The Heredity Experiment

Decisions locked

Choices already made in the design — annotate if any feels wrong.

The journey — follow one method through the machine

Take a single method — say, “adversarial verify with a falsify-first harness on opus” — and walk it through the three stages.

1. The claim on trial. Tonight the lifecycle got a truth rail: green runs bank receipts nobody re-derives. The hypothesis says the same rail can make the lifecycle improve itself: if every receipt names the methods that produced it, “which of our methods actually work” stops being intuition and becomes a number the ledger computes — and if that number predicts, choosing methods becomes a query and the shop’s judgment compounds without you ratifying every lesson. It sounds true. Lots of things sound true. Each stage below is built to kill it cheaply if it’s false.
   ↘ go deeper — the four hypotheses, formalized

   H1 (heredity is representable): ≥80% of dispatches attribute to a finite gene tuple (engine×effort, harness shape, workflow shape, skills, brief class). Kill: <60% attribution — method identity is fiction at our granularity.

   H2 (fitness discriminates): per-gene settlement outcomes show variance beyond sampling noise, stable in sign across a temporal split. Kill: uniform, or sign-flips between halves.

   H3 (fitness predicts): fitness from window one predicts window two with positive lift over base rate (Brier / log-loss). Control: 1000× label permutation must destroy the lift. Kill: no lift, or the scrambled control keeps it.

   H4 (selection lifts the ceiling): the fitness-routed arm beats the judgment-routed arm on pre-registered outcomes in a paired trial. Kill: no difference at the pre-registered bar.

2. Stage A — the past testifies (days, zero new machinery). Our history already carries method provenance: the engine bay’s ledger holds 443 dispatches with the engine used, the kind of brief, whether an adversarial leg ran, and how each ended — green, red, refuted, or infra-death. Compute each method’s fitness from the first half of history; test whether it predicts the second half. Then the load-bearing move: scramble the labels and demand the prediction collapse. This stop is where the whole arc earns the right to exist — or dies for the price of an afternoon.
   ↘ go deeper — instruments & analysis discipline

   Sources: the engine-bay run ledger (engine, slug→brief class, adversarial flag, verdicts, elapsed, repo, retries); quanta’s in-tree + off-tree ledgers; gate history. Infra-deaths (sandbox/infra failures) are excluded from fitness and enumerated separately — they indict the bay, not the method.

   Gene tuple v0: (engine, brief class, adversarial leg). Coarse on purpose; refined only if H1 holds.

   Discipline: the analysis script and outputs are committed; the lead recomputes headlines from raw; an independent leg re-runs it (the reviewer is never the author). Stratify by repo and task class to catch the “good methods got easy tasks” confound.

3. Stage B — receipts learn to cite their methods (one additive field). Only if Stage A shows real signal: receipts gain produced_by — the list of method identities that made the claim. Verification ignores it; old receipts stay valid; switching it off costs nothing. From that day, the ledger accumulates fitness data as a side effect of ordinary work, forever. This is the heredity mechanism: methods become things with lineages and track records instead of prose in doctrine files.
   ↘ go deeper — the field, sketched

   Additive optional array on the fact record: produced_by: [{gene, v}] — gene = a registry id (engine-bay card, saved workflow, skill path, gate harness id, spec template), v = content hash. Emission via quanta ledger bank --produced-by and a gate-tail env contract. Shape-validated, never required; ledger verify ignores it (fitness is not correctness). Ships only through the full build sequence, after your ratification.

4. Stage C — the trial that settles it (the authoritative test). Matched pairs of real tasks from the live dispatch stream. One arm: methods chosen by fitness query. Other arm: methods chosen the way I choose today — doctrine plus judgment. The verifiers who grade outcomes never know which arm produced the work. Pre-registered bar, no peeking, roughly two weeks of volume. If the fitness arm wins — fewer escaped defects, less rework, receipts still holding a month later — the ceiling provably moved. If it loses, heredity stays a metaphor and we keep the better tool with a banked null.
   ↘ go deeper — trial protocol

   Unit: task pairs matched on repo, task class, size band; alternating assignment within stratum. Target: ≥40 pairs before first look (~2 weeks at current bay volume). Outcomes, primary first: escaped-defect rate (found by standing adversarial audits after merge), rework cycles, receipts-still-hold at +30 days, wall clock. Blinding: the fitness query’s recommendation is logged before execution; grading legs see no arm labels. Analysis: paired per stratum against the pre-registered bar; the result banks to the ledger either way — a null is a settlement, not a failure to report.

5. Settlement — what the world looks like if it holds. Method routing runs off the ledger by default; my judgment reserves for the frontier; the escaped-defect trend bends down without a single new page of doctrine being ratified. Your job shifts from mechanic to breeder: you hold the conserved core and set the selection pressure; everything below compounds without either of us being the mechanism. And if it doesn’t hold — the ledger says so, permanently, and the stigmergy arc proceeds without the heredity claim.

Your three calls

What could go wrong — named, with guards

Reference strip

Hypotheses: H1–H4 · Instruments: bay ledger · quanta ledgers · gate history · Controls: label permutation · blinded pairing · Mechanism: produced_by · Decision points: your three calls · Failure honesty: risks · Canonical source: docs/ai/EXPERIMENT-adlc-heredity.md