What it does​

• Regenerates scenarios in the browser from their seed with the pipeline's own generator. No scenario files ship. A test checks the browser path is byte-identical to Node.
• Runs the models in the browser with tfjs on the CPU backend. The same encoding and decoding the pipeline uses feed it, and a test finds it within float32 tolerance of tfjs-node.
• Three modes. Truth shows the engagement. Instantaneous shows the causal LSTM's estimate at a "now" cursor, from the bearings up to then. Time-late shows the bidirectional LSTM's estimate at every sample of the window.
• Every gate's carried pair: iteration 1 (G1 pass) and iterations 2, 2b and 2c (G2 fail).
• The URL holds the view, so every link below reopens exactly what is described.

The 2c BiLSTM runs in about 1 s for one scenario in headless Chromium; the LSTM in about 0.5 s. The eight models total 4.5 MB.

How it works: the viewer.

A typical engagement​

Seed 203, scenario 97612, instantaneous, zig + 10 min. The 2c LSTM sits near its own median errors here.

Before the zig the model still answers, but range is unobservable from bearings alone. The viewer greys that period out. Here the LSTM's range sits well short of the truth before the zig and reaches it only after the turn.

The same scenario in time-late mode shows something the medians hide. The BiLSTM's estimated positions form a jagged path rather than a straight track. Each sample's estimate comes from its own output row, and nothing ties them to one constant-velocity solution. The classical batch solution is one straight track by construction. In this case the BiLSTM's post-zig median speed error, 3.0 kt, is worse than the instantaneous LSTM's 1.8 kt.

The largest speed error​

Seed 203, scenario 95484, instantaneous, zig + 10 min has the 2c LSTM's largest speed error at zig + 10 min among the σ = 1° scenarios: 10.9 kt.

The geometry is hard, not broken. The target starts 2.5 km away and opens at 17.3 kt on a course within about 10° of the line of sight. The bearing barely moves, so speed along the line of sight is weakly observable. The LSTM answers 6.5 kt, pulled towards typical training speeds. The classical solver struggles too: 13 of its 101 post-zig solutions fail to converge, and its median speed error is 6.5 kt. One case cannot say how often this happens; it shows one way the models fail.

A held-out engagement beyond the training ranges​

Seed 201 held-out, scenario 102983, time-late, zig + 5 min starts at 25.8 km. Training ranges ran from 2 to 20 km. The BiLSTM's range estimate starts near 18 km against a true 25 km: in this case the model does not reach ranges beyond those it was trained on. The classical solver has no such limit.

What comes next​

The owner now runs the realism check in Annex A §A.5: do the tracks, turns, speeds and bearing noise look like real engagements? Its record (§A.7) decides which Part B levers, such as larger or more ownship course changes, are worth pre-registering on fresh seeds from 204 onward. Nothing is trained until then.

These three cases illustrate; they do not measure. Frequencies come from the gate experiments, not from cases picked in a viewer.

Verdict: G2 failed (agent review panel, round 2, no rejections). Under SRD §7 the project stops at G2: no-go.

• Range passes for both models: 13.5% [13.0, 13.9] instantaneous and 12.1% [11.7, 12.6] time-late at zig + 300 s on the σ = 1° slice, against 20%.
• Course passes for both: 7.36° [7.05, 7.70] and 6.12° [5.87, 6.41], against 10°.
• Speed fails for both: 1.47 kt [1.37, 1.52] and 1.53 kt [1.47, 1.61] at zig + 600 s, against 1 kt. Only 37% and 33% of scenarios are under 1 kt. Classical least squares reaches 0.61 kt on the same scenarios, at the Cramér–Rao bound.

Experiment of record: experiments/20261016-i2d-g2/. Curves and slices: iteration-2c results.

What changed​

Iteration 2b showed that weighting speed 4× in the loss moved speed by 0.3 kt; the capacity scan showed that a two-layer 128-unit LSTM reaches the 96-unit model's plateau in half the epochs. SRD §7 was amended on 2026-10-15, after 2b's validation result and before its test scoring, to allow one further iteration combining the two, then stop (decisions 18–20, spec 005).

• Model. Two stacked recurrent layers of 128 units (203k parameters for the LSTM, 538k for the BiLSTM, against 41k and 123k before), trained from scratch; the CNN and MLP controls at their sizes.
• Loss. Speed weighted 4×, as in 2b.
• Fresh test split. Seed 203, pinned by content digest in the plan; seeds 201 and 202 decided the earlier attempts; all three are now reserved in the plan checker.
• Cost. The BiLSTM took 27.5 runner-hours over five chained jobs in two runs; the LSTM 12.6 hours. Per-epoch cost rose 1.4× and 2.2× against 2b.

