Nishi Family › Compare › Autonomous Coding & Code-LLM Harness

Nishi Compare · measured, not asserted

Autonomous Coding & Code-LLM Harness

Nishi vs the field — every Nishi cell is measured against real organ source at emit time; each gap names the watch contract that will close it.

Nishi forge vs Cursor and SWE-agent and Qwen2.5-Coder and Aider

Overview

The purpose, declared position and evidence coverage of this domain. Source presence and completed acceptance are different measures.

Where we are. Nishi forge is a real sovereign coding harness: organ-emitted verified context pack, a DETERMINISTIC build plus byte-exact judge (exceed: the verify signal is noise-free), a bounded repair loop on verbatim compiler stderr, a self-retiring vetted-rules registry (exceed), local no-float Coder 0.5B and 1.5B tiers loaded bits-up (exceed), a liar-killed requirements dossier, ChatML prompting, sovereign semantic code retrieval with an externally-validated integer embedding path. It is NOT at frontier capability: retrieval is base-weight quality, prefill is per-token, no best-of-N, no grammar-constrained decoding, single-file scope, hand-authored oracles, and zero external-leaderboard results.

Where we need to go. Stand on a public leaderboard with a fully sovereign stack, then climb it by the levers the record names: batched prefill for throughput, best-of-N against the free verifier, grammar masking into NishiLang, a contrastive retrieval tune on our own verified ledger, multi-file repair, and generated oracles -- each proven by the byte-exact judge, never by an LLM grading an LLM.

The unit. 1 u = one measured session-leg (estate calibration: graphics R21 in one leg 2026-08-15). Local evidence: the code-retrieval organ (nx_code_index, 5-tooth gate, external oracle match) landed in one leg on 2026-07-15; estimates below are relative to that.
Where we are: 7 open rungs. Harness and judge AHEAD on determinism and ruler validity; BEHIND on external numbers, retrieval quality and repo-level tasks. Counts measured at emit below this line.
Cost to throughput and sampling: 2.5 u. CD1 batched prefill (the N6 keystone), then CD2 best-of-N against the verifier it makes affordable.
Cost to a public number: 2 u. CD6 the Aider-Polyglot runner -- measured standing before any capability claim, the grade-1 referee this lane lacks.
Cost to frontier-shaped capability: 9 u. CD3 grammar mask, CD4 contrastive retrieval tune, CD5 multi-file editing, CD7 oracle generation.

Research bar. Aider-Polyglot is measured on 225-task multi-language edit benchmark, scaffold-sensitive (little-coder 2026: 19 to 45 percent by scaffold alone). Theirs: public pass-rate leaderboard. Ours: CD6 publishes our number on this page.

Research bar. SWE-bench is measured on repository-level issue resolution. Theirs: the repo-level frontier. Ours: CD5 is the parity rung; SWE-bench-Lite is the stretch.

Research bar. Large Language Monkeys (arXiv 2407.21787) is measured on coverage scales with samples given a verifier (15.9 to 56 percent at 250 samples). Theirs: best-of-N with a verifier. Ours: CD2 -- our verifier is perfect and free.

Latest recorded release

No valid dated release entry is recorded for this domain.

Release entries describe recorded changes; they do not establish that every capability passed evaluation.

11 of 18 capabilities measured|5 of them measured exceeds|7 open|coverage 611/1000|adoption 10 full / 1 partial

Evidence profile — what the gaps on this board actually are

Measured by nx_swcompare_evidence, read back by nx_evprofile_lib. Every figure is a count with its denominator — there is deliberately no score, no grade and no percentage anywhere in this band, because a stored scalar is a field a seat can edit and a counted partition is not.

evidence|grounded 11/11|unsupported 0|gates green 1/1|proven able to fail 1/1|never bitten 0|green at 0/0 0|open gaps 7|of them unnamed 0|of them proof withheld 0|flips ready 0

proven able to fail counts the gates that have a RECORDED RED — nx_gate_bite mutated the gate subject, rebuilt it, watched the gate go red, and that record is inside the shared TTL. never bitten is its complement over the same denominator: those gates ran and were green, and nothing has ever shown them able to detect anything, so their green is a statement about this run and not about the gate. green at 0/0 is a separate and much weaker observation — the gate printed GREEN on a zero denominator, so its own tooth counter says it examined nothing. A gate can be green, non-zero, and still never bitten; that is the common case and it is now visible instead of implied.

partition: grounded + unsupported = 11 vs present 11 · named + unnamed + withheld = 7 vs open 7 · both reconcile

liar-kill conj=GPQN · all four conjuncts held

graded document: BUILDROOT tree, 8645 bytes · gates map: PRIMARY · stamped 2d 7h ago · source ../knowledge/status/evstamp_coding.verdict

Gap classWhat it is, and the work it names
NONENo gap class fires on this board: every published claim is grounded, the declared gates ran and were green, no watch contract is unnamed or already landed, and the stamp is inside its TTL. That is a statement about evidence, not about depth, scale or polish.
This band reports the referee counts. The per-axis worklist rows — which axis is unsupported, which watch is ready to flip, which gap is unnamed — are printed by nx_swcompare_evidence coding itself and are not carried on the stamp, so this page names the classes and the producer names the rows. That split is stated rather than hidden: a count without a worklist is not actionable, and this band is honest about which half of that it is.

Production map

Follow the dependencies, declared acceptance criteria and recorded priorities. Inspect source binding before treating a rank as executable work.

Ranking source binding: PLAN_MATRIX_BOUND_ONLY. Recorded priorities require current acceptance evidence and resource checks before execution.

Ranking matches the captured plan and matrix only. Latest execution outcome, research freshness, accepted delivery and investment return are unverified.

