Transferable techniques

franken_engine (the registry refuses wording whose actual_wording_state exceeds allowed_state, and emits the exact downgrade text) [Verified, High]; frankengit (SHA-256-bound claims: any mismatch is an automatic demotion, not a reviewer-overridable transition) [Verified, High]; franken_markdown (claims.tsv plus check-claim-discipline.sh) [Verified, High]; franken_node (Ed25519-signed honesty manifest) [Verified, High].

Any project whose README outruns its CI: franken_manim, franken_ocr, frankenpandas, frankenscipy, franken_whisper, and all four website repos [Inference, High].

Marketing drift becomes a CI failure instead of a review-cycle argument. Claim downgrade is automatic and timestamped [Inference, High].

Point franken_engine's matrix (or a minimal reimplementation) at one sibling repo for 60 days: it either blocks a real claim promotion (demonstrated value) or proves too coupled to generalize (falsified portability) [Inference, High].

franken_engine: every claim carries max_observed_freshness_days plus a repro.lock; the gate auto-downgrades rows whose re-verification lags [Verified, High].

franken_networkx (CI green 13 days pre-pin at last check), franken_nlp (stale DSR evidence), frankenredis (fmt-gate failure versus green prose), and any "last verified" statement in the corpus [Inference, High].

"Verified" claims cannot silently age into stale ones. Staleness is typed and visible [Inference, High].

Add a freshness column to one sibling's claims registry with a 30-day ceiling: if it surfaces at least one claim the project believed was current but wasn't, the gate has teeth; if zero surface and a blind audit confirms the flagged claims were already fresh, the ceiling adds bookkeeping without catching drift, and the gate is falsified as overhead [Inference, High].

frankengit (34-row registry, 13 required entry classes including overclaim_correction and non_reproducible_result) [Verified, High]; franken_numpy (67,641-line ledger, the largest counted) [Verified, High]; frankensqlite (perf-negative-results.md with retry conditions) [Verified, High]; franken_tts (NE-002/NE-003 killing the flagship optimizer twice) [Verified, High].

franken_whisper, franken_overlap, frankenterm, frankensympy: all do negative evidence informally [Inference, High].

Failed optimizations stop being re-run by future agents. Reviewers can see what was tried and killed [Inference, High].

Require one sibling project to log its next 10 rejected experiments with retry conditions: either a future session avoids re-running at least one of them (value), or the ledger is never consulted (falsified as ritual) [Inference, High].

franken_overlap: a baseline that exceeds the compute budget is recorded as incomplete, never extrapolated into a comparison [Verified, High].

franken_ocr, franken_tts, franken_whisper: anywhere an official incumbent cannot be run [Inference, High].

"N/A" rows stay honest. Benchmark tables carry no ratio computed against a non-run baseline [Inference, High].

Require one sibling to adopt the rule for 90 days: if the rule blocks at least one extrapolated ratio, it caught a live hole; if zero blocks occur and an independent audit finds at least one extrapolated ratio the rule missed, the rule is decorative, and is falsified [Inference, High].

asupersync: methodology_baselines signoff plus a 105-row artifacts/baseline.json; the gate demonstrably fired on 2026-09-07 against drifted artifacts [CI-observed, High].

frankenredis, frankenpandas, frankenscipy [Inference, High].

Benchmark regression as a signed, waiver-free gate: per-row environment tagging, layout-hyperparameter capture, and drift detection that demonstrably fires. Environment drift (new runner, new toolchain) cannot silently rebase "improvements" [Inference, High].

Replay one sibling's last benchmark comparison with environment tagging added: if any historical "win" reclassifies as environment drift, the gate caught a live confound; if zero reclassify, run a control: inject a known environment drift (runner swap, toolchain bump). If the gate does not fire on the injected drift, the drift detection is decorative, and is falsified as a control [Inference, High].

frankenredis: the program's optimization-evidence template. Every accepted optimization ships a bundle: SHA-pinned golden transcript plus paired/reversed runs plus a keep/reject score gate [Verified, High].

frankenjax (whose per-optimization entries with agent attribution and same-worker/same-binary criterion rows already exist; the transfer is the bundle format: a SHA-pinned golden transcript as a replayable artifact, the paired/reversed run design, and the keep/reject score gate packaged as one reviewable unit) [Verified, High]; plus frankenredis's own future work, frankentorch, frankensqlite [Inference, High].

Optimizations become reviewable artifacts with kill criteria, not commit messages [Inference, High].

Require one sibling's next 5 accepted optimizations to ship bundles: either at least one optimization is rejected at the bundle stage (gate value), or all pass and a blind re-review agrees (no gate value: falsified as a control) [Inference, High].

frankentorch: name the worker, name the harness, quote both estimators, replicate before quoting [Inference, High]. Partial overlap: frankenpandas already enforces the cross-worker rule ("A/B comparisons across different workers are invalid") via its quantified build-variance rule [Verified, High]; the transfer adds the remaining three rules.

Every benchmarking project in the corpus: frankenredis, frankenpandas, franken_whisper, franken_ocr, franken_tts, frankenscipy [Inference, High].

Numbers arrive with their full provenance or don't arrive at all. Cross-worker comparisons die on sight [Inference, High].

Apply the four rules retroactively to one sibling's published ratios: if at least one ratio loses a rule's provenance and re-grades to [NO ADMISSIBLE RATIO], the gate has teeth; if zero lose provenance and a spot re-measurement under the rules confirms the ratios, the project was already compliant: the gate adds nothing there, and is falsified [Inference, High].

franken_whisper (PERF_LEDGER) [Verified, High]; franken_networkx (INCUMBENT-vs-SELF-SPEEDUP contract) [Verified, High]. Typed verdicts: SELF-SPEEDUP is maintenance, never a campaign win; a CAMPAIGN WIN requires the actual incumbent binary side-by-side in the same invocation, with A/A nulls in [0.98, 1.02].

franken_ocr, franken_tts, franken_nlp, frankenredis [Inference, High].