The decision record, with a forecast of 1.4–1.6 kt, was committed four hours before training started. Validation gave 1.42 / 1.52 kt; the test split gave 1.47 / 1.53 kt.

Results on the seed-203 test split (σ = 1°)​

n = 4,689 scenarios at zig + 300 s (93% of the slice) and 3,038 at zig + 600 s (60%).

The repeated G1 checks: both models at 0.35–0.39 × naive on range (pass); controls beaten; shuffled labels at 1.027 × naive (pass). The time-late model is, for the first time, better than the instantaneous one on range and course, and worse on speed; its "not worse" check fails at three late points without an interval (the paired intervals contain the slack), the deficit late, largest at σ = 2° and worst near the window end.

The trend, and why it stops here​

Each step moved the instantaneous model's speed by about 0.3 kt (the time-late model's by 0.6 kt then 0.24 kt), and the 2c-over-2b gain is real: on the same seed-203 rows the paired difference is −0.31 kt [−0.41, −0.23] for the instantaneous model and −0.24 kt [−0.32, −0.15] for the time-late one. But a pass needs the whole interval below 1 kt, a median of about 0.93 kt: a further 37%, more than both steps combined. The panel's statistics reviewer put it this way, and the panel adopted it as the no-go statement:

On the fresh seed-203 test split the instantaneous model's median speed error is 1.47 kt (95% CI 1.37–1.52) and the time-late model's 1.53 kt (1.47–1.61): 12 and 14 standard errors above the 1 kt threshold, with only 37% / 33% of scenarios under 1 kt; the interval would still exclude 1 kt with as few as about 80 scenarios (56 for the time-late model). Range and course pass with wide margins. Each iteration moved speed by about 0.3 kt and the result matched validation and the pre-recorded forecast; reaching the gate would need a further 37%, larger than both steps combined, from a capacity lever the scan already showed flat (at 30 epochs with speed weight 1: 192 units 2.25 kt, 2 × 128 units 2.24 kt) and a loss-weight lever that bought 15%. Classical least squares reaches 0.61 kt on the same rows, at the Cramér–Rao bound: the information is in the data; this family of models does not extract it. Caveats: three different test splits across iterations (split-to-split differences at most 1.4 standard errors), one training seed per model, two levers tried.

What the models did and did not learn​

The panel's geometry and numerics reviewers agree on the shape of the failure.

• Speed error does not follow the geometry. By true range at zig + 600 s the model's error is 1.2–2.0 kt in every band, where classical's falls from 1.6 kt at 20–40 km to 0.1 kt under 2 km. By bearing rate it is flat at 1.3–1.8 kt against classical's 0.4–0.9, and the ratio to classical is largest, 3.7×, in the best-observed band (above 1°/min). By range the ratio reaches 16× below 3 km. Against the per-scenario Cramér–Rao bound the models' median error-to-bound ratio is 1.30, where classical's is 0.66 and an efficient estimator's 0.67, so about twice an efficient estimator in the median and nine times the bound below 3 km.
• The gains were uniform shifts. The 2c improvement over 2b is −0.2 to −0.3 kt in every range band and at every time. By true speed it came from the bands away from the population mean (3–9 kt and 15–18 kt); 9–15 kt and 18–20 kt moved by under 0.1 kt. That is eased regression to the mean (slope of prediction on truth 0.65 → 0.71, about 16% less shrinkage); scatter fell by only 5–9% (2.23 → 2.11 kt). The term that has to move is the one that moved least.
• The floor is scatter. With every speed-dependent bias removed the median error would still be 1.35 kt. A bidirectional model that sees the whole window did no better on speed than the causal one.
• Outside the training range band there is no speed information. On held-out scenarios starting at 20–30 km the models' speed error is 3.4 kt at zig + 600 s, 0.8 × the naive prior, and classical's 3.6 kt is no better. Five minutes after the zig the models beat classical on all three quantities there, because classical fails to converge on a quarter of those scenarios; ten minutes after, classical leads on range and course.
• Range and course are a different story. They follow the bearing-line geometry the network reads, and five minutes after the zig the time-late model is ahead of classical beyond 10 km and at σ = 2° (14.0% / 7.3° / 1.86 kt against 17.1% / 7.4° / 2.29 kt). That lead is robustness to non-convergence: classical fails to converge on 10–16% of those scenarios and is scored infinite there, while on the scenarios it does solve its course is the better one. Given ten minutes, classical leads everywhere. That early robustness is the project's positive result, and it is not what G2 asks.

Speed from bearings is second-order information: it lives in how the bearing rate changes as ownship manoeuvres. Classical least squares uses it explicitly; the networks, given raw bearings and ownship states, learn a population prior refined by about two knots of scatter.

