Skip to content

Research program

Research question: can a small, useful classifier produce exact, transcript-bound verifiable decisions at an operationally useful cost, alongside an independently described hardware provenance claim? The result is a policy assay, not a proof of safety.

Falsifiable hypotheses

ID Hypothesis Required comparison Falsification / next decision
H0 Baseline reporting and integer semantics are internally consistent Exact reference vectors, independent histogram arithmetic, hash tampering, calibrated replay Any mismatch blocks further measurement; bootstrap tests cover software only
H1 A compact integer student preserves useful task quality at a fixed low FPR Task-matched existing guard, float parent, n-gram floor, held-out families Inconclusive FPR bound or unacceptable prespecified quality loss rejects adoption
H2 One chosen real backend proves the entire decision within an agreed budget Identical graph, sequence/chunks, hardware and inputs; include setup and failure rate Unsupported op, host/circuit mismatch or budget breach rejects that candidate
H3 The hybrid binds both genuine evidence objects to the exact same interaction Independent opening, nonce/audience/key/policy/chunk mutations Any accepted mismatch blocks an end-to-end claim
H4 Distillation improves the measured quality/cost frontier Same train/calibration/test identities, teacher supervision ablation, paired seeds No defensible frontier gain means keep the simpler student

Quality, latency and memory acceptance budgets must be chosen with the relying party before H1/H2. They are not invented after results. No real-data, hardware or proof hypothesis has been tested in this bootstrap.

Work packages and exit artifacts

  1. Foundation, implemented: strict schemas; harmless fixture; linear fitting; integer reference; calibration/uncertainty; immutable bundles; caller-pinned replay; offline tests. Exit is reproducible software evidence, not semantic classifier quality.
  2. Admission, next: one policy and relying-party use case; licence/source/lineage ledger; approved grouped split; powered benign test; baseline roster; frozen protocol and explicit compute budget. Exit is an owner-reviewed experiment proposal, with no downloads assumed.
  3. Operator qualification: real backend and one small graph, bit-exact conformance, negative proof verification and full cost record. Exit is a reproduced proof statement and limits; a toy circuit does not qualify an encoder.
  4. Quality/cost frontier: task-matched baselines, integer/float/PTQ/optional QAT, length and group slices, uncertainty and paired comparisons. Exit includes every failed candidate and a model-selection decision.
  5. Hybrid binding: genuine TEE evidence on an authorized host; exact model manifest; proof/attestation/common-input verification; replay and downgrade rejection. Exit requires independent review, not a mocked demo.
  6. Private review packet: short architecture/limitations memo backed by the evidence above. Public release, collaborators and deployments are separate decisions.

Essential ablations

Quantization: float versus integer at the same fixed policy. Length: one chunk versus full max aggregation, long adversarial tails and boundary shifts. Distillation: hard labels alone versus soft teacher outputs, teacher errors and task transfer. Architecture: byte/n-gram floor versus tiny MLP versus one encoder under equal data/selection budgets. Binding: opened-text B1 versus TEE-bound B2, with the different trust assumption written explicitly. Hidden versus public weights: report proof-system cost and disclosure trade-offs without assuming hidden weights prevent adaptive attacks.

Adaptive evaluation and stop rules

Attack with encoding/Unicode transformations, benign quoting/context, role swaps, indirect injection, long prefixes/suffixes, split harmful content, multilingual variants and classifier-aware search. Preserve parent groups and novelty lineage. An adaptive attack set created after reading test outcomes is a new evaluation set, not an unbiased retest. Keep transcript content local and approved.

Stop on evaluator mutation, licence ambiguity, unapproved data/compute, inability to reproduce a selected result, arithmetic mismatch, data contamination or missing transcript coverage. Do not rescue a failing experiment by changing threshold, removing difficult slices or excluding failed proof runs after the fact. If power is insufficient, report inconclusive and propose the additional independently sampled data needed.

Reproducibility standard

Every substantive result identifies source revision plus dirty diff, locked environment and actual versions, dataset/split hashes, policy/threshold, checkpoint and export hashes, seeds, host/resources, full commands, raw counts/costs, exclusions/failures and evaluator revision. Repeat from a fresh directory before a confirmatory claim. A second implementation/reviewer checks critical arithmetic and binding. Agent-generated hypotheses and prose are labeled; measured results require artifacts and replay. No literature result is treated as a local result.