Recorded priority estimates

Order from nx_compare_rank (nx_dr_ocm: (deficit + cost-of-delay + option + enables) x sponsor x self-sufficiency x momentum / cost). FINISH rows are rungs whose symbol is present but whose organ is short of full adoption: listed before new work by this heuristic. Priority is not measured delivery cost or execution readiness. Stamp: # asof=1789541549 domain=coding target_version=0.1 rungs=11 done=4 open=7 finish=0 ranker=nx_dr_ocm

#StageRungPriorityDerivation
#10.1Aider-Polyglot runner (CD6) fg_polyglot_bench9600v=16 m=3 c=5
#2laterBatched prefill (CD1) nsv_prefill_batch3714v=13 m=2 c=7
#3laterMulti-file editing (CD5) fg_multi_file3428v=12 m=2 c=7
#4laterBest-of-N against the verifier (CD2) fg_best_of_n3000v=3 m=2 c=2
#5laterOracle generation (CD7) fg_oracle_gen2400v=6 m=2 c=5
#6laterContrastive retrieval tune (CD4) ci_train_contrastive900v=9 m=2 c=20
#7laterGrammar-constrained decoding (CD3) nsv_grammar_mask857v=3 m=2 c=7

Declared roadmap — contract, acceptance, executor, effort

RungCloses withDefinition of done (pre-declared)ExecutorEst.
Byte-exact judge (CB0)fg_runDeterministic build plus byte-exact judge -- LANDED, the verifier every open rung leans onOrgan0 u
Bounded repair loop (CB1)
after CB0		fg_bankVerbatim compiler stderr fed back, gate 5/5 -- LANDEDOrgan0 u
Local model tier (CB2)load_named_q16Coder 0.5B and 1.5B q8_0 through the no-float stack -- LANDEDOrgan0 u
Semantic code retrieval (CB3)ci_query_top1nx_code_index, jina recipe, 5-tooth gate -- LANDED at base-weight qualityOrgan0 u
Batched prefill (CD1)
after CB2		nsv_prefill_batchPrefill runs the prompt as ONE batched GEMM (m greater than 1) instead of m looped m=1 matmuls; pass-fail tooth: per-token prefill cost must FALL with prompt length on nx_embed_bench at m=8,32,64,125 -- flat means we did not enter the batched regime and must not claim itOrgan1.5 u
Best-of-N against the verifier (CD2)
after CB0,CD1		fg_best_of_nSample N candidates per task, run each through the byte-exact judge, keep the first GREEN; gate proves pass@N is monotone in N on the L4 mini-pack and that a task with no GREEN in N reports NONE rather than the best-scoring wrong answerOrgan1 u
Grammar-constrained decoding (CD3)
after CB2		nsv_grammar_maskThe serve masks logits to the NishiLang grammar's legal next tokens; gate proves every sampled program at least PARSES and a control with the mask off reproduces the current parse-failure rateOrgan1.5 u
Contrastive retrieval tune (CD4)
after CB3		ci_train_contrastiveTrain the embedding head on the forge's own compile-verified query-answer pairs through the no-float autograd; done-rule: top-1 on the held-out ruler EXCEEDS the 3-of-6 base-weight ratchet, declared before the run, or it does not shipOrgan2 u
Multi-file editing (CD5)
after CB1		fg_multi_fileA task may name several organs; the repair loop applies a coordinated patch set and the judge builds the whole closure; gate proves a two-file rename task lands and a patch that breaks a third importer is refused by the buildOrgan3 u
Aider-Polyglot runner (CD6)
after CB0,CB2		fg_polyglot_benchA sovereign runner over the public task set using the byte-exact judge, publishing pass rate per language with the task count; the first number is whatever it is -- an honest standing beats a claimed oneOrgan2 u
Oracle generation (CD7)
after CB0		fg_oracle_genGenerate candidate test oracles for a task and ACCEPT one only when it discriminates a known-good from a known-bad solution (anti-vacuity by construction); gate proves a generated oracle fails the planted mutant and passes the referenceOrgan2 u

Milestones

MilestoneRungsCumulative
M1 · Batched and sampledCD1,CD24.5 u
M2 · On a public leaderboardCD62 u
M3 · Frontier-shaped harnessCD3,CD4,CD5,CD713.5 u
Ladder verdict. NOT DECLARED. This board names no dated best-in-class or frontier target and no rung roles (sotatarget and rungrole rows on its plan); the ranker labels it NO-LADDER until it does, and until then its rungs climb toward a target nobody has written down. Bars: 0 (fresh 0, attested 0, stale 0, unattested 0), verdict NO-BAR against the current month 2026-09.

Inspect a rung and its prerequisites

Declared nodes 11. Rank input binding: PLAN_MATRIX_BOUND_ONLY. Dependency order is authored. Implementation, acceptance evidence, authority and resource readiness are unverified. No action is recommended or dispatched here.

Plan SHA-256 03f6c43c9254afc5dfeb99d984bf5032049fbd99cd45aa19d0ec85961eef93f4. Target rows 0; role rows 0. Existing risks and release worklog retain their own scope; no node completion is inferred.

Use Enter or Space on a rung to inspect its contract. Prerequisite links locate another rung in this list; open its summary to inspect it. Estimates are authored effort, not forecasts.

  1. CB0 — Byte-exact judge

    Prerequisites: None declared; this does not establish execution eligibility.

    Contract: fg_run

    Acceptance: Deterministic build plus byte-exact judge -- LANDED, the verifier every open rung leans on

    Authored effort: 0. Executor kind: Organ. Responsible, accountable and verifier not established. Inspect retained worklog.

  2. CB1 — Bounded repair loop

    Prerequisites: CB0 (acceptance unverified)

    Contract: fg_bank

    Acceptance: Verbatim compiler stderr fed back, gate 5/5 -- LANDED

    Authored effort: 0. Executor kind: Organ. Responsible, accountable and verifier not established. Inspect retained worklog.

  3. CB2 — Local model tier

    Prerequisites: None declared; this does not establish execution eligibility.

    Contract: load_named_q16

    Acceptance: Coder 0.5B and 1.5B q8_0 through the no-float stack -- LANDED

    Authored effort: 0. Executor kind: Organ. Responsible, accountable and verifier not established. Inspect retained worklog.

  4. CB3 — Semantic code retrieval

    Prerequisites: None declared; this does not establish execution eligibility.

    Contract: ci_query_top1

    Acceptance: nx_code_index, jina recipe, 5-tooth gate -- LANDED at base-weight quality

    Authored effort: 0. Executor kind: Organ. Responsible, accountable and verifier not established. Inspect retained worklog.

  5. CD1 — Batched prefill

    Prerequisites: CB2 (acceptance unverified)

    Contract: nsv_prefill_batch

    Acceptance: Prefill runs the prompt as ONE batched GEMM (m greater than 1) instead of m looped m=1 matmuls; pass-fail tooth: per-token prefill cost must FALL with prompt length on nx_embed_bench at m=8,32,64,125 -- flat means we did not enter the batched regime and must not claim it

    Authored effort: 1.5. Executor kind: Organ. Responsible, accountable and verifier not established. Inspect retained worklog.

  6. CD2 — Best-of-N against the verifier

    Prerequisites: CB0 (acceptance unverified), CD1 (acceptance unverified)

    Contract: fg_best_of_n

    Acceptance: Sample N candidates per task, run each through the byte-exact judge, keep the first GREEN; gate proves pass@N is monotone in N on the L4 mini-pack and that a task with no GREEN in N reports NONE rather than the best-scoring wrong answer

    Authored effort: 1. Executor kind: Organ. Responsible, accountable and verifier not established. Inspect retained worklog.

