Make the seed settable, pre-register its reading, and launch both seed-variance campaigns (#51, #135) - #155
Conversation
…ad to mean (#51, #135) Both issues are blocked on the same missing number. #51's matched-operating-point read put the RampNet-vs-YOLO residual at 0.039 F1, and #51's own rule is that differences under ~0.02 should not be read -- but every arm in that comparison is seed 0, the ultralytics default, in all eight args.yaml. #135's power analysis reached the same wall from the other side and said so outright: the binding limit is unmeasured seed variance, n=1. Nothing downstream can be called real without it, RampNet 2.0 included. Neither trainer could vary the seed at all. train.py hardcoded manual_seed(42) with no flag; run_yolo_train_tillicum.slurm never passed one to ultralytics. THE HALF THAT WAS EASY TO GET WRONG. Stage 2 has TWO sources of run-to-run randomness, and only one was ever set. DistributedSampler carries its own seed (default 0) and derives each epoch's permutation from seed + epoch in set_epoch(), independent of torch.manual_seed. A sweep that moved only the torch seed would reuse one data order across every replicate, understating the spread -- and would do it silently, since no log line distinguishes the two cases. So both seeds move together, with one exact exception: at the historical seed 42 the sampler must stay at its historical 0, or the DEFAULT stops reproducing published runs. That pairing is rampnet/seeding.py::sampler_seed_for, tested rather than commented. Also load-bearing, and the reason the klone launcher is a new file rather than a flag on run_train.slurm: train.py writes best_model.pth and latest_checkpoint.pth to the CURRENT DIRECTORY, not to --checkpoint-dir. Three replicates launched from one directory would overwrite each other's best model and, worse, each other's resume state -- a resume file from another seed loads silently as if it were this run's own, converging the arms onto one lineage with nothing in the log to say so. run_train_seed.slurm gives each replicate its own RUNDIR and cd's into it. PRE-REGISTERED, before any replicate finished: docs/seed_variance_51_135.md fixes the decision rule. SD <= 0.010 and the architecture advantage is real at ~4 sigma; SD >= 0.020 and #51 closes with "the supervised baseline is statistically indistinguishable from RampNet at matched operating points"; between is ambiguous and needs both campaigns. The second branch makes our own headline smaller and is accepted in advance. One deliberate deviation from the #51 protocol: save_period=1. The arms ran -1, which is exactly why the epoch-curve follow-up had to be retracted -- no per-epoch weights exist anywhere and cannot be recovered. ~20 GB against 1 TB buys back that analysis. tests/test_seeding.py (21) checks the plumbing at the source level, because every failure here is silent: the default still mapping to sampler seed 0, all three RNGs following the flag, the launcher refusing to default SEED, per-seed RUNDIR isolation, and the Tillicum heredoc's positional list still lining up with its unpack -- an off-by-one there shifts imgsz into epochs and trains a wrong model that finishes green. Full suite: 1,343 passed, 1 skipped. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
… B is decided against (#135) #135 exists to decide whether to spend 1,675-3,350 GPU-hours on the annealed arm. Every input to that decision has been in place since 2026-08-29 -- the rung completed, the LR schedule verified over 100% of the run, manual_gold scored and tied at every epoch -- but the gate written into the pre-registration was never actually applied to the numbers, so the issue stayed open with the decision unmade. scripts/analysis/run_b_gate_135.py applies it from the committed summaries rather than restating it in prose, and writes docs/data/run_b_gate_135.json: PRIMARY max-F1(cosine ep8) - max-F1(Run A ep8) = +0.002994 |z| = 1.02 to 1.86 -> not significant SECONDARY Run A ep3->ep8 -0.006624 ; cosine -0.005158, arrested? NO difference of declines +0.001466, |z| = 0.50 to 0.91 -> not significant VERDICT JUDGMENT CALL The primary misses significance AT THE FAVOURABLE END of the measured s.e. bracket, so the reading does not depend on which value inside it is chosen. That robustness is the reason the bracket is honest enough to decide on: a true paired bootstrap between the two arms needs both arms' per-pano detections and the cosine arm's are not committed, so this uses the s.e. #138 measured across 28 Run A epoch pairs on the same panos and GT. Stated in the script's own docstring, not buried. The pre-registration says a tie on both is explicitly NOT an