vernier eval --emit json output schema

The JSON formatter shipped by vernier-cli (ADR-0015) writes a single document per invocation. The shape is versioned independently of the vernier-cli package version: the document carries a version field whose value is the schema version, currently "1". This page is the field-by-field reference for that schema.

The schema is a stability contract surface for the CLI — additive changes (new fields, new metric rows in known IoU types) keep the same version string; renames or removals require a schema-version bump ("1" → "2"). The vernier-cli package itself is shipping under 0.0.x patches today, but the schema version is decoupled from the package version on purpose: "version": "1" documents archived against any 0.0.x release stay consumable across all subsequent 0.0.x patches. ADR-0015 §"Versioning and stability commitments" pins this commitment.

Top-level fields

Field	Type	Notes
version	string	Schema version. "1" for everything vernier-cli emits today. Bumps on breaking shape changes; see Schema versioning below.
iou_type	string, one of bbox / segm / boundary / keypoints	Mirrors the --iou-type flag and the IouKind variant the kernel ran. Required: --iou-type has no default.
parity_mode	string, one of strict / corrected	The kernel-resolved parity mode. --parity-mode aligned collapses to strict per ADR-0002 (aligned-tier changes are output-equivalent), so the JSON only ever carries the two distinct kernel modes. Default strict (ADR-0015).
max_dets	array of non-negative integers	The M-axis of the accumulator the eval ran on. Defaults resolve via the kernel-canonical path (ADR-0012): [1, 10, 100] for det-family, [20] for kp.
use_cats	boolean	Mirrors --use-cats / --no-use-cats. Default true.
lines	array of Line objects	Per-row stat plan output. Length is fixed per IoU type: 12 for bbox / segm / boundary, 10 for keypoints.
stats	array of floats	The value column of lines, extracted as a flat array. Same length and order as lines. See stats and lines correspondence below.

Object keys appear in the order shown above. The CLI commits to a fixed key order, not insertion order — see ADR-0015 §"Output determinism".

lines[] subfields

Each entry of lines is an object with the following fields:

Field	Type	Notes
metric	string, AP or AR	Which kernel-output tensor this row reads (precision or recall).
iou_threshold	float or null	The IoU threshold this row pins. null means the row averages across the IoU ladder (the 0.50:0.95 rows). Floats are emitted as the threshold value (e.g. 0.5, 0.75).
iou_threshold_label	string	Human-readable label, used by the text formatter and surfaced here for parsing convenience. "0.50", "0.75", or "0.50:0.95".
area	string	Area-bucket label. "all", "small", "medium", or "large" for the canonical COCO grid. Custom AreaRng callers see their label here verbatim.
max_dets	non-negative integer	The M-axis cap this row resolved to. Always one of the entries in the top-level max_dets array.
value	float	The metric value. Sentinel-filtered (cells equal to -1 are excluded from the mean, per the summarize_detection rule documented in coco-summary-stats.md).

Within a Line, the field order is the order shown above.

stats and lines correspondence

stats[i] is lines[i].value for every i. Both arrays have identical length and identical order. The canonical mapping from index i to (metric, iou_threshold, area, max_dets) for the bbox / segm / boundary plan is the 12-stat table in docs/reference/coco-summary-stats.md. The keypoints plan is 10 entries with max_dets = [20] and a 3-bucket area axis (all, medium, large — the small bucket is dropped per quirk D5); ADR-0012 §"Decision outcome" pins the kp plan order.

The redundancy is deliberate: tools that already index into pycocotools' eval.stats array port to summary["stats"] with one line of code. Tools that prefer addressed access (row["metric"] == "AP" and row["area"] == "small") read from summary["lines"].

