research_notes/experiments_FINAL_PROTOCOL_NOTES.md

# Locked standard continual-learning evaluation

The immutable protocol and all executed Python sources are in
`final_standard/source/`; their SHA-256 values are bound by
`final_standard/source_lock.json`. The lock was written on 2026-10-08 at
20:18:11 UTC, before the first final worker loaded official test data. The
manifest SHA-256 is
`af4ac3635681ac02faa2b36181830f45654c3f9549601258e2fe4aa879308825`.
The 260 jobs were launched at 20:18:38 UTC.

There are thirteen methods, ten paired fresh seeds, and two datasets (MNIST and
Fashion-MNIST). Every run has five two-class arrivals. The class permutation is
fixed by the run seed and shared across methods. Output labels are remapped to
arrival rank, but inference always predicts from all ten outputs without a task
ID or a seen-class mask. The training procedure knows the arrival boundaries.
This is a class-incremental benchmark with bounded streams, not a task-free,
unbounded online deployment study.

The globally fixed source partition is 48,000 fitting images, 6,000 validation
images, and 6,000 independent calibration images. The validation partition is
unused by final learners. Its earlier proxy subdivision supported development;
the real calibration images were excluded from every development fit and model
selection. Ordinary baselines may fit all 54,000 nonvalidation images, while
certificate methods spend 6,000 of the same total observed-image allowance on
calibration and fit only 48,000. The no-controller center ablation uses the same
48,000 fitting images as the scalar center, isolating that controller. No
baseline-trained representation is reused by a certificate method.

| Method | Fixed final configuration | Main scope |
|---|---|---|
| SGD | Shared 784–128–128–10 ReLU MLP, SGD 0.1, two epochs per arrival | No memory |
| ER | Same MLP; concatenated current/replay cross-entropy | Reservoir over presentation events |
| DER++ | Current CE + 0.5 all-logit MSE + 0.5 replay CE | Two independent replay draws; pre-update targets |
| ER-ACE | Reference scale-10 cosine head and asymmetric current loss; two epochs | Strong replay baseline |
| ER-ACE, four epochs | Same loss/model, four epochs | Explicitly compute-relaxed control |
| Cumulative RFF ridge | 512 fixed RFF, bandwidth 0.5, ridge 0.001 | Global sufficient statistics; no feature learning |
| Global prototypes | 128 MiniBatchKMeans prototypes/class on normalized pixels | Fits all 54,000 baseline images |
| Scalar center | Same RFF map, projected ridge center, 64 added geometry directions, risk 0.1 | Preserves possibly inaccurate acquired center predictions |
| Center without scalar controller | Same fitting data, feature map, center and geometry update | No retention certificate |
| Immutable prototype regions | 128 prototypes/class, risk 0.1, one mutable MLP outside earlier gates | Atomic two-class admission; refused regions stay unprotected |
| Fresh renewal, risk 0.2 | Cumulative 128 prototypes/class; 30 fresh class tests | Deploy only fully accepted prefixes |
| Fresh renewal, risk 0.1 | Same learner and archive | Predeclared evidence-resolution variant |
| Stagewise joint reference | Reinitialized MLP trained for two epochs on all seen data each stage | Explicit noncontinual access and memory exception |

All methods except the joint reference have a 6 MiB persistent payload budget.
This includes model and feature-map tensors, replay/prototype/Gram/basis state,
active certificates and ledgers, retained calibration archives, source IDs, and
serialized non-tensor checkpoint metadata. Reservoirs are charged at allocated
capacity, rather than their currently filled size. The joint reference's
retained raw images, labels, and IDs are counted and its budget exception is
explicit. Transient fitting buffers, parameter gradients, Python allocator
overhead, and evaluator copies are not called persistent payload; separate
per-process peak RSS captures their memory effect.

Renewal retains one shared prototype array and an accepted-prefix index. Its
unused calibration archive is released one arriving class pair at a time,
partitioned into disjoint future-stage blocks, and counted. A rejected proposal
does not replace the accepted predictor. Before any acceptance, zero logits are
an explicit unprotected fallback. The ordinary global-prototype baseline fits
more images and therefore does **not** use identical learned prototypes. Each
renewal run also saves unconditional predictions from its own exact proposal,
which supplies the identical-prototype controller diagnostic without another
fitting run.

The statistical failure allocation is 0.05 **per seeded run and per policy**,
paid across every predeclared attempt, including failed acquisitions. It is not
a simultaneous 95% claim across twenty dataset/seed repetitions or across
policies. All policy results are reported. Conditional population statements
assume independent representative class-conditional calibration evidence; data
provenance establishes partition separation, not the absence of benchmark-to-
deployment distribution shift. Renewal's statement concerns each accepted
checkpoint and does not bound ever-error on a fixed input over time.

Saved predictions support full class and task matrices, final micro/macro
accuracy, mean acquisition accuracy, average incremental accuracy, backward
transfer, and signed forgetting. Backward transfer averages final-minus-
acquisition accuracy on the first four tasks. Forgetting averages the largest
previous post-acquisition accuracy minus final accuracy on those tasks, so an
improvement may give a negative value. Forward transfer compares a task's
accuracy immediately before its arrival with the same model's initialization
accuracy. Intransigence compares with the paired stagewise joint reference.
Mean, sample standard deviation, and Student-t 95% intervals are reported across
the ten paired seeds; paired-difference intervals use matching seeds. These are
descriptive comparisons without multiplicity-adjusted superiority claims.

Full ten-way softmax ECE uses fifteen equal-width confidence bins and no fitted
temperature. Prototype logits have an arbitrary declared scale of four and
ER-ACE a cosine scale of ten, so calibration metrics compare those output
conventions rather than claiming calibrated posterior models. NLL and Brier
are a separately hashed **post-lock descriptive analysis** requested after
launch, implemented by `analyze_nll_brier.py`; they cannot influence fitting or
model selection.

Each worker uses one PyTorch/BLAS thread in its own process. Four synthetic
resource-probe processes used 1,962,400 KiB in summed peak RSS, below the 6 GiB
admission threshold. Training workers may contend, so their wall times are
labeled accordingly. After all workers exit, serial batch-one and batch-128
inference measurements use three warmups and twenty-one fixed-checkpoint trials.
Reported arithmetic counts are component estimates: dense neural multiply/adds,
RFF products, ridge products/solves, covariance products, and dominant KMeans
mini-batch distance products. Transcendentals, loss/norm operations, QR/SVD/eigh
implementation details, KMeans++ initialization, and bookkeeping are excluded
explicitly. Solver calls, fitting examples, optimizer steps, and measured times
provide additional cost evidence; these estimates are not advertised as full
hardware FLOP measurements.

The benchmark is executed from the captured source with:

```bash
python research/omega/experiments/final_standard/source/run_final.py \
  --output-dir research/omega/experiments/final_standard \
  --data-root research/omega/experiments/data \
  --execute-locked --workers 4
```

Existing completed or failed captures are preserved; the runner does not
silently overwrite or retry them. For a fresh reproduction, use
`reproduce_final.py --captured-root <capture> --output-dir <fresh> --data-root
<data> --workers 4`. That wrapper verifies source, environment, and archive
digests; copies only `source/` and `source_lock.json`; and runs the unchanged
captured implementation into fresh directories. It never duplicates or
overwrites captured prediction/state evidence. The original lock-creation
helper assumes the development directory layout, so it is not used by this
portable reproduction route. The downloaded archive hashes and fixed
partition-ID files remain part of data provenance. Failures, if any, remain
visible in final aggregation.