docs/EXPERT_REVIEW.md

Expert and AI-agent review guide

EIDRION v1.0.0 is a conditional research prototype for maintaining and auditing functional-memory contracts. Its strongest review target is the precise connection between nested orthogonal centers, prefix energy, a bounded trainable residual, and one independently calibrated fixed exceptional set. Its experiments also show substantial practical conservatism. Review the mathematical statement, implementation, statistical provenance, and performance evidence separately; success in one category does not establish the others.

This guide proposes concrete review tasks and future experiments. The larger benchmark designs and research improvements below are unrun suggestions. They are not additional results of this release. Nothing here establishes historical priority, general continual-learning superiority, or a complete solution to catastrophic forgetting.

1. Begin with the evidence boundary

The supplied Hugging Face preparation contains the main research package: manuscript, source, saved JSON records, source locks, figures, previous audit reports, and executable engineering checks. The multi-gigabyte development, prediction, and state companions described in the original package are not included in the supplied attachment. A summary-statistic reconstruction from JSON is possible; reopening every raw prediction array or checkpoint requires those companions. Consult the prepared release's availability and manifest documents before calling any audit complete.

Within the extracted Eidrion_v1.0.0/ root, use a new output directory:

python release_tools/verify_release.py --mode quick --output ../review_quick_new

This checks main-file integrity, 39 grouped finite fixtures, ordinary all-layer training/resume, and generated-data extension tests. It is not a natural-data benchmark or population-calibration experiment. The manifest demonstrates consistency with the supplied hashes; it is not a signature or independent proof of authorship.

For a mathematical review, start with the numbered sources in manuscript/, then theory/nested_operator_risk_ledger.md and theory/numerical_retention_and_response.md. For the statistical alternatives, read prior_art/region_certificate_and_audit.md and prior_art/renewal_controller_audit.md. Historical notes preserve intermediate reviews; bind any current implementation statement to the current source hash and a fresh report rather than copying an intermediate hash from prose.

2. Keep the three contracts distinct

Contract	What is protected	Quantifier over future states	Main price
Nested operator memory	Agreement with the acquired reference's class prediction, including its original errors	All future admitted states share one fixed exceptional set; supports an ever-disagreement bound	Frozen external geometry; protected rank; center energy/radius; independent score calibration
Immutable decision regions	Correct class ownership on a calibrated immutable set	Future appended groups and fallback changes cannot override those owned points; supports an ever-error bound	Frozen features/gates; overlap refusal; growing prototype/decision-list storage and latency
Fresh checkpoint renewal	True-label class risk at each fully accepted checkpoint	One simultaneous confidence event covers every accepted checkpoint; error sets may differ over time	Fresh unopened labeled blocks; all-old-stratum testing; confidence and evidence consumption

The first contract does not convert a mistaken teacher into a correct classifier. The third contract does not inherit either of the first two contracts' fixed-set ever-error statements. An unaccepted initial model, rejected proposal, or proposal-training iterate receives no guarantee merely because a controller object exists.

3. Audit the nested operator contract

The core source is src/avenyra_omega/memory.py, models.py, and calibration.py; src/eidrion/__init__.py supplies the public alias. The mathematical sources are theory/nested_operator_risk_ledger.md and manuscript source 02_nested_memory.txt.

In fixed centered-logit space with fixed features phi, verify the following chain:

1. Protected spans are nested, the center belongs to its span, and each actual center innovation is orthogonal to the preceding span. Appending optional geometry directions is a separately charged rank expenditure.
2. Projection onto every old span recovers its old center: Pi_s M_t = M_s. Orthogonal cumulative innovation energy gives ||M_t - M_s||² = ||M_t||² - ||M_s||². Old residual power does not increase.
3. The complete trainable residual obeys the joint centered-output norm cap. The bound must apply to the entire raw network after any admitted finite parameter update, not merely to a projected last-layer gradient or a tangent approximation.
4. The deviation from an old center is at most p_s(x) * [sqrt(N_t - N_s) + R_t], with N_t = ||M_t||² and p_s = ||(I - Pi_s)phi||.
5. The reference class map, center, span, score definition, tie rules, risk, and failure allocation are fixed before the acquisition's independent calibration sample is opened. The signed distance is the center margin toward the reference class divided by sqrt(2); negative scores must expose loss during consolidation.
6. The lower tolerance threshold c_s comes from the declared order statistic. Every later amplitude satisfies the strict inequality sqrt(N_t - N_s) + R_t < c_s. Equality is not admitted. Atoms and zero-power conventions are part of the proof.
7. All future admitted disagreements lie in the same set {z_s < c_s}. Establish that fixed-set inclusion before asserting an ever-disagreement statement. A bound for each separately selected time is insufficient.