Worked examples

The numbers below are illustrative, not pinned. They match the canonical COCO plans in shape and order; the precision values are plausible for a strong detector on COCO val2017.

iou_type = "bbox" (12 lines)

{
  "version": "1",
  "iou_type": "bbox",
  "parity_mode": "strict",
  "max_dets": [1, 10, 100],
  "use_cats": true,
  "lines": [
    {"metric": "AP", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "all",    "max_dets": 100, "value": 0.527},
    {"metric": "AP", "iou_threshold": 0.5,  "iou_threshold_label": "0.50",      "area": "all",    "max_dets": 100, "value": 0.728},
    {"metric": "AP", "iou_threshold": 0.75, "iou_threshold_label": "0.75",      "area": "all",    "max_dets": 100, "value": 0.581},
    {"metric": "AP", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "small",  "max_dets": 100, "value": 0.341},
    {"metric": "AP", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "medium", "max_dets": 100, "value": 0.566},
    {"metric": "AP", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "large",  "max_dets": 100, "value": 0.683},
    {"metric": "AR", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "all",    "max_dets": 1,   "value": 0.392},
    {"metric": "AR", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "all",    "max_dets": 10,  "value": 0.624},
    {"metric": "AR", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "all",    "max_dets": 100, "value": 0.661},
    {"metric": "AR", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "small",  "max_dets": 100, "value": 0.471},
    {"metric": "AR", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "medium", "max_dets": 100, "value": 0.706},
    {"metric": "AR", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "large",  "max_dets": 100, "value": 0.812}
  ],
  "stats": [0.527, 0.728, 0.581, 0.341, 0.566, 0.683, 0.392, 0.624, 0.661, 0.471, 0.706, 0.812]
}

iou_type = "segm" (12 lines)

The shape is identical to the bbox example. Only iou_type and the value column change.

{
  "version": "1",
  "iou_type": "segm",
  "parity_mode": "strict",
  "max_dets": [1, 10, 100],
  "use_cats": true,
  "lines": [
    {"metric": "AP", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "all",    "max_dets": 100, "value": 0.451},
    {"metric": "AP", "iou_threshold": 0.5,  "iou_threshold_label": "0.50",      "area": "all",    "max_dets": 100, "value": 0.687},
    {"metric": "AP", "iou_threshold": 0.75, "iou_threshold_label": "0.75",      "area": "all",    "max_dets": 100, "value": 0.486},
    {"metric": "AP", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "small",  "max_dets": 100, "value": 0.255},
    {"metric": "AP", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "medium", "max_dets": 100, "value": 0.484},
    {"metric": "AP", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "large",  "max_dets": 100, "value": 0.622},
    {"metric": "AR", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "all",    "max_dets": 1,   "value": 0.354},
    {"metric": "AR", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "all",    "max_dets": 10,  "value": 0.561},
    {"metric": "AR", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "all",    "max_dets": 100, "value": 0.589},
    {"metric": "AR", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "small",  "max_dets": 100, "value": 0.382},
    {"metric": "AR", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "medium", "max_dets": 100, "value": 0.633},
    {"metric": "AR", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "large",  "max_dets": 100, "value": 0.751}
  ],
  "stats": [0.451, 0.687, 0.486, 0.255, 0.484, 0.622, 0.354, 0.561, 0.589, 0.382, 0.633, 0.751]
}

iou_type = "boundary" (12 lines)

Same plan shape as segm. The numerical values differ because the kernel evaluates min(mask_iou, boundary_iou) per ADR-0010, not raw mask IoU.