  7. CD3 — Grammar-constrained decoding

    Prerequisites: CB2 (acceptance unverified)

    Contract: nsv_grammar_mask

    Acceptance: The serve masks logits to the NishiLang grammar's legal next tokens; gate proves every sampled program at least PARSES and a control with the mask off reproduces the current parse-failure rate

    Authored effort: 1.5. Executor kind: Organ. Responsible, accountable and verifier not established. Inspect retained worklog.

  8. CD4 — Contrastive retrieval tune

    Prerequisites: CB3 (acceptance unverified)

    Contract: ci_train_contrastive

    Acceptance: Train the embedding head on the forge's own compile-verified query-answer pairs through the no-float autograd; done-rule: top-1 on the held-out ruler EXCEEDS the 3-of-6 base-weight ratchet, declared before the run, or it does not ship

    Authored effort: 2. Executor kind: Organ. Responsible, accountable and verifier not established. Inspect retained worklog.

  9. CD5 — Multi-file editing

    Prerequisites: CB1 (acceptance unverified)

    Contract: fg_multi_file

    Acceptance: A task may name several organs; the repair loop applies a coordinated patch set and the judge builds the whole closure; gate proves a two-file rename task lands and a patch that breaks a third importer is refused by the build

    Authored effort: 3. Executor kind: Organ. Responsible, accountable and verifier not established. Inspect retained worklog.

  10. CD6 — Aider-Polyglot runner

    Prerequisites: CB0 (acceptance unverified), CB2 (acceptance unverified)

    Contract: fg_polyglot_bench

    Acceptance: A sovereign runner over the public task set using the byte-exact judge, publishing pass rate per language with the task count; the first number is whatever it is -- an honest standing beats a claimed one

    Authored effort: 2. Executor kind: Organ. Responsible, accountable and verifier not established. Inspect retained worklog.

  11. CD7 — Oracle generation

    Prerequisites: CB0 (acceptance unverified)

    Contract: fg_oracle_gen

    Acceptance: Generate candidate test oracles for a task and ACCEPT one only when it discriminates a known-good from a known-bad solution (anti-vacuity by construction); gate proves a generated oracle fails the planted mutant and passes the reference

    Authored effort: 2. Executor kind: Organ. Responsible, accountable and verifier not established. Inspect retained worklog.

Learning and practice paths

No structured learning path is declared for this plan. Existing research, roadmap and worklog remain available above.

Capability comparisons

Compare the field, search individual capabilities and open their source and adoption evidence. Documented presence does not establish comparative quality.

Position map — centrality and distinctiveness

The four-quadrant map the field uses for brand strategy (Dawar and Bagga, HBR June 2015), re-derived from this matrix on every publish. Centrality is the share of the category's feature mass a player covers, each feature weighted by how many hold it; distinctiveness is the average lead over each rival on the rows the player holds; breadth is the depth-weighted share of the whole matrix (the bubble); depth is how deeply the rows held are held; momentum is the day-over-day move off the spine (green rising, red falling, grey until day two); the dashed path runs first day → previous day → today. Dividers are the category means. Axes are fitted to the field of play, so read the tick numerals, not the frame. Rival marks are documented presence, so a rival's position reads the record, never its quality. The picture grades its own readability below; the table beside it is the same data for a screen reader or a second method.

Views. 2D cut: centrality, distinctiveness · 3D cube: centrality, distinctiveness, breadth · axes registered: centrality, distinctiveness, breadth, depth, momentum · a board picks its own in knowledge/compare/coding.cdmap (cut|x|y, cube|x|y|z); absent = the HBR defaults
Position map, two-dimensional cutOne bubble per player. Bubble area is breadth, the ring colour is momentum, the dashed lines are the category means, and both axes are fitted to the field of play with their tick numerals shown. Every value is repeated in the table that follows. 500 600 700 800 900 1000 100 200 300 400 500 600 centrality (permil, fitted 500–1000) distinctiveness (permil, fitted 100–600) mean 769 mean 283 Unconventional Aspirational Peripheral Mainstream Nishi, centrality 573, distinctiveness 477, breadth 500, depth 818, momentum 0 Cursor, centrality 868, distinctiveness 243, breadth 574, depth 794, momentum 0 SWE-agent, centrality 868, distinctiveness 205, breadth 537, depth 743, momentum 0 Qwen2.5-Coder, centrality 721, distinctiveness 340, breadth 555, depth 833, momentum 0 Aider, centrality 819, distinctiveness 152, breadth 462, depth 694, momentum 0 Nishi Cursor SWE-agent Qwen2.5-Coder Aider