Check that the reference is actually the intended acquisition predictor. When a deep residual is compressed into the shared center, calibration must count the mismatch; it must not assume lossless consolidation. If the old reference is discarded after calibration, show that the retained center, basis, feature map, threshold, acquisition energy, radius contract, and confidence ledger still determine every later admission test.

Meaningful adversarial fixtures include an unrestricted global ridge refit that moves an old read, center-preserving deletion of an independent protected direction, a calibration-trained gate that vanishes only on observed calibration points, finite-anchor preservation with an unseen ReLU change, and contradictory labels on the same input mass. These distinguish real premises from an overbroad interpretation. Existing fixtures are in verification/verify_core.py and prior_art/audit_omega_core.py.

The full-network response claim is narrower than an optimization-success claim. With positive power and radius, the smooth radial map preserves the rank of the centered raw-network Jacobian; it can still worsen conditioning. Check finite differences through all raw layers, small-power gradients, large-logit saturation, and a dormant product with both factors initially zero. A finite safe restart can escape that particular zero-gradient state, but does not prove useful response, global convergence, or arbitrary target fitting.

4. Treat numerical validity as a separate proof obligation

theory/numerical_retention_and_response.md supplies an explicit sufficient extension for center-membership defects, innovation leakage, imperfect powers, forward errors, and nonnested span drift. The released float64 diagnostics do not constitute the rigorous global enclosures assumed by that extension.

For a literal numerical certificate, verify bounds on acquisition defects tau_s, old-span innovation leakage, true norm differences, power-envelope error rho_t, forward error epsilon_t, and any accumulated span-drift term. A predeclared additive allowance eta_s ||phi(x)|| is subtracted from the signed Euclidean decision distance before independent calibration. Do not mix this distance with the raw logit margin and lose the sqrt(2) factor.

The sufficient numerical admission conditions include an additive-error allowance and a strict multiplicative-budget inequality. At zero residual power, center/reference argmax agreement alone is insufficient: additive leakage can flip a tiny margin. The conservative numerical convention requires a positive adjusted margin. Increasing eta_s after inspecting calibration scores needs a new valid simultaneous argument or fresh calibration; it cannot be retroactively justified by relabeling a diagnostic.

Reviewers should try long ill-conditioned update sequences, nearly dependent basis columns, repeated reorthogonalization, cancellation in prefix accumulators, cross-device and batch-size changes, and a zero-power input with an arbitrarily small margin. Compare estimates with exact arithmetic or independent high precision on small constructed fixtures. A measured small defect is useful evidence of quality, but a uniform bound must cover unseen inputs and the entire permitted numerical execution domain.

5. Audit immutable decision regions

The source is src/avenyra_omega/regions.py; the construction and tolerance proof are in prior_art/region_certificate_and_audit.md. This mechanism is a calibrated selective nearest-prototype decision list with chronological ownership.

Verify that each current prototype proposal and feature definition are fixed before fresh evidence is opened. For a true class c, an existing correct owner has score -infinity; an existing wrong owner or new local classification mistake has +infinity; only an unowned, locally correct prediction receives its squared distance. The upper tolerance radius uses the declared upper order statistic. A +infinity radius is refused. A -infinity radius can represent coverage already supplied by an older correct owner.

The calibrated bad event is Z > r, so the inference gate is closed, distance <= r. This boundary convention differs from the nested amplitude contract's strict admission threshold. Test a large atom exactly on the boundary and deterministic nearest-prototype ties. Confirm that every routed forward argmax equals its declared owner, including any configured logit scale.

