Short identifiers and status words used across OpenGrad, each with where it is defined. When a document and this page disagree, the linked definition wins and this page is the bug.
| Term | Meaning | Defined in |
|---|---|---|
| B0 | The base model, Qwen/Qwen3.5-2B at revision 15852e8c…, evaluated unchanged. The baseline every stage is compared with. |
README.md, registry/models.yaml |
| M0 | Study 001's supervised fine-tune (SFT) of B0 on Canonical-v2 final (checkpoint 1800). Not promoted. | docs/EXPERIMENT_RESULTS.md |
| M1-v1 | DPO applied directly to B0, with no SFT parent. Its best checkpoints no longer exist. | docs/EXPERIMENT_RESULTS.md |
| M1-v2 | DPO calibration started from M0 (checkpoint 30). The model promoted under tool_use_promotion_v4. |
reports/M1_DPO_EVALUATION.md |
| M2 | On-policy distillation. Never executed (mock-only scaffold); deferred to Study 003. | docs/ON_POLICY_DISTILLATION.md |
| PTQ | Post-training quantization (the GGUF ladder, closed). | reports/PTQ_PHASE_CLOSURE.md |
| QAD | Quantization-aware distillation; decided not required for release. | reports/QAD_DECISION.md |
| MTP | The base model's native multi-token-prediction head, carried or excluded by trainers per policy. | docs/MODEL_COMPONENT_POLICY.md |
A tool-use request has one correct behaviour, its mode (also "decision"): CALL (call a tool),
ANSWER / DIRECT (answer directly, without a tool; DIRECT is the classifier's name for it),
CLARIFY (ask for missing information) and UNSUPPORTED (decline: no tool or knowledge can serve it).
Study 001's held-out set had no ANSWER items, which is why its regression went unseen.
"C1" and "C2" each have three meanings in Study 002. Always qualify them.
| Term | Meaning | Defined in |
|---|---|---|
| arm C0, C2, R1, … | Training arms of the experiment matrix. C0 reproduces Canonical-v2. Arm C1 ("without the synthetic subset") and arm R2 were dropped by study_002_prereg_v14, and their ids are retired. |
04-ARM-MATRIX |
| criterion C1 (coverage) | Every mode in P-CONF has n > 0. |
06-SPLIT-SPEC §C1 |
| criterion C2 (resolvability) | Every row prints n and its resolvable margin; no claim on a row that cannot resolve 10pp. | 06-SPLIT-SPEC §C2 |
| workstream C1 | Building canonical-v3 (provenance repair, schema normalization, classifier labels, decision balance). "C1 is authorised" means this. | 21-C1-IMPLEMENTATION-STATUS |
| L1 … L12 | Lessons from Study 001; L1 is "the held-out set could not see the failure" (no ANSWER items). | 01-LESSONS |
| H1 … H6 | Study 002's hypotheses; H1: refusal-ANSWER label disagreement is enough to cause the collapse. |
02-RESEARCH-QUESTIONS |
| P-DEV, P-CONF, P-UNANS, P-SEALED | Populations: development (checkpoint selection), confirmatory (scored once per arm; the gate reads it), genuinely unanswerable (refusal correctness, H6), never-read reserve. | 06-SPLIT-SPEC §Populations |
| P-DET, P-DET-COVERAGE | Labelled samples for validating the refusal/decision classifier. P-DET-v1 is frozen (581 items); P-DET-COVERAGE-v1/v2 add DIRECT and CALL coverage. | 22-PDET-PROTOCOL, 30, 36 |
| REP-A | A re-run of arm C0 at a fixed seed: the nondeterminism floor every threshold margin is compared with. | 05-SEED-AND-REPRODUCIBILITY-POLICY |
| V1 … V12 | The provenance validators (mode coverage, metric denominators, …, resolvable margin). | 15-PROVENANCE-VALIDATORS |
study_002_prereg_vN |
A preregistration version; v1 is the original contract, v2–v15 are amendments (v15 adopted 2026-10-03). | 03-PREREGISTRATION |
study_002_gate_v1 |
The Study 002 evaluation gate (18 checks at contract 7 since 2026-10-13, ERRATA §35, §37, §38 and §39; over tool_use_promotion_v6). |
11-THRESHOLDS |
| G1 … G17 | Guardrails derived from the claims Study 001 got wrong. | GUARDRAILS |
| Word | Meaning |
|---|---|
| frozen | The bytes are fixed and recorded by hash; the artifact is never edited. A number later found wrong is corrected in reports/ERRATA.md. A frozen study is fixed at a git tag. |
| pinned | A file whose sha256 is recorded somewhere that a check reads (a .sha256 sidecar, a manifest, a code constant). Pinned files are frozen. |
| verified | A claim whose evidence was checked against an immutable anchor: a pinned file's hash, or an external source at an exact revision. verified: true without the anchor is a defect. |
| provisional | Accepted for now, with stated conditions that withdraw it (e.g. a qualification resting on a model-made reference). |
| gold | A reference label frozen as the answer key. Model-made labels are a model reference, never gold; no Study 002 label is gold yet. |
| WIP export | A working annotation snapshot (annotation/wip/), not a gold freeze. |
| canonical | The normalized, source-independent record format (the "canonical IR"), and the corpora built in it (Canonical-v1, Canonical-v2 final, canonical-v3). |
| promoted | A checkpoint that passed the promotion policy version pinned before its evaluation. It says which policy, not that the model is better overall. |
| PASS / FAIL / BLOCKED_* | Gate outcomes. BLOCKED_* means a required input was absent and the gate could not check; it is never a pass. |
| NOT_EVALUABLE / UNDER_POWERED / WITHIN_NOISE | The population cannot measure a dimension / a mode is below n = 200 / a difference is smaller than the row can resolve. None supports a claim. |