bubble area = breadth · ring = momentum (green rising, red falling, grey until day two) · dashed = category means · axes fitted to the field of play: centrality 500–1000, distinctiveness 100–600 of 0–1000 permil (the full range put every player in one corner)
readability of the 2D cut, self-graded by the layout ruler: label overlaps 0 · labels over marks 0 · off-canvas 0 · unresolved labels 0 · mark overlaps 0 (a fact of the data: two players that close are that close) · data spread 620 permil of the plot · quadrant words unseated 0
text contrast, measured with wcag2-ratio (floors from contrast.conf), light theme: labels 16.24 (floor 4.50) · notes 16.24 (floor 4.50) · quadrant words 5.89 (floor 4.50) · tick numerals 5.89 (floor 4.50) · axis titles 5.59 (floor 4.50) · dark theme: labels 13.78 (floor 4.50) · quadrant words 5.55 (floor 4.50) · axis titles 5.97 (floor 4.50) · dark classes under their floor 0 (one figure serves both themes: a dark shortfall is a token to fix, never a class to hide) · classes refused under their floor 0 (a refused class is not drawn; the scale classes are measured, never hidden) · export: SVG PNG (receipt, rendered by the estate's own rasteriser from this page)
Position map, three-dimensional cubeThe same players in an isometric cube. Each axis is fitted to its own field of play and carries numerals on the floor grid and the vertical axis; each bubble drops a dotted line to its floor shadow so height reads as height. Values are in the table that follows. 500 600 700 800 900 1000 100 200 300 400 500 600 425 450 475 500 525 550 575 600 centrality distinctiveness breadth Aider: centrality 819, distinctiveness 152, breadth 462 Nishi: centrality 573, distinctiveness 477, breadth 500 Qwen2.5-Coder: centrality 721, distinctiveness 340, breadth 555 SWE-agent: centrality 868, distinctiveness 205, breadth 537 Cursor: centrality 868, distinctiveness 243, breadth 574 Nishi Cursor SWE-agent Qwen2.5-Coder Aider
axes fitted: centrality 500–1000 · distinctiveness 100–600 · breadth 425–600 permil · farther bubbles are painted first, nearer ones over them · the floor shadow is each bubble's (x, y) at height 0
readability of the 3D cube, self-graded by the layout ruler: label overlaps 0 · labels over marks 0 · off-canvas 0 · unresolved labels 0 · mark overlaps 0 (a fact of the data: two players that close are that close) · data spread 409 permil of the plot
Position map, small multiplesEvery pair of registered axes as one small cut. Each panel fits both axes to the field of play, the dashed lines are the category means, bubble area is breadth and the ring colour is momentum; the legend below names the colours. 500 1000 100 600 distinctiveness vs centrality Nishi: centrality 573, distinctiveness 477 Cursor: centrality 868, distinctiveness 243 SWE-agent: centrality 868, distinctiveness 205 Qwen2.5-Coder: centrality 721, distinctiveness 340 Aider: centrality 819, distinctiveness 152 500 1000 425 600 breadth vs centrality Nishi: centrality 573, breadth 500 Cursor: centrality 868, breadth 574 SWE-agent: centrality 868, breadth 537 Qwen2.5-Coder: centrality 721, breadth 555 Aider: centrality 819, breadth 462 500 1000 650 900 depth vs centrality Nishi: centrality 573, depth 818 Cursor: centrality 868, depth 794 SWE-agent: centrality 868, depth 743 Qwen2.5-Coder: centrality 721, depth 833 Aider: centrality 819, depth 694 500 1000 480 520 momentum vs centrality Nishi: centrality 573, momentum 500 Cursor: centrality 868, momentum 500 SWE-agent: centrality 868, momentum 500 Qwen2.5-Coder: centrality 721, momentum 500 Aider: centrality 819, momentum 500 100 600 425 600 breadth vs distinctiveness Nishi: distinctiveness 477, breadth 500 Cursor: distinctiveness 243, breadth 574 SWE-agent: distinctiveness 205, breadth 537 Qwen2.5-Coder: distinctiveness 340, breadth 555 Aider: distinctiveness 152, breadth 462 100 600 650 900 depth vs distinctiveness Nishi: distinctiveness 477, depth 818 Cursor: distinctiveness 243, depth 794 SWE-agent: distinctiveness 205, depth 743 Qwen2.5-Coder: distinctiveness 340, depth 833 Aider: distinctiveness 152, depth 694 100 600 480 520 momentum vs distinctiveness Nishi: distinctiveness 477, momentum 500 Cursor: distinctiveness 243, momentum 500 SWE-agent: distinctiveness 205, momentum 500 Qwen2.5-Coder: distinctiveness 340, momentum 500 Aider: distinctiveness 152, momentum 500 425 600 650 900 depth vs breadth Nishi: breadth 500, depth 818 Cursor: breadth 574, depth 794 SWE-agent: breadth 537, depth 743 Qwen2.5-Coder: breadth 555, depth 833 Aider: breadth 462, depth 694 425 600 480 520 momentum vs breadth Nishi: breadth 500, momentum 500 Cursor: breadth 574, momentum 500 SWE-agent: breadth 537, momentum 500 Qwen2.5-Coder: breadth 555, momentum 500 Aider: breadth 462, momentum 500 650 900 480 520 momentum vs depth Nishi: depth 818, momentum 500 Cursor: depth 794, momentum 500 SWE-agent: depth 743, momentum 500 Qwen2.5-Coder: depth 833, momentum 500 Aider: depth 694, momentum 500
Nishi Cursor SWE-agent Qwen2.5-Coder Aider
10 panels over 5 registered axes, every pair once (the lower axis on x, the higher on y) · each panel fitted to its own field of play, first and last tick numerals shown · no labels in a small cut, the legend names the colours; the grade below reports the mark terms only (MM summed over the panels, spread averaged)
readability of the small multiples, self-graded by the layout ruler: label overlaps 0 · labels over marks 0 · off-canvas 0 · unresolved labels 0 · mark overlaps 1 (a fact of the data: two players that close are that close) · data spread 486 permil of the plot
PlayerQuadrantcentralitydistinctivenessbreadthdepthmomentumRows heldDays on spineFirst seen
NishiUnconventional5734775008180 since 2026-09-151122026-09-15
CursorMainstream8682435747940 since 2026-09-151322026-09-15
SWE-agentMainstream8682055377430 since 2026-09-151322026-09-15
Qwen2.5-CoderUnconventional7213405558330 since 2026-09-151222026-09-15
AiderMainstream8191524626940 since 2026-09-151222026-09-15
DayMatrix rows reviewedPlayers recorded
2026-09-15185
2026-09-16185

players 5|matrix rows 18|feature mass 61|centrality mean 769|distinctiveness mean 283|axes 5|spine days 2 (shown 2)|rows written today 0|readability defects 2D 0 cube 0|spine knowledge/status/cdmap/coding.spine

How this is scored. Every Nishi mark is measured: the generator reads the real organ source on disk and requires the implementing symbol to exist (no self-grading). A watching tag names the organ and symbol contracted to close a gap — the mark flips itself on the next compare beat when that workstream ships, and the comparewatch- plane row flips with it. A dark tag means the organ file EXISTS but does not declare the contracted symbol: something shipped there under another name, and until the contract is repointed to the real entry point (the plan rung and this row) or the function is renamed, that capability is invisible to this board — a build lost to darkness, named so it is not. The flip is necessary, not sufficient: it proves the symbol exists, never that the capability is good. The bar is the rung's pre-declared done-rule, proven by its gate — a symbol shipped without the behaviour behind it is a defect, and the flip is exactly what makes that defect visible instead of quiet. Competitor marks record documented capability presence — presence, not depth or scale. Adoption is measured too: every measured row carries where its organ stands on the estate's ladder (source → built → promoted → registered → invoked; libraries by importer reach minus validation importers; gates by the execution surfaces that run them). A row is fully adopted only at the top of its ladder; anything short is tagged partial with the exact remedy, so a build nobody promoted can no longer read as shipped. Census stamps: importers asof 1789497476, gate census asof 1789498715 (unix seconds; -1 = census absent).