automatic cancellation, so the decision is recorded as the judgment it is. Not running Run B, because: - The mechanism that justified it is the one that failed. Run B's case was #51's annealed tail; the rung tested exactly that at matched budget, seed and data order and moved manual_gold by nothing measurable. A 30-epoch arm changes length AND schedule, so it could not attribute a difference even if it found one. - The gain is real and in the wrong place: up to 3.98% better auto-label val loss, none of it reaching human F1. That is #84's exchange rate replicating. - THE ARGUMENT THAT WAS NOT AVAILABLE WHEN RUN B WAS SPECIFIED: Run B is n=1, and this issue itself established seed variance as the binding limit. The plausible effect (~0.003 of late-epoch damping) sits below the ~0.01 that is "measured but not attributable" without a seed control, so a single 30-epoch run is uninterpretable at any length. A readable Run B is 3 seeds = 5,025-10,050 GPU-h, ~$4,500-9,000. - Opportunity cost, measured: #151 just produced a +0.115 F1 rig effect on the same model. What is NOT claimed is written down too: not that annealing does nothing (the ~0.003 damping is unresolved, not refuted), not that a 30-epoch run would fail (it was never run), and not epoch 7's +0.0042 -- the largest gap anywhere, which WOULD clear 1.96 at the favourable end of the bracket but is not the pre-registered comparison. The artifact flags it so nobody quotes it as the result. Reopening condition is concrete and already in flight: the seed campaign (PR #155, klone 39515025/26/27) prices it directly. Seed SD <= ~0.002 max-F1 and a 0.003-0.008 effect becomes readable at n=1. Two amendments so the repo does not carry contradictory advice: - stage2_run_b_power_135.md's "run the 30-epoch arm" recommendation predates the rung's results; marked SUPERSEDED in place rather than rewritten. - stage2_epoch_curve_84.md's "the gate does not cancel Run B" section gets the later outcome inline, since that doc is where a reader looks for Run A/Run B. Also states a deliberate omission: #135 asked for the benchmark splits as well as manual_gold, and only manual_gold was scored -- because #138 measured the nine city splits pooled at an unpaired MDE of 0.0219 against manual_gold's 0.0117, so they cannot resolve what manual_gold cannot. tests/test_run_b_gate_135.py (10) covers the branches that did NOT fire, since an implementation that only ever emits the observed verdict is untestable by its own output. Full suite: 1,371 passed, 1 skipped. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…ncher in CI (#51, #135) Jobs 39515025, 39515026 and 39515027 were submitted on 2026-09-03 at 15:01 and were all dead one second later, FAILED with exit 2. The stdout files are zero bytes; the whole record of the failure is 99 bytes of stderr: slurm_script: line 46: unexpected EOF while looking for matching `}' Line 46 was the required-SEED guard, whose message read "the replicate's seed". Inside ${VAR:?...} bash parses the word for quoting even within double quotes, so the lone apostrophe opened a quote and the closing brace was never found. Removing it is the entire fix; the guard still aborts on an unset SEED, which is tested rather than assumed. WHAT IT COST. Nothing in dollars -- klone ckpt is free -- and everything in calendar. The free half of the campaign produced no GPU-seconds for 27 hours while the paid Tillicum half ran normally beside it, so the asymmetry pointed the wrong way: the arm that measures RampNet's own seed spread, which #135 named as the binding limit, is the one that had not started. /gscratch/scrubbed/jfroehli/seedvar/ did not exist. WHY NOTHING CAUGHT IT. Nothing in the suite reads these files as shell. test_seeding.py asserts on their content with regexes, which a syntactically broken script passes happily, and a launcher is the one artifact whose failure is invisible locally -- you learn about it on the cluster, hours later, from an empty log. tests/test_slurm_scripts.py now runs bash -n over every tracked .slurm (23 of them, ~1.5 s, no cluster and no network) and fails on exactly this. Verified by reintroducing the apostrophe: two tests fail, and they name the file. The glob prunes dot-directories deliberately -- .claude/worktrees/ holds whole nested checkouts, and collecting their launchers would scale this test with however many branches happen to be on disk. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Campaign B never started: all three klone replicates died one second after launchChecked the cluster 2026-09-04 ~19:30 UTC, 28 hours after launch. Campaign A (Tillicum, paid) is healthy. Campaign B (klone, free) produced nothing.
