthrough. Calibration scoring belongs to Compass (owns the dimensions). This surface reads their output; it does not duplicate them.
operator surface
The per-founder progression view dramatizes the loop in operator-mode: the readiness constellation, the evidence ledger, the memory versions over time, the open interventions, the next-best-action, and — most importantly — the panel that lists what Core does NOT yet know about this founder.