Capability matrix — measured against source

◉ leads / measured exceed● present◐ partial○ absent · click any capability for its evidence

CapabilityNishiCursorSWE-agentQwen2.5-CoderAider
Verified context pack (organ-emitted, regenerable)Measured: fc_emit exists in runtime/nx_forge_ctx.nx, verified at emit. Cursor codebase-index [cursor-indexing] + Aider repo-map [aider-repomap] are the bar; ours is vetted+auto-regenerated from sources (gate 6/6) Adoption: LIB-WIRED importers=7 nonval=4 — fully adopted (top of its ladder).
Pros Nishi has it, measured on disk; ahead of Qwen2.5-Coder; fully adopted on the estate ladderCons behind Cursor (leads)
●◉●○●
Deterministic build + byte-exact judgeMeasured exceed: fg_run in runtime/nx_forge.nx, verified at emit. EXCEED: integer toolchain -> GREEN is a PROVABLE equality, noise-free repair signal; all named tools rely on flaky test suites [sweagent24] Adoption: LIB-WIRED importers=8 nonval=5 — fully adopted (top of its ladder).
Pros Nishi leads, a measured exceed; ahead of Cursor, SWE-agent, Qwen2.5-Coder, Aider; fully adopted on the estate ladderCons none on the measured axes (rival marks are documented presence, not depth)
◉◐◐○◐
Bounded repair loop (verbatim compiler stderr)Measured: fg_bank exists in runtime/nx_forge.nx, verified at emit. ACT+Debugger/Reflexion/Self-Debug [selfdebug23] are the field (57->65 HumanEval); ours gate-proven 5/5 with a real parse-trap repair Adoption: LIB-WIRED importers=8 nonval=5 — fully adopted (top of its ladder).
Pros Nishi has it, measured on disk; ahead of Qwen2.5-Coder; fully adopted on the estate ladderCons none on the measured axes (rival marks are documented presence, not depth)
●●●○●
Self-retiring vetted-rules registryMeasured exceed: pv_run_witness in runtime/nx_pack_rules_vet_gate.nx, verified at emit. EXCEED: each language restriction carries a witness probe + measured WHY, auto-RETIRES when the compiler heals it; no prompt tool measures its own rules Adoption: GATE:BUILT-UNPROMOTED trial=- — PARTIAL: compiled, never promoted to the serving root: /api/promote it, then roster it.
not adopted: compiled, never promoted to the serving root: /api/promote it, then roster it
Pros Nishi leads, a measured exceed; ahead of Cursor, SWE-agent, Qwen2.5-Coder, AiderCons not adopted yet: compiled, never promoted to the serving root: /api/promote it, then roster it
◉○○○○
Model backend (serve /gen + fenced extraction)Measured: fm_post_gen exists in runtime/nx_forge_model.nx, verified at emit. Table stakes; ours proven over a live loopback round-trip + model->candidate->judge GREEN rival marks uncited — the Yes, Best or Part codes on this row are an observation read with no reference mark behind them Adoption: LIB-WIRED importers=3 nonval=2 — fully adopted (top of its ladder).
Pros Nishi has it, measured on disk; fully adopted on the estate ladderCons none on the measured axes (rival marks are documented presence, not depth)
●●●●●
Local sovereign model tier loaded bits-upMeasured exceed: load_named_q16 in runtime/nx_nofloat_llm.nx, verified at emit. EXCEED (sovereignty): Coder-0.5B AND Coder-1.5B q8_0 (sha=ETag verified) [qwen25coder24] both run through our own no-float integer stack via the arch-config core, deterministic; 1.5B measured 8-10/10 on the inversion ruler (0.5B 1/10) -- scale is the capability lever Adoption: LIB-WIRED importers=33 nonval=8 — fully adopted (top of its ladder).
Pros Nishi leads, a measured exceed; ahead of Cursor, SWE-agent, Qwen2.5-Coder, Aider; fully adopted on the estate ladderCons none on the measured axes (rival marks are documented presence, not depth)
◉○○◐○
Requirements dossier evidence-forked + liar-killedMeasured exceed: rq_row in runtime/nx_forge_req_gate.nx, verified at emit. EXCEED: 14 requirement rows each fork their evidence on disk + a negctl; the census grades itself, cheat-proof by construction Adoption: GATE:LIVE trial=- — fully adopted (top of its ladder).
Pros Nishi leads, a measured exceed; ahead of Cursor, SWE-agent, Qwen2.5-Coder, Aider; fully adopted on the estate ladderCons none on the measured axes (rival marks are documented presence, not depth)
◉○○○○
Best-of-N sampling against the free verifierDARK — runtime/nx_forge.nx EXISTS but does not declare fg_best_of_n: something shipped at this path under another name, and this mark would read open forever. Repoint the contract to the real entry point (the plan rung AND this row) or rename the function; the flip follows on the next beat, and the comparewatch- plane row reads DARK until then. Large-Language-Monkeys [llmonkeys24]: 15.9->56% SWE-bench-Lite [swebench23] at 250 samples WITH a verifier; ours is perfect+free but the loop is not built yet
watching fg_best_of_n
Pros none measured yetCons behind Cursor, SWE-agent, Aider; open contract, nothing on disk yet
○●●○●
Chat-template / instruct promptingMeasured: nsv_chatml_ids exists in runtime/nx_nofloat_serve_core.nx, verified at emit. Qwen chat template is native to the model; ours PROVEN 07-15 (ChatML special-token splice into the no-float serve, byte-identical, L4 ran with it; SSE slot-collision fixed) rival marks uncited — the Yes, Best or Part codes on this row are an observation read with no reference mark behind them Adoption: LIB-WIRED importers=16 nonval=9 — fully adopted (top of its ladder).
Pros Nishi has it, measured on disk; fully adopted on the estate ladderCons behind Qwen2.5-Coder (leads)
●●●◉●
Batched prefill (throughput)DARK — runtime/nx_nofloat_serve_core.nx EXISTS but does not declare nsv_prefill_batch: something shipped at this path under another name, and this mark would read open forever. Repoint the contract to the real entry point (the plan rung AND this row) or rename the function; the flip follows on the next beat, and the comparewatch- plane row reads DARK until then. Per-token prefill; every serving stack batches -- our N6 keystone. NOTE 07-15: a SEPARATE wall (tokenizer O(n^2) linear-scan of 151k merges/pair, ~40min on a 4KB prompt) was ROOT-FIXED to hash+incremental-rank BYTE-EXACT; batched prefill itself still absent rival marks uncited — the Yes, Best or Part codes on this row are an observation read with no reference mark behind them
watching nsv_prefill_batch
Pros none measured yetCons behind Cursor (leads), SWE-agent (leads), Qwen2.5-Coder (leads), Aider; open contract, nothing on disk yet
○◉◉◉●