The stdout files are zero bytes. The entire record is 99 bytes of stderr: Line 46 was the required-SEED guard, and the message read Fixed in 2b7471c: the apostrophe is gone, the guard still aborts on an unset SEED (tested, not assumed), and Resubmitted from the updated checkout at seeds 1/2/3: Campaign A is fine, and its TIMEOUTs are by design
What this cost, and the asymmetry worth notingNothing in dollars — klone 🤖 Generated with Claude Code (claude-opus-5[1m]) |
Review of the whole diff (first review of this PR)The plumbing is right where it counts: both RNG sources move together, the historical What I do not think holds up is the guarding and the reading. Two of the three Nothing here argues for touching the in-flight jobs. HighH1. Verified: I deleted H2. Concretely: at H3. MediumM4. M1. M2. M3. M5. M6. M7. LowL1. L2. L3. L4. L5. L6. L7. Not findings, checked and cleared
|
…ding rule (#51, #135) Review of PR #155. Three findings verified by mutation, all silent. The test that guards the campaign's central invariant did not guard it. test_train_sampler_seed_is_derived_not_hardcoded asserted "sampler_seed_for(args.seed)" in src against the whole file, and train.py already contains that string in a log line. Deleting seed= from the DistributedSampler call -- which is exactly the regression rampnet/seeding.py exists to prevent, and which would make all three replicates share one data order -- left the suite green. It is now asserted on the parsed statement. Two smaller versions of the same mistake: "random.seed(args.seed)" in src is satisfied by np.random.seed(args.seed), so the stdlib call that drives the horizontal-flip augmentation could be deleted unnoticed; and the Tillicum positional-argument count matched only "$VAR", so inserting "${VAR}" mid-list shifted every later field by one and passed. Both now fail as intended. The pre-registered decision rule divided 0.039 by Campaign A's SD alone. That gap is a difference between two single-seed runs, so its standard error is sqrt(s_A^2 + s_B^2). At s_A = 0.010 the table read 4 sigma; with s_B = 0.010 it is 2.8, and with s_B = 0.020 it is inside the band the same table calls indistinguishable -- the conclusion could invert with no change to any input the rule looked at. Campaign B was measuring s_B concurrently and had no reading of its own at all, leaving half the campaign free to be interpreted afterwards. Amendment 1 corrects the sigma, makes the bands disjoint at their endpoints, gives Campaign B a rule against #135's measured 0.0063 paired MDE, and says what happens if B does not finish in time. The original table is kept verbatim. No result from either campaign had been scored when this was written. Also: the doc's headline number, its source document and the script that computes its statistic all live on PR #154, not on main, which is now stated rather than left as a dead link; neither campaign had a written path from its checkpoints to the statistic, and both gaps are named; the Campaign A reproduce command omitted the PYTHON= that Tillicum needs; and Campaign B writes its only artifact to /gscratch/scrubbed, which purges on a ~21-day idle window while the same document says the campaign's calendar is unbounded. Launcher changes are additive only, since both campaigns are in flight: an optional RAMPNET_ENV escape hatch (default path unchanged), a world-size warning, and header notes. The RUNDIR default is deliberately NOT moved off scrubbed -- that would put a resubmission somewhere different from its queued siblings, so the required copy-out is documented instead and the move left as a decision. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Review fixesPushed as 455e1c4. Full suite FixedH1 — H2 / H3 / L1 — the reading rule. Added Band edges and the Campaign B rule are the one place I made a scientific call rather M1 — added a "Where the inputs live" section naming PR #154 as the home of M2 — A1.3, above. M3 — the Campaign A reproduce block now carries M4 — documented, not changed. The launcher header and a new limitations bullet spell M5 — the positional count now matches M6 — M7 — documented in L2 — "best-val" now names L3 — cost corrected: ~$130/replicate as launched ( L4 — L5 — the launcher header now says L6 — the three RNG assertions are line-anchored. Re-ran the mutation: deleting the bare L7 — added the Needs your decision — would affect in-flight runs, so not changedM4, the H2/H3, the amended band edges — flagged above. L3, Skipped as not real
🤖 Generated with Claude Code (claude-opus-5) |
