The Compounding Lab

Open the experiment. The question is concrete: does a capability learned on Mandate A improve work on a different Mandate B after its costs are counted?

A five-step experiment

1. Select continuity, regime change or archive ablation. Choose a future seed and cost assumptions, or supply your own series.
2. Freeze a forecasting policy learned from Mandate A. Inspect its training commitment, policy and full learning cost.
3. Execute B on four disclosed baseline/treatment arms. Inspect each prediction, actual held-out value and error.
4. Review current-stack and treatment evidence separately. Record elapsed review times, identity, rationale and decision.
5. Download the complete Evidence Docket ZIP or portable run JSON. Replay either with the native agent.

Changed inputs clear the frozen capability, results and reviews. Failed computations cannot be accepted by changing a status field. A review is bound to the exact run digest. Imports replay before replacing the current workspace. The last valid run survives reload and can execute offline after the gallery cache has installed. Imported and recovered reviews keep their original timing provenance. A new browser review requires fresh timers for both arms; it cannot relabel another review's reported durations as newly measured browser time.

What the learner actually does

Mandate A provides 20–64 bounded integer observations. A walk-forward selector compares last value, mean, linear least-squares trend and seasonal policies with periods 2–8. A seasonal candidate needs at least two cycles plus one observation before it is considered. Each candidate's validation begins at max(3, period); mean absolute error determines selection, with a fixed policy order breaking ties. The capsule stores the selected policy, training digest, validation scores, number of forecast calls, contract and rollback rule.

The selector receives only A. It neither receives B's observations nor calls an expected-answer function. The policy type and period are frozen; application to each B calibration prefix adapts its level/scale using that prefix alone. Every B suffix remains outside policy selection. Different policies have different training validation windows, which is an explicitly bounded heuristic, not a state-of-the-art forecasting claim.

The four measured arms are:

Arm	Policy	Prior information and charges
B0 null	Repeat last value	B calibration only
B3 static	Linear least-squares trend	B calibration only
B5 current stack	The same walk-forward selector, on each B prefix	All B selection calls charged
B6 treatment	A's frozen policy, applied to each B prefix	All A selection calls charged in full, plus B prediction calls

The default future prefixes have six observations, insufficient to select a five-period policy under the learner's two-cycle rule. A has forty observations. This difference in prior information is the point of the transfer test; it is not evidence that no stronger baseline could infer the period. B1 incumbent, B2 neighbor and B4 strongest single agent are unmeasured. Both B5 and B6 predict the same eight held-out values per task using fixed local arithmetic. No LLM, API, network tool, paid service or external evaluator is used.

Continuity creates new scale, level and phase variants of a disclosed periodic family. Regime change uses linear future series, making the learned seasonal policy fail. Ablation removes the capsule, so B6 reproduces B5 exactly. These are public synthetic examples, not secret held-outs, independent experiments or external economic outcomes. Custom specifications permit different training and future data. Content and identifier duplicates are rejected; this is a guard against obvious reuse, not a general semantic leakage detector. Imported provenance is self-reported.

Accounting and gates

Predictions use integer milliunits, with ties rounded away from zero. The exact formulas are shared by two implementations, Python and JavaScript. The dossier retains every prediction and error.

raw gain = B5 total absolute error − B6 total absolute error

overhead = (B6 forecast calls − B5 forecast calls + full validator replay calls) × assumed call rate + coordination cost

adjusted advantage = raw gain − overhead − (B6 review time − B5 review time) × assumed human rate

Call counts are measured invocations, not equal FLOP costs. Rates are explicit modeled units, not money. All A selection calls are charged in the current trial; there is no hidden amortization. Creation executes the full comparison twice, checking exact equality, and conservatively charges the entire validator pass to the trial. The ledger counts all forecast calls in both passes, including the extra prepared-but-unused capsule in ablation. Whole-run wall time covers both passes. Later UI preparation, exports and requested replays add work outside this creation timing and modeled advantage; no end-to-end deployment efficiency claim is made. Tokens and network tool calls are zero because this algorithm invokes neither. Energy and monetary compute costs are unmeasured, not zero. Human timers measure elapsed time, not verified active attention; CLI inputs are labeled operator-reported. Timings and identities are bound to the digest but are not independently attested.

The local claim requires a positive raw gain, positive cost-adjusted advantage, the declared maximum forecast-error bound, enabled archive and an explicit accepted review. Reject and repair remain HOLD. Large modeled cost or human review overhead can turn a raw win into HOLD. Forecast error is a bounded diagnostic, not a measure of financial, social or deployment risk. The manuscript promotion gate remains HOLD until missing baselines, independent validation, multi-agent scaling, α-WU calibration and delayed real-world outcomes are established.

ECI is E2: executed locally. Replaying in another runtime does not automatically create independent E3 evidence. A local stress example is not a substitute for the paper's independent stress and external validation requirements.

Native CLI and portable files

alpha-agent transfer-run --scenario seasonal --seed 37 --output run.json
alpha-agent transfer-verify run.json
alpha-agent transfer-review run.json --decision accept --reviewer "Operator" \
  --reason "Inspected predictions and costs; accept this bounded result" \
  --control-ms 12000 --treatment-ms 14000 --output reviewed.json
alpha-agent transfer-docket reviewed.json --output evidence-docket.zip
alpha-agent transfer-verify evidence-docket.zip

Use your actual review measurements; the example durations are illustrative. No init, identity secret or network connection is needed for this offline experiment. Use --spec spec.json for custom inputs. Browser JSON imports accept native reports unchanged. Browser-generated ZIPs pass the same native verifier as native exports.

The exact directory contract from manuscript pages 37–38 is:

00_manifest.md                 07_replay_logs/
01_claims_matrix.md            08_cost_ledgers/
02_environment.md             09_safety_ledgers/
03_benchmark_tasks/            10_validator_reports/
04_baselines/                 11_alpha_wu_calibration/
05_agialpha_runs/              12_summary_tables/
06_proof_bundles/              checksums.json

Empty scientific obligations are explicit status records. The α-WU record is uncalibrated with a null value. All raw results replay; changing a prediction, capsule, claim, review binding, file checksum or ZIP path is rejected. ZIP verification uses fixed in-memory paths, entry/expansion limits and no filesystem extraction. JSON imports reject duplicate keys, extra fields, non-integer/non-finite values, oversized documents and excessive nesting. Hashes establish content integrity relative to a trusted copy, not signatures, reviewer identity or external truth.

Relationship to existing demonstrations

Proof Bloom continues the claim → jobs → proof → review → Chronicle path, with ECI terminology corrected. Its existing capability promotion and transitive revocation protocol stays compatible. The Compounding Lab is a separate portable experiment for future-task advantage. It does not silently insert a forecasting capsule into Bloom's registry or grant a permission to execute arbitrary code. See the manuscript map.