Grammar-constrained decoding (logit mask)DARK — runtime/nx_nofloat_serve_core.nx EXISTS but does not declare nsv_grammar_mask: something shipped at this path under another name, and this mark would read open forever. Repoint the contract to the real entry point (the plan rung AND this row) or rename the function; the flip follows on the next beat, and the comparewatch- plane row reads DARK until then. Outlines/XGrammar [xgrammar24] (100x, near-zero overhead); sovereign-exclusive for us since only our OWN serve can mask logits to the NishiLang grammar
watching nsv_grammar_mask
Pros none measured yetCons behind Cursor, SWE-agent, Qwen2.5-Coder; open contract, nothing on disk yet
○●●●○
Own-model training (compiler-filtered data)Measured exceed: nfa_backward in runtime/nx_nofloat_autograd.nx, verified at emit. EXCEED (determinism): 100% integer no-float train+infer is bit-exact reproducible; HONEST scale = toy (modelwright 476); the forge ledger is the compile-verified training set rival marks uncited — the Yes, Best or Part codes on this row are an observation read with no reference mark behind them Adoption: LIB-WIRED importers=47 nonval=7 — fully adopted (top of its ladder).
Pros Nishi leads, a measured exceed; ahead of Cursor, SWE-agent, Aider; fully adopted on the estate ladderCons none on the measured axes (rival marks are documented presence, not depth)
◉○○◉○
Code retrieval + reranking (CoREB)Measured: ci_query_top1 exists in runtime/nx_code_index.nx, verified at emit. NOW HAVE 07-15: nx_code_index semantic organ retrieval (jina-recipe arXiv:2508.21290 [jinacode25], NXCI1 sovereign store, gated 5-tooth incl cross-request determinism + corrupt-store-refused); base-weight quality MEASURED 3/6 top-1 (ratchet, behind trained rerankers); e5 contrastive tune (llm2vec recipe) = the quality closer Adoption: LIB-WIRED importers=4 nonval=2 — fully adopted (top of its ladder).
Pros Nishi has it, measured on disk; fully adopted on the estate ladderCons behind Cursor (leads), Qwen2.5-Coder (leads)
●◉●◉●
Sovereign code embeddings, external-oracle validatedMeasured: nsv_embed exists in runtime/nx_nofloat_serve_core.nx, verified at emit. NEW 07-15: our 100%-integer last-token-pooled embed path RAN jina-code-embeddings-0.5b [jinacode25] (Qwen2.5-Coder backbone [qwen25coder24], cc-by-nc ORACLE) and REPRODUCED its bf16 2x2 signature 822/176/122/617 vs oracle 817/124/120/553 pm (matched cells within <=5pm; tuned margins 646/495 vs base 72/49) -- a sovereign-stack result no API embedder exposes; Qwen col=2 (jina fine-tuned ON that backbone = the bar) Adoption: LIB-WIRED importers=16 nonval=9 — fully adopted (top of its ladder).
Pros Nishi has it, measured on disk; ahead of Cursor, SWE-agent, Aider; fully adopted on the estate ladderCons behind Qwen2.5-Coder (leads)
●○○◉○
Repository-level / multi-file editingDARK — runtime/nx_forge.nx EXISTS but does not declare fg_multi_file: something shipped at this path under another name, and this mark would read open forever. Repoint the contract to the real entry point (the plan rung AND this row) or rename the function; the flip follows on the next beat, and the comparewatch- plane row reads DARK until then. SWE-bench issue resolution [swebench23] is the frontier; forge is single-file organ scope today
watching fg_multi_file
Pros none measured yetCons behind Cursor (leads), SWE-agent (leads), Qwen2.5-Coder, Aider (leads); open contract, nothing on disk yet
○◉◉●◉
Test / oracle generationDARK — runtime/nx_forge.nx EXISTS but does not declare fg_oracle_gen: something shipped at this path under another name, and this mark would read open forever. Repoint the contract to the real entry point (the plan rung AND this row) or rename the function; the flip follows on the next beat, and the comparewatch- plane row reads DARK until then. The field generates tests/oracles; ours are hand-authored gates -- the generator is a future rung rival marks uncited — the Yes, Best or Part codes on this row are an observation read with no reference mark behind them
watching fg_oracle_gen
Pros none measured yetCons behind Cursor, SWE-agent (leads), Qwen2.5-Coder, Aider; open contract, nothing on disk yet
○●◉●●
Retrieval quality closer (contrastive tune on the compile-verified ledger)DARK — runtime/nx_code_index.nx EXISTS but does not declare ci_train_contrastive: something shipped at this path under another name, and this mark would read open forever. Repoint the contract to the real entry point (the plan rung AND this row) or rename the function; the flip follows on the next beat, and the comparewatch- plane row reads DARK until then. The retrieval row above is PRESENT at base-weight quality (3 of 6 top-1, a ratchet); trained rerankers lead -- the named closer is the e5/llm2vec-style contrastive tune over the forge's own verified pairs, through the no-float autograd, so the ruler moves by measurement not by swapping in a foreign model rival marks uncited — the Yes, Best or Part codes on this row are an observation read with no reference mark behind them
watching ci_train_contrastive
Pros none measured yetCons behind Cursor (leads), SWE-agent, Qwen2.5-Coder (leads), Aider; open contract, nothing on disk yet
○◉●◉●
External leaderboard standing (Aider-Polyglot run)DARK — runtime/nx_forge.nx EXISTS but does not declare fg_polyglot_bench: something shipped at this path under another name, and this mark would read open forever. Repoint the contract to the real entry point (the plan rung AND this row) or rename the function; the flip follows on the next beat, and the comparewatch- plane row reads DARK until then. HONEST STANDING: zero external-suite results today; little-coder (2026) proved sub-10B is benchmarkable on Aider-Polyglot (19 to 45 percent by scaffold alone, N=225) -- the contract is a sovereign runner over the public task set with the byte-exact judge, publishing pass rates the field can compare rival marks uncited — the Yes, Best or Part codes on this row are an observation read with no reference mark behind them
watching fg_polyglot_bench
Pros none measured yetCons behind Cursor (leads), SWE-agent (leads), Qwen2.5-Coder (leads), Aider (leads); open contract, nothing on disk yet
○◉◉◉◉
Rival-claim provenance. Every Yes, Best or Part mark in a rival column is a claim about someone else's product. rows with rival marks 16 · cited 9 · uncited 7. A cited row carries a reference mark that resolves to a pinned mirror (the refs gate measures that); an uncited row is an observation read and is badged in its evidence until a reference lands. The badge is the receipt, never the proof: a mark proves a mirror exists, not that the mirror supports the code.

