Adapt without losing a baseline.
Compare unchanged, new-only and replay candidates. Preserve provenance, measure retention and defend a controlled update.
Useful starting points: replay selection, task-family curves and release and rollback. All remain directly accessible.
Project 14 · version 1. Save no-adaptation, new-only and replay comparisons with the same baseline, update budget, buffer size and frequency setting. Each capture stores numerical checkpoints, not an authenticated training history. This project does not alter published path requirements.
A trained candidate and a replayable release decision
A 2–3–1 tanh network has 13 learned parameters. Its two inputs are x and a task tag. Old targets are y = x; new targets are y = −x + 0.5. Train the old baseline, freeze it, then compare candidates from exactly those weights.
Old training: 0/100 updates. Candidate: 0/200 updates. Each update processes 32 fitting rows; repeated rows are counted as work, not new evidence.
Replay uses 16 new rows and 16 buffer draws per update; new-only uses 32 new rows. Equal updates match total fitting examples, not new-task exposure. No adaptation consumes the declared comparison slots but performs no gradients. The 160 receipts reuse 32 unique training rows; buffer size counts retained receipts. If task tags and IDs are encoded numerically, five float64 fields would occupy 40 bytes per receipt; actual JavaScript strings/objects and JSON overhead are excluded.
| Monitoring MSE | Baseline | Candidate |
|---|---|---|
| old | Not frozen | 0.38584 |
| new | Not frozen | 0.98536 |
Shared 0–2 MSE scale; 0 bars clipped. Exact values above. The separate 16-row monitoring grids never enter gradients and are repeatedly visible development evidence.
Inspect replay receipt membership
Old receipts: 8; new: 8. Oldest retained time: 26.
| Row ID | Task | Stream time | x | y |
|---|---|---|---|---|
| 7 | old | 41 | -0.067 | -0.067 |
| 2 | old | 27 | -0.733 | -0.733 |
| 0 | old | 69 | -1.000 | -1.000 |
| 10 | old | 37 | 0.333 | 0.333 |
| 11 | old | 26 | 0.467 | 0.467 |
| 4 | old | 85 | -0.467 | -0.467 |
| 12 | old | 115 | 0.600 | 0.600 |
| 3 | old | 56 | -0.600 | -0.600 |
| 108 | new | 159 | 0.067 | 0.433 |
| 111 | new | 146 | 0.467 | 0.033 |
| 102 | new | 147 | -0.733 | 1.233 |
| 100 | new | 109 | -1.000 | 1.500 |
| 109 | new | 148 | 0.200 | 0.300 |
| 100 | new | 129 | -1.000 | 1.500 |
| 101 | new | 118 | -0.867 | 1.367 |
| 112 | new | 35 | 0.600 | -0.100 |
Versioned gate · adaptation-release-v1
Require old MSE regression ≤ 0.02 and new MSE improvement ≥ 0.01. Thresholds are fixed before comparison. These are public development gates, not a safety guarantee.
Saved comparisons and rollback
0/6 candidates. Known comparison count: 0. Matched three-way evidence: outstanding.
| Candidate | Strategy | Updates | Old MSE | New MSE | Gate |
|---|
State: development; no rollback recorded. A frozen selection is not an automatic promotion.
Your evidence and decision
Protocol: identify task/data splits and matched checkpoint. Retention/adaptation: cite both metrics and curves. Cost: count example uses and memory separately. Provenance: state all interventions and known exposure. Decision: justify weights versus context/memory, limits, release and rollback.
Open the implementation studio and dependency-free starter. Full source files remain local; pasted code never executes here.
Report not submitted. Submission checks matched evidence and text presence, not reasoning or code quality. Human artifact review and learner pilots remain pending.
Separate objective transfer cases
Open each case explicitly. Four criteria and at most three submissions per case. The first attempt is retained; retries remain practice. These public local checks do not certify independent implementation.