…e bundle truncation travels with its numbers (#135) Two review passes on this PR (2026-08-18 and 2026-09-03) left thirteen findings open, and the gate commit 4171a6e added an artifact that contradicts the document it sits beside. This is the fixes pass those reviews said would follow. THE HEADLINE CLAIM WAS READ AGAINST THE WRONG BAR The pre-registration says |delta| / s.e. >= 1.96 per pair. The Results section instead compared the largest delta to #138's MDE of 0.0063, which is 2.80 x s.e. at 80% power -- 1.43x looser -- and concluded "tied at every epoch". Those are different bars: an effect below the MDE is not thereby non-significant. run_b_gate_135.json, added in 4171a6e, already records the consequence: epoch 7's +0.0042 gives |z| = 1.45 to 2.64 over the measured s.e. bracket, and the artifact flags it as clearing 1.96 at the favourable end. The accurate statement is narrower and is now what the document says. The pre-registered primary (epoch 8, +0.0030, |z| = 1.02 to 1.86) is a tie at both ends of the bracket -- which is the reading the decision rests on and which does not depend on which s.e. inside it is chosen. Six of the other seven epochs are ties at both ends. Epoch 7 is undetermined, not a tie, and resolving it needs the cosine arm's per-panorama detections, which are not committed. The deviation from the pre-registration is stated where the numbers are, along with the fact that the pre-registration's own "Exact commands" could not have run the test as written. THE DECISION IS UNCHANGED, AND WAS NEVER LOAD-BEARING ON THE TIE Run B at n=1 cannot be told apart from a seed draw at any length. That argument holds whether or not epoch 7 separates, which is why overstating the tie bought nothing. THE NINE CITY BUNDLES ARE CUT AT 0.55, AND EVERY UNPAIRED ROW INHERITS IT Measured: all nine hold 0.0% of their detections below 0.55 (minima 0.5501-0.5607) while manual_gold reaches 0.0501 with 28.8% below. So "the #54 operating point of 0.30" is a no-op on nine of ten splits, max-F1 there peaks on an already-truncated curve (max_f1 == f1 exactly on eight of nine cities and on POOLED cities), and the pooling gain is understated. Measured against the one untruncated arm in committed data (--reference rampnet_1pass, all ten splits to 0.05): POOLED-all MDE 0.0105 against manual_gold's 0.0121, a 14% gain rather than 7%. That arm is single-pass and missing seam detections, so 14% bounds the correction rather than being it; there is no clean uniform-0.30 RampNet arm in the repo, which is now stated beside the number. The conclusion is unchanged on either reading: pooling is not a lever. The artifact now records reference_min_confidence and protocol_threshold_binds per split and truncated_members per pooled row, so this is visible rather than inferable -- it sat undetected through two reviews precisely because the only evidence was indirect. benchmark_power_135.json regenerates with ZERO changed values: the additions are new keys plus self_pair's max-F1, which is now null. Every number the documents quote still stands. REST OF THE FIX LIST - self_pair max-F1 was identically zero by construction (max-F1 re-picks its own threshold, so shifting the read-out point cannot move it). It read as a measured zero with zero uncertainty in all 24 rows; it is now null with a note. That block bounds F1. - The headline table's rows 1 and 4 were F1 standard errors under a max-F1 header (0.0042/0.0117 and 0.0039/0.0109). Now max-F1: 0.0041/0.0114 and 0.0039/0.0108. - Four-decimal drift against the artifact, all corrected: +0.1119 -> +0.1117, +0.1492 -> +0.1489, c = 123 -> 124, "9 of 356" -> "10 of 357", discordance ranges to 2.50-4.26 / 2.50-6.66 / 6.48-6.66, MDE upper 0.0081 -> 0.0082, and the rung doc's 0.961/3.979 -> 0.962/3.980 (the trap that file's own provenance note warns about). - The recall table printed six of the nine pairs the script computes, and the three it dropped included the only "not resolvable" verdict. All nine now print, and the table's verdicts carry the conditional the max-F1 table already had: they rest on a discordance range taken from three cross-detector pairs, none of them epoch-vs-epoch. - "Every derived number in this document is in benchmark_power_135.json" was false. Four classes of exception are now named, and test_committed_json_matches_the_doc_headline_ numbers pins every headline value so the prose cannot drift again. - dump_peaks_from_cache.py hardcoded a run_a_epoch_N label whatever --summary-csv it was given, so a second arm's dumps were either invisible to the reader or overwrote Run A's committed ones. Adds --label-prefix; --verify now fails loudly on a fingerprint the summary does not contain instead of silently checking nothing; exclude_border is read from the extractor rather than restated; the docstring no longer recommends the one output directory the same file's help text forbids. - stage2_epoch_curve_84.md's status block still stated the superseded curve shape. It now carries the #135 amendment: the plateau is 2-6, not 2-8. - The rung's pre-registration claimed nothing above Results had been edited; two commits had. Says