Delivery and evidence

Inspect risks, technical debt, rendered observations, experiments and references. Read scope and limitations alongside every result.

Risk register

RiskLikelihood x impactMitigation
Best-of-N on a model that pattern-copies (0.5B) buys nothing -- scale is the lever the record already measuredlikely x mediumCD2 reports per-tier; the 1.5B tier (8-10 of 10 on the inversion ruler) is the default subject.
A retrieval tune that improves the training ruler and not the held-out one is overfitting wearing a winpossible x highCD4 done-rule is pre-declared on the held-out ruler; a miss stays UNWIRED as the BEIR arm did.
A leaderboard number taken once under a favourable scaffold is a sample, not a standingpossible x mediumCD6 publishes N and the scaffold hash with every number; re-run on the beat.
On these two registers. Rows are declared in the domain's plan file and carry the debt id, which is the join key back to the sovereign debt plane — that plane, not this page, is the authority on state. Reconciling them automatically (the regen reading the plane and refreshing these rows) is a named, owed rung; until it lands, treat an id here as a pointer to look up, not a status to trust.
Honest verdict. Nishi forge is a REAL sovereign coding harness -- organ-emitted verified context pack, a DETERMINISTIC build+byte-exact judge, a bounded repair loop feeding verbatim compiler stderr, a self-retiring vetted-rules registry, a local integer no-float Coder tier (0.5B AND 1.5B) loaded through its own stack, and now a sovereign semantic code-retrieval organ (nx_code_index) -- all gate-proven and evidence-forked. It is <b>not</b> yet at frontier CAPABILITY: the local 0.5B executes FORM not reasoning (L4 mini-pack candidates now COMPILE+RUN but pattern-copy the exemplar -> byte-judge 0/3; scale is the lever -- the 1.5B tier measured 8-10/10 on the independent inversion ruler), and best-of-N sampling, grammar-constrained decoding, batched prefill, repository-level editing, and own-model training at scale remain absent or partial. Retrieval quality is base-weight (3/6 top-1, a ratchet) -- present but behind trained rerankers until the e5 contrastive tune. The genuine <b>exceeds</b> are mechanism-clear: (1) a deterministic integer toolchain makes the verify signal NOISE-FREE (byte-exact GREEN is a provable equality; flaky-test harnesses like SWE-agent/Cursor cannot) -- INDEPENDENTLY CONVERGED ON by codefactory/llmfwbench's proof-by-sabotage; (2) the vetted-rules registry MEASURES each language restriction and auto-retires it when the compiler heals (no prompt tool measures its own rules); (3) NEW -- our 100%-integer embedding path REPRODUCES a 3rd-party SOTA model's signature (jina-code-embeddings bf16 2x2 817/124/120/553 pm matched within <=5pm on the matched cells), a sovereign-stack result no API embedder can claim. EXTERNAL CORROBORATION of the whole thesis: little-coder (2026) moved a matched Qwen3.5-9B from 19.11% to 45.56% on Aider-Polyglot by varying ONLY the scaffold (N=225), confirming scaffold-model fit is first-order and that sub-10B is benchmarkable -- their retry-with-failing-output lever (+18% of passes) is our next adoption. HONEST STANDING: AHEAD on judge determinism + ruler validity; BEHIND on external-leaderboard numbers (zero suite results -> stand on Aider-Polyglot next), retrieval quality, and repo-level tasks. The climb: 1.5B/best-of-N against the free verifier, e5 retrieval tune, batched prefill, then own-model training on the compile-verified ledger.

Person · product · place — not yet measured for this domain

Every compare carries this layer. Declare knowledge/compare/coding.ppp (rows surface|nishi or c1..c4|label|url|connect naming OUR live surface and each rival's front door), run nx_ppp_probe domain coding, and this section fills itself on the next beat: the same ruler on both sides — privacy and CX (third-party hosts, tracker classes, cookies, security headers), design and longevity (design hygiene, computed WCAG contrast, render-blocking resources, unsized media, script weight, theme and motion queries), findability (landmarks, skip link, on-site search, breadcrumb, headings, internal links).

The field — discovered, not chosen

Rows written by nx_field_discover from coding.seeds: the industry's own lists (Wikipedia wikitext, GitHub topics, awesome lists) read mechanically, every candidate counted across seeds. The matrix columns above are a SEAT'S pick; this band is the population they were picked from, and the stats line measures one against the other. A rival here is a lead, never a verdict — it earns a column when its capabilities are read and pinned.

field|seeds=4|scanned=4|fetched=4|reused=0|failed=0|named=4|candidates=4|mentions=4|capped=0 rival|Cursor|1|1|col0-cursor|https://cursor.com/blog/secure-codebase-indexing|named rival|SWE-agent|1|1|col1-swe|https://arxiv.org/abs/2405.15793|named rival|Qwen2.5-Coder|1|1|col2-qwen2|https://arxiv.org/abs/2409.12186|named rival|Aider|1|1|col3-aider|https://aider.chat/docs/repomap.html|named
RankRivalSeedsMentionsFirst seedKindLink
1Cursor11col0-cursornamedhttps://cursor.com/blog/secure-codebase-indexing
2SWE-agent11col1-swenamedhttps://arxiv.org/abs/2405.15793
3Qwen2.5-Coder11col2-qwen2namedhttps://arxiv.org/abs/2409.12186
4Aider11col3-aidernamedhttps://aider.chat/docs/repomap.html

field candidates 4|shown 4 of 4|matrix columns in the field 4 of 4|discovered rivals with no column 0|malformed rows 0 (counted, never rendered)|read-capped 0

Gaps from the record — what the estate does that no board carries

The record census (nx_goalmap record) reads the invoked-tool population and every plan queue row and files each organ or directive that NO matrix, plan or gates row names. A row here is a callout the boards missed: adjudicate it onto a board or declare it infrastructure. Census state BLIND (age 74396 s), sources read 4 of 7 declared — a BLIND census is a FLOOR: unread sources can only add rows.

kindnameboardsourceevidence

rows shown 0|this board's directives 0|estate-wide un-boarded organs 492 (listed in full on /compare/ecosystem)|census rows 611|malformed 0 (counted, never rendered)