What the panel found​

Five reviewers, two rounds. Round 1 returned approve-with-findings from all five with no blocker or major finding; every reviewer reproduced the speed medians independently, the reproducibility reviewer re-gated, re-evaluated and re-predicted byte for byte, and the red team found no path from a pipeline error to the verdict. Round 2 approved with wording minors only; the corrections are in this post. Record: review.md.

Nineteen minor findings were answered or fixed. Two are worth repeating: the output scale for speed (ownship's 15 kt maximum) puts targets above 15 kt beyond the normalised range, where the models shrink most (not saturation: the head is linear and predictions reach 19.9 kt), and this lead was never tried; and the models were trained with one seed each, so seed-to-seed variation in these medians is unmeasured.

What was not tried​

Within the SRD's hyperparameter cap, three leads stayed on the table:

1. Derived features (bearing rate and the bearing-line intercept), built in iteration 2 and never trained on.
2. Rescaling the speed target to unit variance, or a dedicated speed head.
3. A per-leg steady-track fit as an input or an auxiliary target, giving the network the second-order quantity it does not extract.

None of them is a lever the record shows moving speed, and the SRD's budget is spent. A follow-on would need a pre-registered validation result with the whole interval below about 0.9 kt from at least two training seeds before another test split is scored.

Where this leaves the question​

The SRD asked whether a network trained on simulated bearings-only data can produce usable range, course and speed. On range and course, yes, with intervals well inside the thresholds and, early in the track, better robustness to least squares' non-convergence. On speed, no: three iterations reduced the error from 2.1 to 1.5 kt against a 1 kt threshold that classical meets with margin. The gate is failed and the project stops at G2.

Verdict: G2 failed (agent review panel, round 2, no rejections).

• Range passes for both models: 14.6% [14.1, 15.2] instantaneous and 15.8% [15.2, 16.2] time-late at zig + 300 s on the σ = 1° slice, against 20%.
• Course passes for both: 8.22° [7.88, 8.54] and 8.66° [8.29, 9.03], against 10°. The time-late course failure of iteration 2 is gone.
• Speed fails for both, by less: 1.77 kt [1.68, 1.83] and 1.77 kt [1.67, 1.86] at zig + 600 s, against 1 kt. The same weights before 2b, scored on the same scenarios, were at 2.02 kt and 2.31 kt. Classical reaches 0.60 kt.

Experiment of record: experiments/20261015-i2c-g2/. Curves and slices: iteration-2b results.

What changed​

The SRD allows one more round after a stalled G2. Iteration 2 left three leads (research R19): derived features, speed in the loss, and the training budget. The user chose the loss (decisions 15–17 in spec 004):

• Loss weight. The speed term of the loss is weighted 4× (loss_weight_speed: 4). Speed contributed about 15% of the loss before; now it is about 40%.
• Continued, not restarted. Every model (the two carried ones and the three controls) starts from its committed phase-B weights, to 220 epochs at most with patience 6. Because the loss changed, tma-train re-scores the best epoch under the new loss before it resumes; otherwise the old best, on the old scale, could never be beaten and the old weights would be kept.
• A fresh test split. The iteration-2 test split (seed 201) decided the first attempt and is never scored again. The gate is judged on the test split of seed 202, generated with the same count and ranges and pinned by content digest in the training plan. Train and validation are unchanged.

Three tooling checks promised at the end of iteration 2 arrived with 2b and were on for this gate: tma-g2 checks the plan's content digests itself, refuses a phase decision that is not a git ancestor of the carried weights, and checks that every estimator was scored on the same rows, not just the same number of them. The evaluation page describes them.

The carried set was fixed before training: the 2b models, unconditionally. Phase B is scored beside them on the fresh split and reported, never gated.

Results on the fresh test split (σ = 1°)​

n = 4,682 scenarios at zig + 300 s (93% of the slice) and 2,973 at zig + 600 s (59%).

Phase B's numbers here differ from the iteration-2 verdict (2.09 / 2.38 kt) because this is a different split; the intervals overlap.

The repeated G1 checks​

The time-late model's knife-edge G1 fail from iteration 2 (0.514) is now a clear pass. Its deficit since G1 was a training deficit, not an architectural one: given the same budget it matches the instantaneous model on every check. The "not worse" check has no interval; the two exceedances (ratios 1.055 and 1.056 against 1.05) have paired intervals that contain 1.05, so this fail is not distinguishable from a pass. What is real is the G1 caveat persisting: the time-late model is slightly worse than the instantaneous one at 17 of 19 points, not better.