what was edited and what was not. - Provenance gaps stated rather than implicit: the 21 restarts and 35.06 h come from sacct -D with no dump committed (a clean clone can count 18 event files, 11 with steps); the 560.9 GPU-hours are pending #147's compute_log.jsonl; the LR verification block gets its one-line reproduce command; whether the rung's eval cache survived is unknown and is now said so. - The reopening condition pointed at docs/seed_variance_51_135.md, which is not on this branch; it points at PR #155. #151's +0.115 F1 is marked as not re-derivable here. The "~0.01 measured but not attributable" threshold the decision's third reason turns on is marked as a working assumption rather than a measurement. - Run A's epoch-7 max-F1 differs between two documents in this PR (0.9110 from summary.csv, 0.9107 post-#140). Both arms were scored under the pre-#140 matcher so the comparison is internally consistent; that is now said, in both places. - Dead Scored.gt_pano removed. observed_and_se(paired=...) asserts shared panorama order, the way mcnemar already asserts its own. The test's (prediction_confidence(p) or -1e9) mapped a legitimate 0.0 confidence to -1e9; it checks for None. - Deliberately NOT changed, with the reason recorded in the code: the single Generator threaded through every group, and the redundant res_a/res_b bootstraps. Either change shifts the draw stream and moves every standard error these documents quote, for no gain in correctness. The --splits caveat is documented instead. - train.py's epoch-boundary checkpoint window is described in the ResumeSkipSampler docstring rather than fixed: latest_checkpoint.pth is written after validation, so a preemption there loses the whole pass (the committed events show it happening twice). Suite: 1,373 passed, 1 skipped. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
… about the wrapper #138 and #154 merged while this branch was open, and #138 rewrapped the train sampler as ResumeSkipSampler(DistributedSampler(...)). Three conflicts in stage_two/train.py, all additive on both sides except one that is genuinely semantic: train_sampler = ResumeSkipSampler( DistributedSampler(..., shuffle=True, drop_last=True, seed=sampler_seed_for(args.seed))) The seed has to sit on the INNER DistributedSampler. On the wrapper it would be accepted and inert -- the sweep would run with one data order across all three replicates and nothing in any log would say so, which is the exact failure this branch exists to prevent. The other two conflicts are the argparse block and the startup log line; both sides added different things and both are kept. Two tests then failed, and both were right to: - tests/test_seeding.py asserted the assignment was a DistributedSampler call. It is now a wrapper. Rather than loosen the assertion to a substring -- which is what made this test vacuous in the first place -- it walks into the wrapper via _unwrap_to() and fails loudly if the target is ambiguous or absent. - tests/test_resume_skip_sampler.py executes parse_args in a restricted namespace, which now needs HISTORICAL_SEED injected. Re-verified by mutation, both caught: deleting seed= from the inner sampler, and moving it to the wrapper where it would be inert. The second is a NEW failure mode that only exists because of this merge. Full suite: 1,481 passed, 1 skipped. bash -n clean on the launcher. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
An amendment to a pre-registration is only worth the audit trail attached to it. This records who ratified it and the one fact that decides whether it is legitimate: it was ratified on 2026-09-04, before any replicate from either campaign had been scored. Nothing about the rule changes. The cut points remain 0.010 and 0.020 as originally written; the amendment applies them to sqrt(s_A^2 + s_B^2) rather than s_A alone. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
1 participant
Makes the training seed settable in both trainers, pre-registers how its variance will be read, and launches the two campaigns that measure it. Campaign A has been running since 2026-09-03; Campaign B was relaunched 2026-09-04 after its first three jobs died one second after submission on a bash parse error (see the review comments below). Job IDs below.
Why now
#51 and #135 hit the same wall from opposite sides.
seed: 0— the Ultralytics default — in all eightargs.yamlfiles. Train a supervised YOLO baseline (YOLO11 / YOLO26) on the RampNet dataset — isolate architecture vs. data #51's own rule is that differences under ~0.02 should not be read.Neither trainer could vary the seed at all:
train.pyhardcodedmanual_seed(42)with no flag, andrun_yolo_train_tillicum.slurmnever passed one to Ultralytics.Everything downstream inherits this. With no noise floor, no RampNet 2.0 improvement can be called real either.