References

Beyond a link list. Every reference below resolves twice — the publisher's copy and, where banked, the estate's own non-rottable library mirror with a content pin — and carries its evidence class plus the exact claim on this page it grounds. Keyed marks like [key] in the matrix notes jump here. A dash means honestly absent, never assumed.
  1. [swebench23] Jimenez, Yang, Wettig, Yao, Pei, Press, Narasimhan. SWE-bench: Can Language Models Resolve Real-World GitHub Issues? ICLR 2024 (arXiv:2310.06770). publisher · read in our library knowledge/fetched/cmp_coding_swebench23.html · pin hcd170cab3a56f891ee94070d5748d923685f2a84a91201d6336a221436fd0f9f · accessed 2026-08-18 · published-paperGrounds: The "Repository-level / multi-file editing" row: SWE-bench issue resolution over real GitHub repositories is the frontier task that row names, and the reason forge's single-file organ scope grades 0 against 2/2/1/2; it is also the benchmark behind the SWE-bench-Lite figure in the "Best-of-N sampling against the free verifier" row.
  2. [sweagent24] Yang, Jimenez, Wettig, Lieret, Yao, Narasimhan, Press. SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering. NeurIPS 2024 (arXiv:2405.15793). publisher · read in our library knowledge/fetched/cmp_coding_sweagent24.html · pin h1589f706b0380d637196a65970b23cb89037465fe6d883af1007e34a41d8bdbf · accessed 2026-08-18 · published-paperGrounds: The SWE-agent column (@cols): the agent-computer-interface harness those column codes describe, and the flaky-test-suite verify signal the "Deterministic build + byte-exact judge" EXCEED row contrasts against (a byte-exact GREEN is a provable equality, a test-suite pass is not).
  3. [qwen25coder24] Hui, Yang, Cui, Yang, Liu, Zhang et al. (Qwen Team). Qwen2.5-Coder Technical Report. arXiv:2409.12186, 2024. publisher · read in our library knowledge/fetched/cmp_coding_qwen25coder24.html · pin h449fe8f906c0ebde1f5c7f7957692a28dff03a1a935e12882a0767c271555390 · accessed 2026-08-18 · published-paperGrounds: The Qwen2.5-Coder column and the "Local sovereign model tier loaded bits-up" row: Coder-0.5B and Coder-1.5B are this model family, run q8_0 through the no-float integer stack; it is also the backbone named in the "Sovereign code embeddings, external-oracle validated" row.
  4. [llmonkeys24] Brown, Juravsky, Ehrlich, Clark, Le, Re, Mirhoseini. Large Language Monkeys: Scaling Inference Compute with Repeated Sampling. arXiv:2407.21787, 2024. publisher · read in our library knowledge/fetched/cmp_coding_llmonkeys24.html · pin h922be22acaa0999ba5690c50a680d29228db6c0e73aa6c89ffb769acb7cab68c · accessed 2026-08-18 · published-paperGrounds: The "Best-of-N sampling against the free verifier" row: the 15.9 to 56 percent SWE-bench Lite result at 250 samples with a verifier is this paper's headline number; the row grades our loop absent although our verifier is exact and free.
  5. [selfdebug23] Chen, Lin, Schaerli, Zhou. Teaching Large Language Models to Self-Debug. arXiv:2304.05128, 2023. publisher · read in our library knowledge/fetched/cmp_coding_selfdebug23.html · pin h967b70ab50990ac9fd716a27160f07d6d494283bdefdfa432a05d55e9b99af7c · accessed 2026-08-18 · published-paperGrounds: The "Bounded repair loop (verbatim compiler stderr)" row: Self-Debug is one of the three named field methods (ACT+Debugger, Reflexion, Self-Debug) that feed execution feedback back to the model; ours feeds verbatim compiler stderr under a bounded loop, gate-proven 5/5.
  6. [jinacode25] Kryvosheieva, Sturua, Guenther, Xiao. Efficient Code Embeddings from Code Generation Models (jina-code-embeddings). arXiv:2508.21290, 2025. publisher · read in our library knowledge/fetched/cmp_coding_jinacode25.html · pin he933c046f06596bf19f75f0acfe3e3d28c5b50a860227d76cf492f824637de49 · accessed 2026-08-18 · published-paperGrounds: The "Code retrieval + reranking (CoREB)" and "Sovereign code embeddings, external-oracle validated" rows: the jina-code-embeddings recipe (Qwen2.5-Coder backbone, last-token pooling) is the external ORACLE our 100 percent integer embed path reproduced within 5 permil on matched cells, and the arXiv id the matrix cites.
  7. [xgrammar24] Dong, Ruan, Cai, Lai, Xu, Zhao, Chen. XGrammar: Flexible and Efficient Structured Generation Engine for Large Language Models. arXiv:2411.15100, 2024. publisher · read in our library knowledge/fetched/cmp_coding_xgrammar24.html · pin h16539c51e47d62e3672b6fa34da90194c587a5708df9c346ef0ff5723c44c56f · accessed 2026-08-18 · published-paperGrounds: The "Grammar-constrained decoding (logit mask)" row: XGrammar (with Outlines) is the named engine whose near-zero-overhead logit masking sets the bar for the sovereign-exclusive NishiLang-grammar mask this row declares absent.
  8. [aider-repomap] Aider documentation: Repository map -- a concise tree-sitter derived map of the classes and functions in the whole git repository with their signatures, ranked to fit the context window. publisher · read in our library knowledge/fetched/cmp_coding_aider-repomap.html · pin h77708d7004b0755018668f7db831cb09b7cb5883da4e46d90bd65fe5683b07f0 · accessed 2026-08-18 · vendor-docGrounds: The Aider column and the "Verified context pack (organ-emitted, regenerable)" row: the Aider repo-map is one of the two named bars for that row; ours is vetted and auto-regenerated from sources rather than ranked from a symbol graph.
  9. [cursor-indexing] Cursor. Securely indexing large codebases (engineering blog, 2026-01-27): Merkle-tree change detection, syntactic chunking, embeddings computed asynchronously, per-codebase vector namespace with obfuscated paths. publisher · read in our library knowledge/fetched/cmp_coding_cursor-indexing.html · pin h150ddab0184554edda2844c71304792378f8034a3e88da5556912c1024d3640c · accessed 2026-08-18 · vendor-docGrounds: The Cursor column and the "Verified context pack (organ-emitted, regenerable)" row where the Cursor codebase-index is graded 2 (Best): this is the vendor's own description of that index; ours grades 0 as a regenerable vetted pack, not a semantic index of the whole tree.

generated by nx_swcompare_matrix (sovereign NishiLang organ) from knowledge/compare/coding.matrix · source checks show implementation presence; runtime and user-outcome evidence are reported separately · JavaScript supports page controls