"Faster" claims carry their class. Self-speedups cannot be promoted into campaign wins [Inference, High].

Reclassify one sibling's headline ratios under the doctrine: if at least one claimed win demotes to SELF-SPEEDUP, the doctrine caught promotion drift; if zero demote, the doctrine is a re-labeling exercise for that project, and is falsified as a new control [Inference, High].

franken_snowflake (credential-less runs emit franken_snowflake.live_gate.v1 skip events) [Verified, High]; frankenscipy (CI control job fails if the live SciPy oracle is absent) [Verified, High]; franken_tts (XFAIL is not SKIP; GREEN WITH SKIPS is a distinct state) [Verified, High]. Runs that cannot execute honestly emit a typed skip event instead of silently passing; the absence of the oracle is itself a gate failure.

Every project with optional live dependencies: franken_nlp (HuggingFace access), franken_ocr (model weights), frankensim's DSR lanes [Inference, High].

"Green with skips" is a distinguishable state from "green". Skip events are distinguishable from passes in CI output and release notes [Inference, High].

Convert one sibling's silent skips to typed skip events for one release cycle: if any event reveals a lane that never ran in production, the honesty was load-bearing; if all lanes ran and the project's docs already distinguished skips from passes, the conversion adds ceremony without information, and is falsified [Inference, High].

franken_drone_geometry_reconstruction (five-label vocabulary: source-present, reference-implemented, publicly-invokable, locally-qualified, production-admitted; "no lower label implies a higher one") [Verified, High]; frankensim (15-entry L1–L5 registry: L1=3, L2=11, L3=1, L4=0, L5=0) [Maintainer claim, High].

franken_remote, franken_surveillance_system, franken_native_capsule: capability-rich, validation-poor projects [Inference, High].

"It exists" stops meaning "it works". The maturity distribution itself is publishable [Inference, Medium].

Label one sibling's capabilities under the vocabulary: if any capability the README treats as mature lands below "publicly-invokable," the vocabulary caught drift; if none do, the vocabulary is a re-labeling of an already-calibrated README, and is falsified as a new control [Inference, High].

frankensim (fs-evidence: "evidence color" and "no-launder" rules found in no other packet) [Maintainer claim, High]. Measurement evidence carries typed colors; composition rules prevent weak evidence from being laundered into strong claims.

frankenpandas (359-lane aggregation), franken_networkx (multi-ledger verdicts), frankenterm (attestation bundles) [Inference, High].

Aggregated verdicts cannot inherit the strongest color of their components. The weakest link is visible [Inference, Medium].

Re-color at least three of one sibling's aggregate verdicts under the algebra: if any aggregate's color weakens, the algebra caught laundering; if all stay unchanged, the algebra restates existing practice, and is falsified as a new control [Inference, High].

franken_alignment (dispatch plus receipts plus a self-disqualification rule: code landing after the last executed gate opens a self-disqualification ledger) [Verified, High].

frankenjax: the genuine RCH-dependent case (zero GitHub Actions workflows; all green claims execute on the maintainer's private RCH fleet, attested only by checked-in artifact JSON) [Verified, High]. franken_node: with the distinction that node already has a hash-chained decision-receipt lineage, so the transfer is the self-disqualification rule for post-gate code and retention of failed attempts, not receipts per se [Inference, High].

"It ran on the private runner" becomes a checkable receipt instead of a prose claim. Post-gate code is explicitly disqualified, not silently trusted [Inference, High].

Require frankenjax to retain and publish its next 10 batch receipts with source-hash binding: if any receipt contradicts the prose claim it supports, the binding caught drift; if all agree, trust is receipted but untested; and if a receipt can be regenerated post-hoc with altered hashes without detection, the binding is decorative, and is falsified [Inference, High].

frankenpandas: 359 lanes aggregate into typed verdicts (FASTER / SLOWER / NULL_UNDECIDABLE / DROPPED_HIGH_CV / PARITY) as 201 / 20 / 74 / 62 / 2 [Verified, High]; one row retracts its own 1.20x win as build variance, in the same row [Verified, High].

frankenredis, frankenscipy, franken_whisper [Inference, High].

"201 faster" cannot be quoted without "62 dropped for high CV". The funnel is the number [Inference, High].

Re-aggregate one sibling's benchmark history as a full funnel: if the decidable-subset statistic changes the headline, the funnel was load-bearing; if the headline is unchanged and the project already disclosed its dropped lanes, the funnel restates existing disclosure, and is falsified as a new control [Inference, High].

franken_alignment (planted-mutation testing: deliberately introduced defects must fail after controls pass) [Verified, High]; frankensympy (caught and remediated its own fails-open harness) [Verified, High]; franken_code_browser (six named defects: demo hardcoding, proof laundering, refusal farming, commit pumping, gate self-weakening, follow-up laundering) [Verified, High]. The evidence gate itself gets audited adversarially: planted defects, fails-open harness hunts, named failure modes of honesty culture.

Every project with a claims gate: franken_engine, frankengit, franken_networkx, frankensim [Inference, High].

Gates get tested against adversaries, not just against the happy path. A gate that cannot catch a planted defect is known to be decorative [Inference, High].

Plant 3 defects in one sibling's gate (a laundered claim, a weakened gate, a pumped commit) and see if the gate catches them: caught 2 or more means the gate is adversarially sound; caught 1 means partial soundness, inconclusive, re-run with a fresh defect set before judging; caught 0 means the gate is theater, and is falsified [Inference, High].