The half that was easy to get wrong
Stage 2 has two sources of run-to-run randomness and only one was ever set.
DistributedSamplercarries its ownseed(default 0) and derives each epoch's permutation fromseed + epochinsideset_epoch(), independent oftorch.manual_seed. A sweep that moved only the torch seed would reuse one data order across every replicate — understating the spread, and doing it silently, since no log line distinguishes the two cases.So both seeds move together, with one exception that has to be exact: at the historical seed 42 the sampler must stay at its historical 0, or the default stops reproducing published runs. That pairing is
rampnet/seeding.py::sampler_seed_for, tested rather than commented.Why the klone launcher is a new file
train.pywritesbest_model.pthandlatest_checkpoint.pthto the current directory, not to--checkpoint-dir. Three replicates launched from one directory would overwrite each other's best model and — worse — each other's resume state: a resume file from another seed loads silently as if it were this run's own, converging the arms onto one lineage with nothing in the log to say so.run_train_seed.slurmgives each replicate its ownRUNDIRandcds into it.run_train.slurmstays untouched as the preserved record of the published run and #135's rungs, the same reasoning that kept the Tillicum YOLO launcher separate from its klone original.Pre-registered, before any replicate finished
docs/seed_variance_51_135.mdfixes the decision rule on the sample SDsof the three Campaign A replicates:sThe second branch makes our own headline smaller and is accepted in advance.
What is running
Campaign A — YOLO seed variance, Tillicum. Three replicates of
y11x_tilesat seeds 1/2/3, config identical to the as-runargs.yamlin every respect except the seed. Jobs 274367 (running, verifiedseed=1in its log header), 274369, 274370, each withCHAIN=5to cover the 24 hnormal-QoS ceiling. ~3.0 h/epoch on one H200; ~$21.60 per 24 h link, ≤$389 for all three, against a $1,500/month cap with $0 used this cycle.Tillicum was chosen over free klone for a measured reason: the tiles arm consumes 8.5 MB/s there against a filesystem measured at 8.3–11.8 MB/s — it sits on the ceiling — so three concurrent replicates would contend for the wall itself. On Tillicum the same arm uses 4% of available bandwidth and is genuinely GPU-bound.
Campaign B — RampNet seed variance, klone. Three replicates of the committed recipe (1 epoch / 9,378 steps, constant lr 1e-5, global batch 16) at seeds 1/2/3. Jobs 39583887/39583890/39583891 on
ckpt-all(relaunched 2026-09-04; the original 39515025/26/27 died at submit time on the apostrophe parse bug fixed in 2b7471c), free and preemptable, resumed by--requeueplustrain.py's ownlatest_checkpoint.pth. Calendar, not money, is the risk:ckpt-all's duty cycle was 3.9% in 2026-08.One deliberate deviation:
save_period=1The #51 arms ran
save_period: -1, which is exactly why the epoch-curve follow-up had to be retracted — no per-epoch weights exist for any arm and they cannot be recovered. Keeping every epoch costs ~20 GB against a 1 TB allocation and buys back an analysis that is currently foreclosed.Tests
tests/test_seeding.py(21 tests) checks the plumbing at the source level, because every failure mode here is silent:42literal survivesSEED— an unset replicate would be a silent duplicate of the published runRUNDIRisolation, and no checkpoints under klone's 10 GB home quotasys.argvunpack — an off-by-one there shiftsimgszintoepochsand trains a wrong model that finishes greenFull suite: 1,343 passed, 1 skipped.
--seedverified end to end in the real klonesidewalkcv2environment, andseed=1confirmed in job 274367's Ultralytics arg dump.Stated limitations
seed: 0arm is klone L40S. The SD is over the three same-hardware replicates only; the seed-0 arm is a separate cross-hardware check, not a fourth sample.y11x_tilesis replicated — the measured SD is not automatically the other arms'.deterministic=Trueguarantee, so its spread is an upper bound on seed effect alone.Related: #154 (the parity result this exists to adjudicate), #51, #135.
🤖 Generated with Claude Code (claude-opus-5[1m])