{
  "version": "1",
  "iou_type": "boundary",
  "parity_mode": "strict",
  "max_dets": [1, 10, 100],
  "use_cats": true,
  "lines": [
    {"metric": "AP", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "all",    "max_dets": 100, "value": 0.398},
    {"metric": "AP", "iou_threshold": 0.5,  "iou_threshold_label": "0.50",      "area": "all",    "max_dets": 100, "value": 0.612},
    {"metric": "AP", "iou_threshold": 0.75, "iou_threshold_label": "0.75",      "area": "all",    "max_dets": 100, "value": 0.421},
    {"metric": "AP", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "small",  "max_dets": 100, "value": 0.231},
    {"metric": "AP", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "medium", "max_dets": 100, "value": 0.426},
    {"metric": "AP", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "large",  "max_dets": 100, "value": 0.553},
    {"metric": "AR", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "all",    "max_dets": 1,   "value": 0.317},
    {"metric": "AR", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "all",    "max_dets": 10,  "value": 0.503},
    {"metric": "AR", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "all",    "max_dets": 100, "value": 0.527},
    {"metric": "AR", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "small",  "max_dets": 100, "value": 0.341},
    {"metric": "AR", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "medium", "max_dets": 100, "value": 0.564},
    {"metric": "AR", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "large",  "max_dets": 100, "value": 0.689}
  ],
  "stats": [0.398, 0.612, 0.421, 0.231, 0.426, 0.553, 0.317, 0.503, 0.527, 0.341, 0.564, 0.689]
}

iou_type = "keypoints" (10 lines)

Per ADR-0012 / quirk D5, the keypoint summary drops the small area bucket and the multi-cap AR rows. max_dets resolves to [20] by default. The plan order is AP / AP@.50 / AP@.75 / AP_M / AP_L / AR / AR@.50 / AR@.75 / AR_M / AR_L.

{
  "version": "1",
  "iou_type": "keypoints",
  "parity_mode": "strict",
  "max_dets": [20],
  "use_cats": true,
  "lines": [
    {"metric": "AP", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "all",    "max_dets": 20, "value": 0.642},
    {"metric": "AP", "iou_threshold": 0.5,  "iou_threshold_label": "0.50",      "area": "all",    "max_dets": 20, "value": 0.864},
    {"metric": "AP", "iou_threshold": 0.75, "iou_threshold_label": "0.75",      "area": "all",    "max_dets": 20, "value": 0.708},
    {"metric": "AP", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "medium", "max_dets": 20, "value": 0.598},
    {"metric": "AP", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "large",  "max_dets": 20, "value": 0.711},
    {"metric": "AR", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "all",    "max_dets": 20, "value": 0.706},
    {"metric": "AR", "iou_threshold": 0.5,  "iou_threshold_label": "0.50",      "area": "all",    "max_dets": 20, "value": 0.901},
    {"metric": "AR", "iou_threshold": 0.75, "iou_threshold_label": "0.75",      "area": "all",    "max_dets": 20, "value": 0.764},
    {"metric": "AR", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "medium", "max_dets": 20, "value": 0.659},
    {"metric": "AR", "iou_threshold": null, "iou_threshold_label": "0.50:0.95", "area": "large",  "max_dets": 20, "value": 0.768}
  ],
  "stats": [0.642, 0.864, 0.708, 0.598, 0.711, 0.706, 0.901, 0.764, 0.659, 0.768]
}

Determinism guarantees

ADR-0015 §"Output determinism" pins the following invariants for the JSON formatter; every consumer that archives or diffs the output relies on them:

• Fixed key order. Object keys are emitted in the schema-defined order documented above (top-level and per-Line), not insertion order and not sorted alphabetically. The lines array is in plan order — the same order Summary::pretty_lines() produces — never sorted by metric name. Reordering keys in a future PR is a schema-version bump.
• No timestamps. No generated_at, started_at, wall_clock_seconds, or any other field whose value changes between two runs of the same input. The eval inputs (GT, DT, flags) define the result's identity.
• No environment leakage. No host, user, working directory, or vernier-build-metadata fields. The schema's version is the contract surface; the commit of vernier that produced the file lives in Cargo.lock / the release tag, not in the document.
• Round-trip-safe float formatting. value fields use Rust's default {} for f64, which is the shortest representation that round-trips through f64::from_str. Stable across the workspace MSRV (rustc 1.83+) and across platforms.
• Atomic file writes. When --emit json=PATH is used, the file is written to PATH.tmp.<pid>, fsynced, and renamed atomically. A concurrent reader either sees the previous contents in full or the new contents in full — never a half-written file.