Install a group atomically only after every claimed class passes. Later groups cannot overwrite an old owner, and mutation of proposal arrays cannot mutate an admitted copy. On failed acquisition, every opened test still consumes its failure allocation. Check that checkpoint restore preserves ownership, radii, labels, feature semantics, and the complete attempt ledger.

Decisive controls use the identical fitted prototype bank with ordinary global nearest-prototype inference, the identical chronological architecture with a predeclared uncertified gating rule, and a matched fallback learner with replay. Report wrong old ownership, current-class overlap refusal, covered class mass, unowned mass, latency versus number of groups, and all retained prototype/feature/metadata bytes. Existing fixtures are in prior_art/audit_omega_regions.py.

6. Audit fresh checkpoint renewal

The source is src/avenyra_omega/renewal.py; its contract and corrections are documented in prior_art/renewal_controller_audit.md and prior_art/final_core_api_audit.md.

For each candidate, bind a checkpoint, output labels, preprocessing, feature map, and inference semantics before opening any unused blocks. Evaluate every persisted old stratum plus every newly claimed stratum through that exact inference path. For E errors out of n, acceptance uses the binomial lower-tail test P[Binomial(n, beta) <= E] <= delta; verify its direction independently and test its equivalence to the one-sided Clopper-Pearson upper bound. Very small delta should use a numerically appropriate inverse survival function rather than computing 1 - delta first.

Each attempted test spends confidence and consumes its source observations, including failed tests and refused candidates. An early refusal cannot return an already inspected block to the unused reserve. Source IDs must identify underlying originals, not augmented views. ID uniqueness alone does not prove independence: audit the actual fitting, feature extraction, tuning, and candidate-selection data flow.

On rejection, confirm that the previous predictor itself is preserved. A string candidate ID is insufficient if model tensors were changed in place. A controller checkpoint must load with its matching deployed predictor, required strata, protected strata, evidence ledger, and budget. Test omission of an old stratum, omission of a claimed new stratum, source-ID reuse, duplicate IDs within one call, budget exhaustion, first-candidate refusal, and corrupted deployment metadata. Existing independent cases are in prior_art/audit_omega_renewal.py.

Always include the finite counterexample in which different accepted checkpoints err on disjoint subsets. It prevents the per-checkpoint risk theorem from being incorrectly advertised as a fixed-set ever-error theorem. Report the unchanged proposal learner deployed unconditionally alongside the validated learner using identical fitting data/prototypes; this isolates the cost of acceptance/refusal from improvements in the underlying classifier.

7. Review the experiments with causal controls

Read experiments/FINAL_PROTOCOL_NOTES.md, FINAL_RESULTS.md, BASELINE_DEFINITIONS.md, final_standard/source_lock.json, and research_status.json. Verify the lock and all predeclared method/seed lists before analyzing results. The source benchmark has two datasets, ten paired seeds, thirteen methods, five two-class arrivals, and ten-way inference without task identity.

The new certified methods have no best mean-accuracy wins on the two tested image datasets. Fresh renewal at risk 0.20 differs from two-epoch ER-ACE on MNIST by +0.051 percentage points, with a descriptive paired 95% interval spanning -5.789 to +5.891; that does not establish an advantage. The Fashion-MNIST result is substantially worse. Preserving these negative findings makes the scalar/evidence bottlenecks testable.

Ordinary controls fit 54,000 examples; certified policies fit 48,000 and allocate 6,000 to calibration. The center/no-controller comparison has matched fitting data. Four-epoch ER-ACE is a compute-relaxed control. The small two-epoch joint reference is not an optimized upper bound. Review paired differences and all-seed distributions rather than selecting a favorable dataset, risk, or seed after seeing outcomes.

Keep persistent learner payload, process RSS, peak training memory, arithmetic component estimates, and archival evidence size separate. Count frozen backbones, cached features, unused calibration examples, labels/source IDs, failure ledgers, optimizer state, and simultaneously live proposal/deployed models as appropriate to the declared boundary. Resource-matched claims require explicit time and memory measurements; a component arithmetic estimate is not measured total FLOPs.

