Issues

• Serving host: rebuild the served model's trained manifest with --reworded

  owner: devP3type/bug
  jonpol01/d1a#379 · Filter by author jonpol01 opened on Oct 31, 2026 · Accuracy and robustness

• Data blocklist: a listed name that starts a longer identifier is missed

  owner: devP2review:needs-humantype/security
  jonpol01/d1a#378 · Filter by author jonpol01 opened on Oct 31, 2026 · D1A core (refactor and future-proofing)

• D1A-E2B-it QAT v0.7: PyTorch, Core ML and the intents suite

  owner: devP3type/feature
  jonpol01/d1a#372 · Filter by author jonpol01 opened on Oct 30, 2026 · On-device

• D1A-E4B-it QAT: typesafe-v1 −7.8 and transfer-v4 calibration against v0.6

  owner: devP3type/test
  jonpol01/d1a#361 · Filter by author jonpol01 opened on Oct 28, 2026 · Accuracy and robustness
  4

• On-device: D1A-E2B-it on Gemma 4 QAT weights (MLX int4 with adapter, Core ML generating 4-bit)

  owner: devP3type/bug
  jonpol01/d1a#341 · Filter by author jonpol01 opened on Oct 24, 2026 · On-device
  1

• DGX Spark readiness: D1A in production on Linux aarch64 + Blackwell from next week

  P3type/test
  jonpol01/d1a#317 · Filter by author jonpol01 opened on Oct 22, 2026 · Portable foundation

• Self-learning tiers: learning from use without a strong machine

  owner: devP3type/feature
  jonpol01/d1a#299 · Filter by author jonpol01 opened on Oct 19, 2026 · Accuracy and robustness

• Shadow-serve the rate-tracking severe flag, then a blind confirmation (#237 next step)

  P3type/feature
  jonpol01/d1a#296 · Filter by author jonpol01 opened on Oct 18, 2026 · Accuracy and robustness
  5

• D1A-12B (QAT 4-bit, one model on the serving Mac)

  P3type/feature
  jonpol01/d1a#270 · Filter by author jonpol01 opened on Oct 13, 2026
  12

• Pre-registered: does the Grok review bot label like the hermes-agent maintainers?

  P3type/test
  jonpol01/d1a#252 · Filter by author jonpol01 opened on Oct 11, 2026 · Accuracy and robustness
  18

• Self-learning: severity flag (learned decision threshold), offline-fitted, served beside the answer: E4B 15/30 → 25/30 P0/P1 at 14.6% flagged

  owner: devP3type/feature
  jonpol01/d1a#237 · Filter by author jonpol01 opened on Oct 10, 2026 · Accuracy and robustness
  7

• Core ML pilot: D1A E2B on the Apple Neural Engine (convert, verify flips, benchmark on the serving Mac)

  P3type/perf
  jonpol01/d1a#224 · Filter by author jonpol01 opened on Oct 9, 2026 · On-device
  10

• Release the SWE verifier's adapter and results (on the maintainer's public go)

  P3review:needs-humantype/security
  jonpol01/d1a#223 · Filter by author jonpol01 opened on Oct 9, 2026 · Hybrid recipe (Kev skills + D1A's own data)
  2

• Private eval mirror returns 404: the OOD, breadth, tasksource and longdoc suites cannot be scored or screened

  P3type/test
  jonpol01/d1a#200 · Filter by author jonpol01 opened on Oct 7, 2026 · Accuracy and robustness
  1

• Track E: new training data for the uncovered use cases (prompt injection, asking for missing information, evals)

  P3review:needs-humantype/docs
  jonpol01/d1a#199 · Filter by author jonpol01 opened on Oct 7, 2026 · Accuracy and robustness

• Severity offsets for the PR labeler: v0.5's P1 drop is a prior shift (sev 75.6 → ~78.6, P1 3 → ~18/47, no training)

  P3type/bug
  jonpol01/d1a#198 · Filter by author jonpol01 opened on Oct 7, 2026 · Accuracy and robustness
  14

• Borrow Kev's newer eval suites (decision-v8, transfer-v6, public-pool-v6) as eval-only, pending licensing

  P3type/test
  jonpol01/d1a#190 · Filter by author jonpol01 opened on Oct 7, 2026 · Accuracy and robustness

• Port three Kev changes after 2026-09-30: MLX long-state prefill, torch long rows, held-out temperature refit

  P3type/feature
  jonpol01/d1a#189 · Filter by author jonpol01 opened on Oct 7, 2026 · Accuracy and robustness
  3

• D1A's own architecture: a decision model designed from scratch (parked, not planned)

  enhancementowner: leadP3type/feature
  jonpol01/d1a#174 · Filter by author jonpol01 opened on Oct 6, 2026 · D1A's own architecture (not planned)
  1

• Outcome memory: pre-registered test on the live PR-labeler stream at ≥400 new outcomes

  owner: devowner: leadP3type/feature
  jonpol01/d1a#149 · Filter by author jonpol01 opened on Oct 5, 2026 · Accuracy and robustness
  16

• Routing learns from outcomes: feed routed-task results into d1a.feedback

  owner: devowner: leadP3type/bug
  jonpol01/d1a#128 · Filter by author jonpol01 opened on Oct 5, 2026 · Model routing
  7

• Anytime verifier r2: train on partial runs for stop-and-restart under a budget

  owner: leadP3review:needs-humantype/test
  jonpol01/d1a#120 · Filter by author jonpol01 opened on Oct 5, 2026 · Agent verifier (Gemma 4 paper)

• LATER: Live captions + speaker tags + D1A line reads (anime-style scenes, overlapping speech)

  enhancementP3type/feature
  jonpol01/d1a#82 · Filter by author jonpol01 opened on Oct 2, 2026 · Vision and voice

• Multimodal decision data + training stage

  enhancementP3review:needs-humantype/feature
  jonpol01/d1a#80 · Filter by author jonpol01 opened on Oct 2, 2026 · Vision and voice

• Zero-shot feasibility: D1A questions over an image or audio state

  enhancementP3review:needs-humantype/feature
  jonpol01/d1a#78 · Filter by author jonpol01 opened on Oct 2, 2026 · Vision and voice
  1