What the speed weight did, and did not do​

• Speed moved by 12–23% (2.02 → 1.77 kt; 2.31 → 1.77 kt), and range and course improved with it. Nothing got worse.
• It is still 1.8 kt against 1 kt, with the whole interval above 1.67 kt. The models are about three times classical, which is at the Cramér–Rao bound on these bearings: the panel recomputed the per-scenario bound on this split (median 1.05 kt for the bearings to zig + 600 s) and classical's error is 0.62 of it at the median, where an efficient estimator gives 0.67.
• The gain is less scatter as much as less shrinkage. The slope of predicted on true speed rose from 0.58 to 0.65 and the residual spread fell from 2.38 to 2.21 kt; correcting the shrinkage alone would not help (it would give 2.3 kt). Target speeds above ownship's 15 kt map to normalised targets above 1, where the shrinkage is worst; the output scaling was not changed.
• The gap is where classical is near-perfect. By true range at zig + 600 s (all noise):

Classical's 0.60 kt median is carried by the short ranges, where a few hundred metres of range error fixes speed to a fraction of a knot. The models are 1.4–1.5 kt there, flat across bands they resolve well, and only beyond 20 km do they beat classical on speed.

• The range advantage beyond 10 km is narrower than iteration 2 suggested. At zig + 300 s the models are ahead at 10–20 km (14.2% against 18.6%, all noise) and at 20–30 km (24.1% against 27.4%), but the gap is carried by classical's non-converged tail (117 of 1,806 scenarios at 10–20 km on the σ = 1° slice, scored as infinite error). On converged scenarios alone the two are level, and at zig + 600 s classical is better in every band.

What the panel found​

Five reviewers again: TMA geometry, estimation and numerics, evaluation statistics, reproducibility, and red team. All five returned approve-with-findings in round 1, and none rejected. Every one of them reproduced the speed medians independently, and none found a path from a pipeline error to the verdict.

One latent major, fixed. When a continuation changes the loss, tma-train re-scores the best weights under the new loss before it resumes. It used that value in memory but never wrote it to the manifest, so a chained job (resuming with its optimizer state) compared new-loss epochs against the old-loss number. Had a continuation's first job ended before any epoch beat the re-scored value, later jobs would have kept the old weights silently. It did not happen in this run (the reviewer checked every job boundary in the CI logs), but it could have. The manifest now records the value a later epoch must beat and the source of a continuation, a chained-job test covers the case, and tma-train --check verifies that a finished continuation's kept weights are explained by its record.

What else the panel established:

• The fresh split is fresh: per-scenario seeds derive from the dataset seed and index, a seed-201 and seed-202 collision would need an index difference of 10¹⁸, and no test scenario shares a seed or parameter set with any of the 100,000 seed-201 scenarios.
• The fail margin is 16–20 standard errors. Only 30% of scenarios have a speed error under 1 kt; for the interval merely to touch the threshold the test set would have to be about 28 scenarios.
• The re-scoring worked as specified: the reviewers recomputed the re-scored losses to five decimals and confirmed each model's best epoch was chosen under the new loss. The shuffled control's weights are unchanged from phase B because no epoch beat its re-scored loss, which is the right outcome for a control.
• Classical's non-converged scenarios (1.4%) are scored as infinite error and included; its 0.60 kt is not flattered.

Nineteen minor findings were answered or fixed in two rounds, mostly record-keeping: the manifests of the committed 2b models do not say they are continuations (the facts are in the review), gate-deciding seeds are now reserved in the plan checker, and the write-up wording above follows the geometry reviewer's corrections.

Record: review.md.

Next​

SRD §7 was amended on 2026-10-15 to allow one further round, iteration 2c, and then stop. The capacity scan (research R20) showed that a two-layer 128-unit LSTM reaches the 96-unit model's plateau in half the epochs. Iteration 2c trains that model from scratch with the speed weight, on yet another fresh test split (seed 203). The decision record, with the forecast (1.4–1.6 kt) and the stopping rule, is committed before training: experiments/20261015-i2d-decision/.

Verdict: G2 failed (agent review panel, round 2, no rejections).

• Range passes for both models. Median range error at zig + 300 s on the σ = 1° slice is 16.5% [16.1, 17.1] instantaneous and 18.5% [17.7, 19.1] time-late, against a 20% threshold.
• Course passes for the instantaneous model only: 9.46° [9.02, 9.90]. The time-late model fails at 10.57° [10.16, 11.04], against 10°.
• Speed fails decisively for both: 2.09 kt [2.00, 2.18] and 2.38 kt [2.28, 2.48] at zig + 600 s, against 1 kt. Classical reaches 0.62 kt on the same scenarios.

Experiment of record: experiments/20261014-i2-g2/. Curves and slices: iteration-2 results.

