Skip to the content.

TensorBuild deep recheck: measurement contracts worth borrowing

Observed: 2026-08-29
Source: local://tensor-build-main
Snapshot: snapshot-sha256:64986cf8ff942cdcd6178491d3d9af0199c354ce41f998b89ced2b0286f6772d over 6,012 regular files / 216,644,047 bytes
Durable receipt: study_5152edd9388ef5a396f526004fa64ccfd0796171db04810d91a49eed606a17b4
Inventory: local-tensor-build-2026-08-29.json
Prior study: tensor-build-local-study-2026-08-15.md

Verdict first

TensorBuild remains most useful to fak as a source of measurement contracts, not as a TensorRT implementation donor. This recheck found 13 bounded, previously unowned mechanisms worth testing in fak: non-vacuous gate-domain receipts; evidence-lineage independence; cost-normalized evidence acquisition; population-sampling discipline; directional slice floors; addition-versus-replacement benchmark topology; offered-versus-exercised tool census; post-pass GPU-clock coverage; anti-splice vendor evidence; functional kernel-selection identity; claim/null/prior sensitivity arms; evidence-polarity handling for incomplete inputs; and comparison criteria derived from run identity.

Seven other mechanisms already have exact fak owners, two are already present without follow-up, and four are deliberately watched or excluded because they assume TensorRT artifacts that fak’s native execution path does not have. The resulting backlog is issues #10268-#10271 and #10278-#10286. No source code or expressive prose was copied.

Provenance and license fence

The source directory is not a trustworthy Git checkout. Git resolves an unrelated enclosing repository, so that revision is not cited. The declared module path was owner-qualified and private; its remote was unavailable through authenticated repository lookup and anonymous ls-remote; issue, PR, release, discussion, blame, and commit-history claims are therefore unavailable, not empty. No root LICENSE, COPYING, NOTICE, or CLA is present.

Disposition is INSPIRE-ONLY: retain independently specified effects, tests, and operating boundaries; do not transfer implementation or prose. The snapshot digest and per-file SHA-256 anchors below are the only reproducibility identity available.

Completeness and change audit

The machine inventory indexes 5,972 files / 888 directories / 215,914,591 bytes / 1,482,541 text lines after deliberately skipping 40 generated files (729,456 bytes) in internal/build and internal/conn/web/dist. It classifies 2,309 runtime files, 1,619 tests, 825 documentation files, and 16 immediate subsystems.

Two independent deep scopes covered the source:

Scope Denominator Reproducibility pin
Worldview, docs, results, operator contracts 3,174 files / 108,356,033 bytes scope-sha256:76b52a40616ddf26327b51587ec58007e3dba7863f549921eb683064045480fc
go.mod, internal/, and cmd/ runtime/tests 2,353 files scope-sha256:3f37fa97cce240d75d1fbdbd6f529778b4b6318b90d74ad1333dde907f7139b1

All 32 source-file hashes cited by the deep readers were independently recomputed and matched. The prior raw snapshot had 6,358 files / 218,328,563 bytes. Exact inventory comparison found no substantive subsystem change: the current source is missing only 344 transient .dos files (1,681,440 bytes) and two transient .fak files (3,076 bytes). Runtime/test classifier shifts are inventory-tool classification changes, not evidence of source churn.

Architecture and worldview

TensorBuild treats proof shape as part of the product:

  1. A gate reports the population it could have judged, not only its verdict.
  2. Evidence acquisition is separate from later policy adjudication.
  3. Agreement is meaningful only across independent evidence roots.
  4. Benchmark arms must match how a treatment is deployed and must prove the measurement is sensitive enough to say no.
  5. Incomplete inputs affect positive observations and absence claims differently.
  6. Hardware and vendor reports carry coverage and coherence, not just selected numbers.

The source also gives agents a command-oriented front door through AGENTS.md, llms.txt, and one binary surface. That is already consonant with fak and did not justify another API.

Candidate matrix

PRESENT, PARTIAL, and ABSENT describe current fak coverage. Disposition says whether the mechanism belongs in the default product, an optional module, a reusable recipe, a watch list, or the exclusion set.