When companion captures become available, use the read-only prediction, checkpoint-contract, and aggregation scripts in prior_art/ with fresh outputs, following REPRODUCE.md. Do not use partial mode to certify a full release. Preserve missing files, execution failures, refusals, repaired archives, and source-lock mismatches in the report. Reconstructing stored predictions is different from retraining the 260-run benchmark.

8. Prioritized development roadmap

Priority	Research target	Required scientific advance	Decisive success criterion
1	A directional fixed-set retention bound	Replace the global scalar energy penalty with a valid anisotropic functional constraint without losing a fixed independently calibrated exceptional set	Strictly more useful admitted acquisition at matched risk/evidence/memory while preserving theorem premises and exact counterexample resistance
2	Certified representation migration	State a common observation/function space and charge old-center, power, margin, and forward-error changes under encoder/span migration	A mathematically justified transfer or fresh recertification protocol, followed by improved plasticity over a frozen-geometry control
3	Economical valid evidence reuse	Prove validity for a declared family or adaptive data-access protocol; simple recycling of outcome-inspected blocks is insufficient	Lower fresh-evidence cost with controlled type-I error and complete failed-attempt accounting, compared with fresh-block renewal
4	Rigorous numerical admission	Implement outward error enclosures or a fully charged deterministic error domain for the numerical supplement	Zero-power and ill-conditioned tests pass with verified upper bounds, not only empirical small residuals
5	Query-aware protected-rank allocation	Estimate a declared future-query family and price its information/control relevance without assuming unknown future statistics	Better risk/plasticity per counted byte across held-out query laws, with uncertainty and misspecification reported
6	Modern-model evaluation	Extend valid contracts, provenance and resource accounting to stronger representations and longer streams	Reproducible gains against strong matched controls on predeclared modern benchmarks; configurations alone do not count

A directional improvement must retain uniform validity over the actual future candidate family. An empirical covariance ellipsoid by itself need not contain rare margin-sensitive directions. Begin with exact finite-dimensional examples where its domination condition can be proved, then test non-Gaussian tails, nearly invisible decision directions, and adversarial adaptive proposals.

Representation migration changes the function used to define both old scores and powers. Reusing the old threshold after an encoder update is invalid without a proved domination relation or fresh recertification. Start with isometric coordinate changes preserving the entire protected span and center, then controlled nonisometric maps with explicit error charges. Basis deletion that preserves only today's center is already defeated by a supplied counterexample.

For evidence reuse, distinguish repeatedly checking the same fixed score's already-certified threshold from using old sample outcomes to select a new score or candidate. The former is part of the nested contract; the latter needs a new validity argument. Candidate routes include finite-family simultaneous validation, a predeclared reusable-holdout mechanism, or an explicitly designed sequential procedure. Each proposal must specify what adaptivity it permits, how unopened information remains independent, and what confidence/utility cost it pays. These are research directions, not implemented guarantees.

theory/evidence_capacity_closure.md gives a useful reference model for the fifth target: known conditional Gaussian covariance, specified queries, noiseless linear measurements, and a fixed quadratic cost yield query-weighted residual eigenmodes. It also distinguishes a later revealed repair request from a controller seeing only stored evidence. Audit this information pattern before transferring the closed forms into a learning policy. Dimension is not bit capacity; quantization and noisy observations change the model. Unknown future queries or estimated covariance do not inherit exact Gaussian optimality.

9. Suggested larger protocols — all unrun

Existing extension assets are documented in extensions/README.md; generated-data tests ran, while actual CIFAR-100/ResNet18 extraction, full reserved Permuted-MNIST, pretrained transformer adaptation, and GPU/distributed benchmarking did not. Begin with the provided frozen-feature paths before attempting a new learned-encoder contract.

For each new protocol, register the following before accessing official held-out evaluation outcomes or sealed calibration values:

• Task/class/domain order, underlying source identities, fixed output universe, and inference inputs. Permutations or augmented views must preserve original identities; reserve independent originals before making views.
• Fitting, development, calibration, and final-evaluation partitions; frozen backbone/checkpoint/preprocessing hashes; every permissible calibration access; risk and confidence allocations over all planned attempts.
• Methods and causal controls, seeds, tuning allowance, epochs/optimizer settings, feature/prototype/replay sizes, and total retained memory including encoder and evidence. Declare whether a comparison is memory-matched, data-matched, compute-matched, or a compute-relaxed diagnostic.
• Primary endpoints, statistical summaries, refusal/exhaustion behavior, missing-run policy, and immutable source/data locks. Save proposal and deployed states separately before/after every attempt.