What changed from iteration 1​

• Noise and dropouts.
  • Bearing noise σ ∈ {1°, 2°}, half the scenarios each.
  • Dropouts drawn per sample, at 0–10% per scenario.
  • G2 is judged on the σ = 1° slice. The σ = 2° slice and the two dropout bands are reported.
• A rate-limited turn.
  • Ownship's zig is now flown at 0.5–3°/s instead of instantaneously.
  • The zig time is the end of the turn, so "zig + 300 s" means five minutes after ownship is steady on its new leg.
  • This changed the generator (0.2.0), so G0 was re-run and passed again (panel rounds 1–3, 20261013-g0-i2-r2).
• Provenance.
  • Every dataset header names the generator's commit.
  • Every model records the header of the data it was trained on.
  • tma-train refuses to resume on different features. tma-g2 refuses to score the test split before the phase decision is committed.
• Confidence intervals.
  • Every gated median carries a seeded bootstrap 95% interval.
  • A check whose interval contains its threshold is inconclusive, not a pass or fail.

Two training phases, one test score​

The test split was protected by a two-phase rule, fixed before training (research R5).

Phase A ran the G1 configurations unchanged: 60 epochs, patience 3, one seed per model. On validation, the instantaneous model passed range (17.4%) and failed course (10.47°) and speed (2.24 kt), so phase A stalled. The decision record was committed before phase B started: decision.md.

Phase B made one change, the training budget. Every model continued from its phase-A weights to 150 epochs with patience 6. This was the user's choice over the pre-registered default (derived features), and the record says what it tests: whether undertraining explains the gap.

Phase B improved range (16.6%) and course (9.56°) on validation, but not speed (2.12 kt). It was carried by the pre-declared rule: its range error was no worse than phase A's by more than phase A's interval half-width.

The test split was then scored once, for phase B, with phase A beside it.

Results on test (σ = 1°)​

n = 4,707 scenarios at zig + 300 s (93% of the slice) and 2,995 at zig + 600 s (59%).

The repeated G1 checks​

These are judged on all rows.

The time-late fail sits on a knife-edge. The interval's lower bound is 0.50035, and with other bootstrap seeds it falls between 0.4995 and 0.5006. It does not decide the gate.

Why speed fails​

The panel ruled out the pipeline:

• Three reviewers re-predicted phase B independently and got exactly 2.09 kt and 2.38 kt.
• The model's mean speed is right: 11.58 kt predicted against 11.69 kt true.

The errors come from shrinkage towards the mean instead:

• The slope of predicted against true speed is about 0.57.
• Slow targets are over-estimated and fast ones under-estimated: errors are 2.6 kt at 0–5 kt and 2.9 kt at 15–20 kt, against 1.5 kt at 10–15 kt.

The information is there. A per-scenario Cramér–Rao bound on the actual bearings gives a median achievable speed error of 0.64 kt, and classical reaches 0.62 kt. The models are about 3.3 times the bound.

Three things point to how the models are trained, not to a ceiling of the approach:

• Speed gets little weight in the loss. All four loss terms are weighted equally, and speed contributes about 15% of the loss.
• Training stopped early. In phase B the training loss was still falling (LSTM 0.400 → 0.377 over its last 14 epochs). Validation loss moved by about ±3% from epoch to epoch, more than the per-epoch gain, so patience 6 ended the run.
• More epochs barely moved speed. The phase-B change improved range and course but not speed.

Where the models do well​

Range bands at zig + 300 s, all noise levels:

Short ranges carry classical's advantage. Beyond 10 km the models match or beat it on range, and fail grossly far less often. The same holds at σ = 2°: the instantaneous model's range error moves only from 16.5% to 18.4%, while classical's doubles from 8.8% to 17.1%.

On held-out geometry, with initial range 20–30 km (outside training, reported only), the models beat classical on every median:

What the panel found​

Five reviewers: TMA geometry, estimation and numerics, evaluation statistics, reproducibility, and red team. All five returned approve-with-findings in round 1, and none rejected.

Two major findings, both resolved:

• Phase A's predictions overwrote phase B's on disk. tma-g2 wrote both phases' prediction files under the same names. The verdict was unaffected, because it was scored in memory, but re-scoring from disk would have given phase A's numbers. Fixed: the other phase's files are now tagged.
• The models were stopped by noise, not converged. Accepted. The verdict is stated for this architecture, these inputs and this budget.

A gate-rule fix. The G2 code failed a check whose interval straddled its threshold whenever the point estimate was at or above it. The spec says that case is inconclusive. The code now follows the spec, and a check passes only if its point estimate is also below the threshold.

Re-gating the committed metrics with the new code gave identical checks, and the reproducibility reviewer re-ran the whole gate byte for byte.