The contract: byte-equal output for byte-equal input, across runs, machines, and elapsed time. The one well-defined exception is the version field, which only changes on schema revisions.

Schema versioning

version tracks the schema, not the vernier-cli package version. The CLI commits to:

• Backward-compatible additions ship under the same version value. New fields appended to the top level or to lines[] for an existing IoU type do not bump version.
• Renames, removals, or shape changes bump version to the next integer string. The older shape remains the CLI's default until a future release switches the default; the per-formatter knob --emit json[,version=N] (ADR-0015 §"Formatter: JSON") is the planned opt-in for newer shapes ahead of that switch. Today the only shipped schema is "1" and --emit json,version=N has no other accepted values.
• Older schemas remain consumable for one schema generation past the default switch. A pipeline that pins vernier-cli to a specific 0.0.x patch and stores its output for replay can read the same bytes back through every subsequent 0.0.x patch that keeps "version": "1" as the default, and through the first generation after the default flips. The generation after that, parsers may stop accepting "1".

The schema version is decoupled from the package version on purpose — vernier-cli ships under 0.0.x patches today (no SemVer guarantees on the package), but "version": "1" is a hard contract: reshaping it outside the discipline above breaks the regression-archive use case the JSON formatter is built around.

Schema v2 — partitioned output (--manifest)

ADR-0046 adds a fan-out lane: when vernier eval --manifest PATH is passed, the binary emits a partitioned JSON document under "version": "2". Un-partitioned eval keeps emitting v1 verbatim — that is the load-bearing byte-stability contract of ADR-0046.

Rule: without --manifest, the output is byte-for-byte identical to the v1 document above. The v2 envelope is only emitted when a manifest drives the eval.

v2 top-level fields

Field	Type	Notes
version	string	"2" for partitioned output.
label	string or null	--label value stamped on this run. null when omitted. vernier aggregate joins by this field when present.
iou_type	string	As v1.
parity_mode	string	As v1.
max_dets	array of non-negative integers	As v1.
use_cats	boolean	As v1.
overall	Overall object	The un-partitioned summary. Bit-identical to a v1 document over the same (GT, DT) pair — load-bearing parity contract per ADR-0046.
slices	array of Slice objects	One entry per partition slice, in canonical order (axis ascending, value ascending, __unassigned__ last; joint cells follow the marginals).

overall subfields

Field	Type	Notes
lines	array of Line objects	Same shape and order as v1's lines.
stats	array of floats	Same shape and order as v1's stats.
n_images	non-negative integer	Number of dataset images behind the overall summary.
n_detections	non-negative integer	Total detection count behind the overall summary (a=0 "all" bucket counted once per detection).

slices[] subfields

Field	Type	Notes
axis	string	The manifest axis name. For joint cells, the ::-joined tuple (weather::time_of_day).
value	string	The categorical level. For joint cells, the ::-joined value tuple (fog::night). __unassigned__ for unassigned.
n_images	non-negative integer	Number of dataset images in this slice.
n_detections	non-negative integer	Detection count in this slice (same accounting as overall.n_detections).
lines	array of Line objects	Same shape and order as v1's lines, evaluated over this slice's image subset.
stats	array of floats	Same shape and order as v1's stats, evaluated over this slice's image subset.

Determinism guarantees (v2)

The v1 determinism rules carry through verbatim. Additionally:

• Slice order is canonical, not insertion order: (axis ascending, value ascending, __unassigned__ last) for marginals; joint cells follow the marginals in the same canonical order applied to the joined axis / value strings.
• overall is bit-identical to a v1 run on the same (GT, DT) pair — the partitioned dispatch invokes the same accumulate + summarize calls over the un-filtered grid.