Proposed tier	What it would test	Essential comparisons	Required reporting
Reserved CIFAR-100 with frozen ResNet18 features	More classes and a stronger fixed representation	Identical-data unconditional prototypes, cumulative ridge, replay, fresh renewal; nested variant only after implementing its feature path	All classes per accepted checkpoint; evidence resolution, refusal rate, backbone bytes, fit/evidence/total latency
Reserved-original Permuted-MNIST	Domain arrivals and original-source leakage control	Same-original-identity replay/ridge/prototype controls; fresh renewal with disjoint preallocated evidence blocks	(task,class) risk strata, old/new-domain accuracy matrix, unique original counts, calibration exhaustion
Strong frozen vision features with longer class streams	Capacity and certificate conservatism under better geometry	Matched encoder and fitting subset; matched representation baselines; scalar/directional variants only under valid contracts	Paired-seed performance, protected rank/energy, decision margins, calibration use and refusal causes
Pretrained transformer with bounded adapters	Whether the entire trainable raw adapter network can obey an external fixed envelope usefully	Frozen model, ordinary matched adapter/replay, bounded adapter with identical fit and evaluation data	A task/output/risk definition fixed before calibration, model/cache/evidence bytes, acquisition versus retention, conditioning and cap saturation
GPU/multi-device numerical study	Contract consistency and numerical error across hardware	CPU/high-precision small-fixture oracle, multiple precisions/batches/devices	Measured discrepancies and rigorously charged allowances separately; no bitwise-match claim without evidence

For transformer tasks, define the certified output before experimenting. A fixed finite class/logit vector fits the present theorem. Autoregressive sequences, expanding vocabularies, token-conditioned shifts, and unrestricted generation introduce extra quantifiers and distributions; they need a derived contract rather than being treated as routine API substitutions.

A useful primary comparison is whether a new certificate admits more correctly acquired classes at the same retained memory, risk, confidence, and observed-data allowance. Final accuracy alone can conceal refusal. Report acquisition, average incremental and final accuracy, backward transfer, signed forgetting, per-class risk, accepted checkpoints, unaccepted proposals, early uncertified fallback, evidence consumption, and any true-label/reference-agreement distinction. Use matched paired seeds, report uncertainty, and predeclare any confirmatory superiority test and multiplicity correction.

10. Issue and review format

Reports should be independently actionable. Identify whether the issue is a theorem defect, assumption violation, numerical implementation defect, statistical-provenance defect, resource-accounting error, reproduction/packaging defect, or performance concern.

Title:
Category:
Release version and file hashes:
Contract/claim being tested:
Exact source or mathematical passage:
Assumptions preserved and assumptions intentionally changed:
Environment, dependencies, device/precision and thread settings:
Minimal reproducer, source IDs and fixed random seed:
Expected result under the stated contract:
Observed result and quantitative discrepancy:
Fresh output paths and artifact/checkpoint hashes:
Impact: theorem, operational claim, reported result, or documentation only:
Suggested correction or next decisive experiment:

For a theorem counterexample, give all dimensions, matrices/functions, distributions, tie rules, and admissible states; show exactly which claimed inequality or quantifier fails while its stated premises hold. For a statistical objection, identify the information that leaked and the sample-selection path. For an empirical improvement, preserve the old source lock, register the new protocol, and compare against matched controls using every planned seed. A corrected theorem, a newly valid contract, and a better benchmark result are different contributions and should receive separate status labels.

AI agents can assist with independent derivations, exact small fixtures, implementation reading, metric reconstruction, and adversarial test generation. Their agreement or fluent review is not proof, historical-priority verification, peer review, or a substitute for reproducible evidence. The most valuable review ends with a precise invariant, counterexample, verified artifact, or explicitly scoped open question.