Round 2 approved with only minor findings open. Record: review.md.

What G2 cannot tell us​

• The speed point covers 59% of scenarios. Only windows that run at least ten minutes past the turn are scored at zig + 600 s, so the figure describes longer windows (decision 13). At zig + 300 s, where 93% are scored, speed is still 2.30 kt.
• Medians hide the range dependence. The range pass is not carried by short ranges, which are the models' weak band. But a single median mixes a 10–20 km band where the models beat classical with a 2–5 km band where they are four times worse.
• One seed per model (decision 11). The intervals cover test-set sampling, not training-to-training variation.
• The course pass is narrow. Its interval's upper bound is 9.90° against a 10° threshold.
• Targets never manoeuvre. G2 says nothing about tracking a target's own zig.
• Not done: scoring the G1 models on a turn-only set (reported only, not gated). It is deferred to iteration 2b.

Next: iteration 2b​

The SRD allows one more round after a stalled G2 (decision 5). The leads, in research R6, R18 and R19:

1. Derived features. Bearing rate and the bearing-line intercept, already built (feature_set: derived), give the network quantities classical TMA works from.
2. Speed in the loss. Weight it per term, or scale its target to unit variance.
3. Per-leg steady-track fit. Within each ownship leg, fit tan B(t) = (x₀ + vₓt)/(y₀ + v_yt) to the bearings. The fitted shape constrains relative velocity without needing range, and the residual flags a target manoeuvre.
4. Budget and stopping. A smoothed stopping criterion or a longer run, since training loss was still falling.

Before 2b scores a gate, three tooling fixes are due:

• tma-g2 checks the plan's content digests itself;
• the decision commit must be an ancestor of the weights' commit;
• the same-rows check compares row identity, not just counts.

Verdict: G1 passed (panel round 3) with a caveat. Both model variants beat the naive baseline by more than 2× at the gate point, with the whole 95% interval below 0.5: instantaneous 0.419 [0.407, 0.431], time-late 0.450 [0.438, 0.464] (medians 16.1% and 17.3%; naive 38.4%, classical 4.6%). The shuffled-label model scores at naive. The machine verdict is inconclusive, because the time-late model is measurably worse than the instantaneous one (1.075 [1.046, 1.108] × at zig + 300 s) although it sees more data; the agent review panel passed G1 under a rule fixed before the deciding run (pre-registered rule 1, SRD §5) and recorded that as the caveat. Experiment of record: experiments/20261012-g1-r2/.

It took two attempts. In attempt 1 (experiments/20261011-g1-r2/) the time-late model was cut off by a wall-clock cap after 25 of its 60 epochs and failed (0.551 [0.538, 0.565]). One pre-registered continuation of that model, to early stopping at 50 epochs, gave the pass. Full curves and slices: iteration-1 results.

The question​

G1 (SRD §5) passes if, for both model variants, the median relative range error 5 minutes after the ownship zig is below half of the naive baseline's, on the same test scenarios. The naive baseline predicts the training-set mean range, course and speed, so a model that beats it by 2× has learned something about the geometry, not just the prior.

The two variants answer different questions. The instantaneous model is causal: its estimate at a given time uses only bearings up to that time, as a real-time tracker would. The time-late model sees the whole recorded window, including bearings after the gate point, and so is a post-hoc reconstruction.

Each ratio is judged with a paired bootstrap 95% interval: pass if the whole interval is below 0.5, fail if it is wholly at or above, otherwise inconclusive. Further checks guard the result: a model trained on shuffled labels must score like naive (leakage), each recurrent model must beat a non-recurrent control, and time-late must not be worse than instantaneous. How these combine is on the evaluation page.

What was trained​

Four small models over the same masked 240-row input window (see models), plus one leakage run, trained on GitHub-hosted runners (4 cores, no GPU). Every model had max_epochs 60, patience 3 on validation loss and a 5.5 h wall-clock cap (max_wall_s 19,800; 2 h for the shuffled run). Epochs are counted from 0, so "60 (best 58)" means the 59th epoch gave the kept weights.

Data: 100,000 scenarios (seed 101), σ = 0.5°, no dropouts, initial range 2–20 km log-uniform; 80,000 train, 10,000 validation, 10,000 test, plus 10,000 held-out scenarios at 20–30 km.

The bilstm-time-late row is attempt 1. Under the protocol below it was continued once, from the epoch-23 weights with no wall-clock cap, and stopped early after 50 epochs (best 46); the other four models are unchanged. One configuration per model; nothing was tuned. In attempt 1 the time-late configuration was started twice with identical settings (one job cancelled, one completed).