Worked example (v2)

{
  "version": "2",
  "label": "run_2026_05_14",
  "iou_type": "bbox",
  "parity_mode": "strict",
  "max_dets": [1, 10, 100],
  "use_cats": true,
  "overall": {
    "lines": [/* 12 entries, same shape as v1 */],
    "stats": [0.527, 0.728, 0.581, 0.341, 0.566, 0.683, 0.392, 0.624, 0.661, 0.471, 0.706, 0.812],
    "n_images": 5000,
    "n_detections": 31_500
  },
  "slices": [
    {"axis": "weather", "value": "clear",          "n_images": 3700, "n_detections": 22_900, "lines": [/* ... */], "stats": [/* 12 */]},
    {"axis": "weather", "value": "fog",            "n_images": 1100, "n_detections":  7_300, "lines": [/* ... */], "stats": [/* 12 */]},
    {"axis": "weather", "value": "__unassigned__", "n_images":  200, "n_detections":  1_300, "lines": [/* ... */], "stats": [/* 12 */]}
  ]
}

See docs/reference/manifest-schema.md for the manifest input and docs/reference/aggregate-schema.md for the vernier aggregate companion output.

Schema v2 — partitioned LRP output (--metric olrp --manifest)

ADR-0046 (with ADR-0043 / ADR-0044 / ADR-0045 for the LRP semantics) extends the partition fan-out to the LRP / oLRP metric. When --metric olrp --manifest PATH is supplied, the binary emits a partitioned LRP JSON document under "version": "2", distinguished from the AP v2 envelope by a "metric": "olrp" discriminator.

Rule: --metric olrp without --manifest keeps emitting the un-partitioned LRP shape under "version": "1". The partitioned-LRP envelope is only emitted when both --metric olrp and --manifest are present.

LRP v2 top-level fields

Field	Type	Notes
version	string	"2" for partitioned output.
metric	string	"olrp" — discriminator for parsers that switch on (version, metric).
label	string or null	--label value stamped on this run. null when omitted.
iou_type	string	As v1.
parity_mode	string	As v1.
use_cats	boolean	As v1.
overall	LrpOverall object	The un-partitioned LRP block. Bit-identical to a single optimal_lrp_* call over the same (GT, DT).
slices	array of LrpSlice objects	One entry per partition slice, in the same canonical order as the AP v2 envelope.

Note: the LRP envelope does not carry the AP max_dets top-level field. LRP runs at a single max_dets_per_image rung (the top of the resolved ladder); the value is implicit in the kernel-canonical defaults and is recorded inside the resolved config for documentation rather than as a wire-level array.

LrpOverall subfields

Field	Type	Notes
olrp	float	Mean per-class oLRP across classes with at least one positive GT (per ADR-0043).
olrp_loc	float	Mean per-class oLRP_Loc across classes with at least one TP at the optimal tau.
olrp_fp	float	Mean per-class oLRP_FP (same denominator as olrp_loc).
olrp_fn	float	Mean per-class oLRP_FN (same denominator as olrp).
n_empty_classes	non-negative integer	Number of classes with no positive GTs (excluded from the headline means).
n_images	non-negative integer	Number of dataset images behind the overall report.
n_detections	non-negative integer	Total detection count behind the overall report.
config	LrpConfig object	Resolved configuration (kernel, tp_threshold, tau_grid_len) — every report self-describes per ADR-0044.

LrpSlice subfields

Field	Type	Notes
axis	string	The manifest axis name. For joint cells, the ::-joined tuple.
value	string	The categorical level. For joint cells, the ::-joined value tuple. __unassigned__ for unassigned.
n_images	non-negative integer	Number of dataset images in this slice.
n_detections	non-negative integer	Detection count in this slice.
olrp	float	Per-slice headline oLRP, computed by restricting the LRP decompose walk to this slice's image set.
olrp_loc	float	Per-slice oLRP_Loc.
olrp_fp	float	Per-slice oLRP_FP.
olrp_fn	float	Per-slice oLRP_FN.
n_empty_classes	non-negative integer	Per-slice count of classes with no positive GTs (a slice may be more sparsely class-covered than the overall dataset).