Mechanism Source anchor Current fak witness Verdict / disposition Owner
Claim-to-proof artifact reachability typed repository graph and reachability audit docs reachability exists; proof-artifact liveness remains open PARTIAL / DEFAULT #6876
Work-class × cost × outcome join token-spend audit exact joined accounting shipped PRESENT / DEFAULT #6874
Scoped negative/inconclusive experiment ledger typed result identity screening-only experiment receipts shipped PRESENT / DEFAULT #6875
Gate domain and admit-rate non-vacuity AGENTS.md:54-74; AD-35:204-230 hook candidates distinguish unreported from zero but omit eligible/admitted distributions PARTIAL / DEFAULT #10268
Persist evidence, reproject policy ground-truth design 234-268 replayable immutable audit evidence already owned PARTIAL / DEFAULT #3978
Derived-lineage independence ground-truth design 272-308 run correlation rejects self-agreement but not shared roots or cycles PARTIAL / DEFAULT #10269
Information gain per acquisition cost ground-truth design 312-365 tool-call control has gain but no hypothesis-bound cost ranking PARTIAL / DEFAULT #10270
Powered population audit AD-35:8-17,92-122,234-245 denominators exist; powered sampling contract does not ABSENT / RECIPE #10271
Directional per-slice uncertainty M9:121-156 slice sample counts and vetoes exist without asymmetric bounds PARTIAL / DEFAULT #10278
Addition versus replacement topology M8:102-120,186-201 next-best alternative exists without treatment topology PARTIAL / DEFAULT #10279
Expected-null walking skeleton plan 301-315 extension promotion already requires it PRESENT / DEFAULT none
Offered versus exercised tool census AD-29:92-117 calls and catalogs are separate; zero rows are lost PARTIAL / RECIPE #10280
Revision-range closeout debt changelog and agent closeout rules exact range-aware owner exists PARTIAL / DEFAULT #5645
Builder-sealed engine inputs internal/enginekey no equivalent TensorRT artifact; native controls and blobs are already bound PARTIAL / EXCLUDE none
Cache hit = deserialize + one token engine cache validation no serialized native engine-plan cache PARTIAL / WATCH none
Dynamic min/opt/max shape profiles engine-key profile tests TensorRT compilation concern, not fak-native execution PARTIAL / EXCLUDE none
Bounded binary-plan introspection engine-plan reader GGUF structure/bounds preflight plus smoke execution exists PRESENT / DEFAULT none
Trained versus produced preprocessing identity vision result contracts exact identity-bound owner exists PARTIAL / DEFAULT #4032
Post-pass GPU clocks and coverage buckets internal/gpu/clocklock.go:110-216 nativeperf environment evidence omits per-pass post clocks PARTIAL / OPTIONAL-MODULE #10281
Anti-splice vendor memory evidence internal/gpu/enginemem.go:440-821 memory provenance lacks cross-section pass coherence PARTIAL / OPTIONAL-MODULE #10282
Functional kernel-selection digest internal/tactic/protocol.go:316-369 controls exist without one mutation-proven selector identity PARTIAL / DEFAULT #10283
Claim/null/prior sensitivity triad internal/tactic/remeasure.go:378-421 nativeperf has control/candidate but no mandatory drifting null PARTIAL / DEFAULT #10284
Evidence-polarity incomplete input internal/fusion/fusion.go:451 incomplete evidence cannot pass, but positive-preserve/absence-withdraw is not typed PARTIAL / DEFAULT #10285
Byte, numeric, and non-finite equality export/golden comparator tests exact non-finite-safe comparison owner exists PARTIAL / DEFAULT #8635
Comparator criterion from run identity internal/export/parity.go:479-552,774-786 criterion and identity are recorded but independently supplied PARTIAL / DEFAULT #10286
Named order-independent multi-output comparison export parity native forward currently exposes one activation/logit output surface PARTIAL / WATCH none

Source anchors

These anchors make the conceptual transfer falsifiable without reproducing source text:

Path SHA-256
AGENTS.md 3f399b6f7f51e57b5c8ae14331ea3697b3be4cf9fe77201d80bc413775a5ee4d
docs/design/DESIGN-agentic-ground-truth-engine.md f2bb5967679b24ecad2b955717101669e9d6f192c03e5f351fa40aeb54cfb847
results/agentic-data/AD-35-the-price-of-agreement.md b69f2d2af35b92a1656b8c59ac1d4e09545a19e6abf110eaae72e6c508f8eea3
results/agentic-data/AD-29-bringup-turns-tools-and-the-conn-loop.md 7c046806607637da658c2e3afe859b0ef7a1a522387b052a538a089cd402127a
results/measurement/M8-baseline-must-be-the-next-best-alternative.md 8e245018067f76d0547cd4ab1533838884c05ff33592141a8ce5a42c2b8e2007
results/measurement/M9-the-agentic-lift-blends-and-the-human-one-does-not.md b88ce7684b3d388f755cac3328f09aca588d1af5d041ca3fa87a884512ae36b4
internal/gpu/clocklock.go f950841c45569c8b303a48ac70decf1a361dd406c9dc219355692731483f194d
internal/gpu/enginemem.go 857eaf16d921befcadf8b378dda0ba4b8b902ad643535179f59a33a052ba2042
internal/tactic/protocol.go cc21bcf670e544b60baa908813d3a2c35b2b14b825b6e63295bcc2841ffcf860
internal/tactic/remeasure.go 2e07c8752ea55a935add40e376c48e309a02535dd38984b9839605ec88c3a08f
internal/fusion/fusion.go bbce0ec9aac19b8c5b0333c33dc18dccd2d4cbdf119ff3cd1038c4a8051868eb
internal/fusion/verdictpolarity_test.go 1dd540c3697f19d74f24e57fb2a62be871ddf6eeaca49ddc0a622109ca73f2f3
internal/export/parity.go c29297bb051329987f24e321017289a1fcba67eff6dbbea900b958d3d26bd885

Ablation and negative knowledge

The TensorRT-specific candidates were not rejected because they are weak implementations. They were rejected because their useful effect depends on a compiled engine-plan abstraction that fak-native inference does not own. Adding builder-shaped keys, min/opt/max optimization profiles, or deserialize-and-one-token cache validation now would create a false parallel architecture.

Similarly, named multi-output comparison remains a watch item until fak-native forward exposes multiple named outputs. The correct near-term borrow is comparison identity and non-finite-safe semantics on the output surface fak actually has.

The native-performance leaves preserve the repository invariant: execution stays fak-native and the live path prefers Qwen3.8. Hardware witnesses must run on sanctioned lab nodes and name the engine in the receipt; none may introduce a silent llama.cpp fallback.

Best-default and coverage frontier

Axis Best default now Bounded superset
Gate health domain plus decision counts constant-decision exemptions only with typed rationale
Evidence quorum independent root lineage richer evidence classes after a real consumer needs them
Tool acquisition raw gain, cost, hypotheses, and selected/next-best learned estimates only after calibrated receipts exist
Benchmark admission matched topology, directional floors, sensitivity control hardware-specific probes as optional modules
Partial evidence polarity plus covered source range domain-specific parsers behind the common contract
Native identity decoded canonical selector record plus digest multi-output identity when the runtime exposes it

Problem and value check

For fak maintainers and agents; Problem: the earlier study captured only three broad operational mechanisms; Today: source-level measurement contracts and their native-runtime applicability were not deeply reconciled; Better because: the complete source denominator, license fence, ablation decisions, exact FAK witnesses, and issue ownership are durable; Witness: the inventory, immutable study receipt, 32/32 source-hash read-back, and issue set #10268-#10271 / #10278-#10286.

Independent witness and remaining uncertainty

Worker summaries were treated as claims. The coordinator recomputed every cited source hash, read the relevant FAK seams, reran exact open-and-closed issue searches for all 13 new titles, and read back the created issues. The local source lacks managed worker transcript paths, so no transcript-level DOS verification was possible; source artifacts, FAK files, GitHub objects, and the durable study record are the independent evidence surfaces instead.

Refresh when a trustworthy source revision or license appears, when any watched TensorRT-shaped abstraction gains a real fak-native counterpart, or when one of the filed leaves changes the on-axis evidence. Do not reinterpret absent upstream metadata as evidence that no upstream work exists.