Every run (FR-009)​

• 38131616871, train/i1, 11 Oct: all five models. Instantaneous, CNN, MLP and shuffled completed; these four are the models of record. The time-late job was cancelled after ≈ 5 h 15 min; the g1 job then stopped at once (time-late: training has not finished). Test split not scored.
• 38152208243, train/i1, 11 Oct: time-late only, attempt 1 (5 h 24 min job, AMD EPYC 9V74); the other four reused. Its g1 job scored the test split but could not push its results branch; informational only (same models as the attempt-1 re-score).
• 38153046011 and 38154023864, train/ckpt-smoke: checkpoint smoke tests (MLP); g1 stopped at model collection. Test split not scored.
• 38165442543, train/i1: cancelled 25 s after it was queued; no job ran.
• 38178916962: re-score (evaluate-g1.yml) of the committed models, failed before any model was scored. The test features, regenerated at schema 0.5.0, carry a different header digest from the 0.4.0 training data, so prediction refused them ("normalisation differs"). Fixed by PR #27 (accepts a schema-version-only relabel).
• 38182021648: re-score with PR #27, attempt 1 of record, experiments/20261011-g1-r2/.
• 38208186337, train/i1-continue, 12 Oct: the pre-registered time-late continuation, over two run attempts (jobs listed under attempt 2 below). Its final g1 job scored the test split into experiments/20261012-g1/, the training run's own folder (train/i1 code, no bootstrap; its verdict.json reads fail): informational only.
• 38275763335: re-score of all five committed models on main code (evaluate-g1.yml), attempt 2 of record, experiments/20261012-g1-r2/.

Attempt 1: results at the gate point (zig + 300 s, test split)​

Experiment 20261011-g1-r2, re-scored from the committed weights by run 38182021648. n = 9,784 scenarios; 216 excluded because their window ends before the gate point; the same scenarios for every estimator.

G1 threshold: 0.5 × naive median = 19.2%.

Classical non-converged rate at the gate point: 2.0% (non-converged rows count as +∞). Classical is far ahead of every model (a median 3.5× lower than the instantaneous model's); G1 asks only whether the models beat the prior, and closing that gap is G3's question.

Intervals​

Paired bootstrap of the ratio model / naive median range error at zig + 300 s (2,000 resamples of the 9,784 test scenarios, seed 20261011). A G1 check passes only if the whole 95% interval is below 0.5.

The classical ratio (in brackets) is a point value; no interval was computed for it. The interval reflects which test scenarios were drawn, not how a re-trained model would score (one training seed per model).

Checks​

Machine verdict: fail, because g1-time-late and time-late-not-worse failed.

Error against time since the zig. Naive is flat (38–49% median throughout). Classical falls fast: 22.6% one minute after the zig, 4.6% at five minutes, 1.6% at fifteen. The instantaneous model falls from 24.5% at one minute to 16.1% at five and levels off near 15% at ten to fifteen minutes. The time-late model has the same shape but sits above it everywhere (26.0%, 21.2%, 19.0%), although it sees more data at every point. Rounds 1–2 read this as the signature of an undertrained model, not of a pipeline that hides the future from it (see the panel's checks below); round 3 withdrew that explanation for the continued model, which trained to early stopping and is still worse (see the caveat). Interactive curves are on the results page.

By range band and zig angle​

Median range error at zig + 300 s, with each model's ratio to naive in brackets. Bands are by the mean range over the window, so a few test scenarios fall outside 2–20 km. Slices with n < 200 are not interpreted.

The per-band ratios mostly measure naive, not the models: naive guesses the middle of the prior, so it is very wrong at short range and nearly right at 5–10 km. In absolute terms the models show the prior's U-shape (best at 5–10 km, worse at both ends), whereas classical's error grows steadily with range, as the geometry predicts (lower bearing rate at long range). Larger zigs help classical (6.4% → 3.5%) and, weakly, the instantaneous model (17.1% → 15.3%); the time-late model is flat across zig angle (21.0–21.6%).

Held-out geometry (20–30 km initial range, reported, not gated)​

10,000 scenarios outside the training range; at zig + 300 s, n = 9,769 (231 excluded).

Outside the training range the models lose most of their advantage over naive (ratios 0.71 and 0.77, against 0.42 and 0.55 in range), while classical keeps a median of 19% at the cost of more non-converged solutions. The panel read this as the models being held towards the training prior rather than extrapolating the geometry. G2 is to include a range-extrapolation check.

Attempt 1: why the time-late model was undertrained​

All models shared the same budget in time, not in epochs: a 5.5 h wall-clock cap per model, inside GitHub's 6 h job limit. The bidirectional LSTM runs the window in both directions and has twice the instantaneous LSTM's parameters (82,948 vs 41,476); one epoch took about 775 s against 276 s. The cap therefore stopped it after 25 epochs while the instantaneous model ran all 60. Its validation loss was still falling (0.592, 0.565, 0.571 over the last three epochs; best at epoch 23).

