Skip to content

Latest commit

 

History

History
206 lines (177 loc) · 13 KB

PROJECT_STATUS.md

File metadata and controls

206 lines (177 loc) · 13 KB

ABI project status

Status date: 2026-11-14

Active work

R108 ended incomplete at its fixed 25-minute control budget, under reported GPU power/thermal throttling (requested 10 W versus 80 W default, 210 MHz). It preserves 13/96 initial responses, no replay, no recipients and zero training. Matched R99 LoRA/distillation checkpoints remain unchanged. The complete verifier rejects qualification. This is a resource prerequisite, not a measured acquisition-quality failure or native LayerCake defect. Its immutable research tag and public reconstruction receipt are published: 565 assets and 182 host bindings match; 49 focused checks pass. The complete scientific verifier still rejects the interrupted attempt.

R107 completed 18 no-training executions with exact raw replay. The primary residual-neutralization contrast supports modest parent interference on natural text (+0.03224 BPB, 95% CI [+0.02064, +0.04419]), but neutralization increases parent repetition. The fixed coherence route was untrained in R99; distillation's matching parent behavior there does not prove trained-route fluency or ABI superiority. All quality/product gates remain unchanged. This is an acquisition-lineage diagnostic, not a native host defect. Its public reconstruction matches 585 assets and passes 55 focused checks; tagged CI passes. The scientific tag is immutable. Same-machine/reused-environment limits remain explicit.

R106 completed at its narrow grammar-control stop: 178/200 versus 190/200, exact live raw replay. No recipients or training ran. This supplied-token likelihood diagnostic is distinct from autonomous generation; its negative result and external-label limitations do not promote an English core or establish a native LayerCake defect. The full acquisition breakthrough remains open. Its research tag and reconstruction receipt are public: 581 assets match and 52 focused checks pass. Tagged CI is green; same-machine/cache limitations remain.

R105 rejected one frozen conventional NLI evaluator: 779/1,200 external scope cases and 62/64 approval cases, below both registered gates. Initial/replay raw logits are byte-identical. Full-study verification requires both executions after a preserved missing-replay hostile-audit finding. No teacher/recipient/training calls occurred. English-control qualification and the compact-core breakthrough remain open; native LayerCake is unchanged. The immutable research prerelease is public. Same-machine public reconstruction matches 40 evidence assets and eight evaluator files, recomputes the rejection and passes 41 focused checks. Tagged CI passes; independent review remains open.

R104 completed with 155 finite evaluator cases passing but raw control qualification failing at 29/48. Four task floors fail; two additional approval-scope limitations are disclosed. Even removing all approval flags yields at most 30/48. The legacy score of 44/48 also misses the clarification floor. No recipients or training ran. The requested qualified control and recipient comparison remain unmet; native LayerCake is unchanged. The weight-space note keeps competing acquisition methods in view without claiming superiority or adding a new host contract. The sealed R104 release is public. Same-machine public reconstruction matches 566 assets and 729 bindings and passes 43 focused checks; tagged GitHub CI passes. Publication does not change the negative scientific verdict.

R103 stopped at raw control qualification: 41/48 against 44/48, with three task-floor failures. No recipient comparison or training ran. An observed temporal false-positive in the locked contradiction proxy is documented, not silently repaired. Control/evaluator validity must be established before drawing the intended recipient/interface conclusions.

The published R102 study completed and strictly recomputed its matched factorial. Neither registered factor met its support criterion and no checkpoint passed quality. Repeat-data LoRA scored 51 -> 46/96 and adapters 32 -> 39/96; renamed data scored 14 -> 34 and 10 -> 24. The source, eight states, raw generations, failed preflight and descriptive error audit are preserved. No native LayerCake defect or new English core is established. Its immutable research tag and complete required payload set are public. Fresh public-only clones matched 782 assets and 833 bindings; strict recomputation and 42 focused tests passed. The separate receipt records the same-machine scope. Current CI passes after the preserved dependency-only repair.

R101 is complete and strictly recomputed: sampled full/padded/cached execution agreed across four checkpoints, and all 288 unassisted sequences replayed exactly. Its primary oracle-prefix remaining-fact contrast was null (24/34 before and after; 95% CI [-0.14706, 0.14706]). The full-model reference-token fit was 91.44% on training and 78.91% on development. This narrows the next question toward acquisition fit and generalization without proving data insufficiency or a native LayerCake defect.

The additive R99-D1 fit audit completed without changing any model: training-example passes were 7/24 ABI plan, 12/24 distillation, 13/24 LoRA and 0/24 parent. Correct-prefix fact-token accuracy was high but visible-response fit remained limited. The next bounded R100 control changed only optimizer parameter scope to all model parameters, with new development prompts. It completed at 20/48, equal to retained LoRA; adapter distillation reached 14/48, ABI plan 6/48 and parent 0/48. The required score stayed 44/48. Its training fit was 17/24 versus a required 22/24. All registered screens failed. No ABI product, host regression, or simple parameter-scope explanation is established.

R98 and its completion follow-up are preserved source failures: 206/240 and 226/240 mechanical passes, respectively. Neither trained a student. C1 removed truncation but still missed the email family gate.