LrpConfig subfields

Field	Type	Notes
tp_threshold	float	IoU/OKS floor for TP — resolved per ADR-0044 (0.5 for every kernel by default).
tau_grid_len	non-negative integer	Length of the confidence-threshold grid the LRP search ran on (101 for the canonical default grid).
kernel	string	Canonical kernel name ("bbox" / "segm" / "boundary" / "keypoints").

per_class is omitted by default

The un-partitioned LRP v1 envelope ships a per_class array (one row per category) carrying the deployable tau plus the four per-class decomposition fields. The partitioned LRP v2 envelope omits per_class from both overall and slices by default, for two reasons:

• Size at LVIS scale. A 1203-category dataset crossed with 8 slices would balloon the document by ~10k per-class rows — the bulk of which downstream vernier aggregate flows do not consume.
• Wrong surface for the partitioned use-case. The partitioned document is the comparative table — the headline numbers per slice are the reason to slice. Per-class detail is the un-partitioned --metric olrp run's job; users who want both spawn both.

A future --per-class opt-in flag is anticipated if a workload ever needs the per-class table embedded in the partitioned envelope.

Determinism guarantees (LRP v2)

The v1 / AP-v2 determinism rules carry through verbatim: fixed key order, no timestamps, no environment leakage, round-trip-safe float formatting, atomic file writes. Additionally:

• Slice order is canonical, same rule as the AP v2 envelope: (axis ascending, value ascending, __unassigned__ last) for marginals; joint cells follow the marginals.
• overall is bit-identical to a v1 LRP run on the same (GT, DT) pair — the partitioned LRP dispatch invokes the same matching engine once and runs the decompose walk over the un-filtered image set for overall.

Worked example (LRP v2)

{
  "version": "2",
  "metric": "olrp",
  "label": "run_2026_05_14",
  "iou_type": "bbox",
  "parity_mode": "strict",
  "use_cats": true,
  "overall": {
    "olrp": 0.412,
    "olrp_loc": 0.183,
    "olrp_fp": 0.092,
    "olrp_fn": 0.137,
    "n_empty_classes": 0,
    "n_images": 5000,
    "n_detections": 31_500,
    "config": {"tp_threshold": 0.5, "tau_grid_len": 101, "kernel": "bbox"}
  },
  "slices": [
    {"axis": "weather", "value": "clear",          "n_images": 3700, "n_detections": 22_900, "olrp": 0.398, "olrp_loc": 0.177, "olrp_fp": 0.086, "olrp_fn": 0.135, "n_empty_classes": 0},
    {"axis": "weather", "value": "fog",            "n_images": 1100, "n_detections":  7_300, "olrp": 0.487, "olrp_loc": 0.221, "olrp_fp": 0.121, "olrp_fn": 0.145, "n_empty_classes": 2},
    {"axis": "weather", "value": "__unassigned__", "n_images":  200, "n_detections":  1_300, "olrp": 0.453, "olrp_loc": 0.198, "olrp_fp": 0.114, "olrp_fn": 0.141, "n_empty_classes": 4}
  ]
}

See also

• docs/adr/0015-vernier-cli.md — the source of truth for the CLI surface, the formatter abstraction, and the determinism contract.
• docs/adr/0046-slice-and-aggregate.md — partition manifest and the v1 → v2 schema bump.
• docs/reference/coco-summary-stats.md — the stats[i] ↔ slice mapping for the bbox / segm 12-stat plan.
• docs/adr/0012-oks-keypoints-surface.md — keypoints plan order and kernel-canonical max_dets = [20].
• docs/adr/0010-boundary-iou-isolated-subsystem.md — boundary IoU metric definition consumed by iou_type = "boundary".
• docs/adr/0043-lrp-oracle-and-namespace.md / docs/adr/0044-lrp-thresholds-and-tau-grid.md / docs/adr/0045-lrp-keypoints-shipped.md — the LRP / oLRP semantics, per-kernel defaults, and keypoints support consumed by --metric olrp.