R99 introduces source-substring sentence selection and a separately frozen prospective development population. Normalized source qualification passed 230/240 (30/30 emails). The first learner then stopped on loss-scale overflow. A bounded numerical audit isolated that failure. The matched FP32 comparison then completed: ABI plan 10/48, distillation 20/48, LoRA 27/48, parent 0/48, versus the required 44/48. No learner passed; the autoregressive-plan branch is closed. The entire 87,710,428-parameter parent is counted; no earlier LayerCake performance certificate is inherited.

Verdict

FULL_MOONSHOT_NOT_PROVEN

The R101 diagnostic review is versioned as abi-research-r101-2026-11-05. Its local review passed 262 tests with four documented skips; the supported CI surface passed 83 and package builds passed. Use the R101 review packet for its exact claim limits. Public-only checkouts matched 35 direct assets and 813 frozen bindings, strictly recomputed the result and passed 43 focused tests. The receipt discloses reused public directories and same-machine evidence reconstruction; no independent hardware or new live inference run is claimed.

The preceding R100 bounded research snapshot is published as abi-research-r100-2026-11-05. Its local review passed 236 tests with four documented skips; the supported 57-test CI surface, package builds and documentation checks passed. See the R100 review packet. Public-only checkouts then matched 52 assets and 784 frozen inputs, recomputed both negative decisions and passed 62 focused tests; see the bounded reconstruction receipt. No live generation or different-hardware replication is claimed.

The preceding R99 snapshot is published as abi-research-2026-11-05. Its local review suites passed 219 tests with four documented skips; package builds, documentation checks and GitHub CI passed. Fresh public clones then matched all 738 frozen inputs and 50 required assets, recomputed the negative result, and passed 45 focused/adversarial tests. See the reconstruction receipt. This same-machine evidence replay is not a scientific promotion, fresh model training, live generation replication, or independent-hardware validation.

Evidence dashboard

Surface Status Claim ceiling
R7 public release Publicly reconstructed bounded technical pass Named capability runtime/conformance environments only
V1089 Local machine evidence through Phase 7; same-machine Phase 8 rehearsal Separate lineage; not a public final-artifact certificate
Phase 2 human ratings 0/21,000 Workflow prepared; no human-quality result
Phase 5 clean replay Blocked on six disclosed deleted baseline tensors Historical bounded certificate retained; clean replay not complete
R97 1,399/1,400 prospective two-hop reasoning; strict, causal, hostile, and byte-replay checks pass One bounded reasoning family and compatible LayerCake parent
R21 Public prerequisite pass; hidden summary gate failed Not promoted to general generation
R22 Source normalization prerequisite failed No student or LayerCake run
R23 Prospective bounded semantic supplied-content pass Not broad English or autonomous discovery
R98 / C1 Source gates failed; immutable negative evidence No student quality evaluation
R99 Normalized data qualified; matched FP32 learner comparison failed ABI plan underperformed both trained controls; no fluent core promoted
R99-D1 / R100 Limited autonomous training fit; full-model control 20/48 on fresh prompts, equal to LoRA Parameter unfreezing alone did not solve this bounded recipe; cause remains unresolved
R101 Measured execution-path agreement and exact unassisted replay; null primary continuation contrast No autonomous gain, native host defect, or sole acquisition cause established
R102 (public research) Completed matched renaming/exposure factorial; all quality screens failed No further same-recipe sweep or product promotion; recipient/interface suitability remains open
R103 (public prerequisite) Control qualification failed at 41/48; no recipient or training run Validate evaluator boundaries and qualify a control before recipient attribution
Independent hardware Not executed Open
Minimum-information proof Not executed Open
General teacher-to-LayerCake English extraction Not established Open
General LoRA/distillation superiority Not established Open

What a reviewer can evaluate now

  • exact public R7 reconstruction and its bounded hostile controls;
  • the repaired supported replay surface for retained V1089 and R97 evidence;
  • R97's prospective causal reasoning-transfer result as its own lineage;
  • R21-R23's disclosed passes and failures as separate bounded experiments;
  • the frozen 21,000-row human workflow, which remains entirely unrated; and
  • whether missing inputs fail closed rather than being replaced by trusted status booleans.

Remaining decisive gates

  1. Complete 21,000 independently attested blinded human preferences.
  2. Reproduce the eligible immutable release on genuinely independent CPU and CUDA hardware.
  3. Resolve the Phase 5 replay deficit prospectively or restore the six exact content-addressed tensors.
  4. Produce one coherent final artifact instead of composing evidence from R7, V1089, and R97.
  5. Prove broad teacher-origin English acquisition, automatic discovery and segregation, and matched comparisons with LoRA/distillation.
  6. Execute the registered smallest-sufficient-information program.

Review identity

The current diagnostic/control review tag is abi-research-r100-2026-11-05. Use the R100 result and reproduction commands and its hash-addressed public manifest; this tag certifies no product.

The preceding bounded English review tag is abi-research-2026-11-05, peeled commit e230fe298d46db2894680b6d518fa0bf98751c41. Follow the R99 reproduction packet, fetch its required Git LFS objects, and report missing assets as failures. Its expected verdict is ABI_ACQUISITION_PILOT_FAILED.

The historical documentation/binding-repair tag abi-peer-review-candidate-v3-2026-09-18 remains unchanged. Its independent review handoff and lineage-specific LayerCake revisions do not become R99 evidence merely because they share this repository.

For full detail, continue with CLAIMS.md, research status, and the independent-review handoff.