The panel compared like with like. On the post-zig validation rows (the rows that are scored), the time-late model at epoch 23 had a loss of 0.406; the instantaneous model at the same epoch had 0.478 and needed until epoch 31 to reach 0.405. At equal training, time-late was ahead.

What the panel found in attempt 1 (rounds 1–2)​

Five independent agent reviewers (TMA geometry, estimation and numerics, evaluation statistics, reproducibility, red team) reviewed attempt 1 over two rounds; the full record, with every finding and response, is in experiments/20261011-g1-r2/review.md.

In plain words:

• The pipeline is sound. Padding does not leak into the bidirectional pass (padded and unpadded windows give identical outputs); predictions line up with their scenarios; the causal models ignore bearings after the time they estimate; the time-late model does use later bearings (cutting them off makes it worse, 21.2% → 27.2%).
• The instantaneous pass is real. No truth reaches the inputs or the normalisation, naive is not artificially weak, every comparison uses the same 9,784 scenarios, and the shuffled-label model trained down to the "know nothing" level and scores like naive.
• Reproducible. A reviewer regenerated the data and re-scored the committed weights locally: the verdict file came out byte-identical, naive and classical metrics byte-identical, model metrics within 1.3 × 10⁻⁵ relative.
• The time-late fail is a budget effect (one reviewer rejected in round 1 until this was recognised). Calling it "ML not extracting the signal" would overstate it. The first result is recorded as: G1 not passed as run; the instantaneous variant met the criterion; the time-late variant failed under a 25-epoch wall-clock budget and is plausibly undertrained; no evidence that the time-late signal is absent.
• But no rule allowed a re-run. Turning "undertrained" into "try again" needed an explicit amendment, decided after seeing the result and written down before any new training. That is what the user decided (below), and round 2 approved the protocol with all five reviewers at approve-with-findings.

Smaller findings, reported rather than fixed now: pre-zig rows, never scored, carry 64% of the time-late model's training loss; model manifests lack a git commit field (provenance is recorded in the review instead); training budgets should be set in epochs, not wall-clock time, because runner CPUs vary; the shuffled model's median course error (82.9°) is a little below chance (≈ 90°), a weak bearing-as-course prior rather than leakage.

The decision and the continuation protocol​

The user decided on 12 October, after the first G1 result and before any further training, to continue the time-late model once, to the same 60-epoch budget the instantaneous model had. SRD §5 was amended to allow exactly this, and the protocol was fixed in review.md before the run started:

1. One change only: no wall-clock cap. Training resumes from the committed epoch-23 weights (digest checked) and runs to epoch 60 or early stopping (patience 3). Epoch 24 of the first run is dropped.
2. Unchanged: data, splits, training seed, architecture, loss, learning rate, batch size. The other four models are reused untouched; the instantaneous model is not retrained.
3. Optimizer: Adam's state was never saved in attempt 1, so it restarts once, at epoch 24, and epochs 24–26 do not count towards patience. From then on Adam's state is saved with every checkpoint and restored if a job is interrupted, so a resumed run matches an uninterrupted one (tested: identical weights).
4. Scoring: the verdict of record comes from re-scoring all five committed models with evaluate-g1.yml on main; the test split is not touched until training completes.
5. Outcomes, fixed in advance: both intervals wholly below 0.5 and shuffled labels at naive → G1 pass, even if time-late-not-worse still fails (the record then says "time-late not shown better than instantaneous"). A ratio fail of the continued model is final: no further continuation, seed or architecture change, and the project stops per SRD §5.
6. Record: attempt 1 (0.551) stays recorded as a valid attempt. Every job of the continuation is listed with its run id, epochs covered and runner CPU.

Attempt 2: the continued time-late model​

Experiment 20261012-g1-r2, re-scored by run 38275763335 on main code (commit bb4525b). The four reused models are byte-identical to attempt 1; their metrics differ from attempt 1's re-score by at most 8.4 × 10⁻⁶ relative (float32 inference on AVX2, then AVX-512), and naive and classical are identical.

Training. Run 38208186337, branch train/i1-continue, commit 7c1c579. Before training, the seed step checked the epoch-23 weights (digest 2c5f8784…) and a configuration that differed from attempt 1 only in max_wall_s (19,800 → 86,400).

The first job's log reads "resuming bilstm-time-late after epoch 23 (optimizer state restarts; patience not counted for epochs 24–26)", the second's "resuming bilstm-time-late after epoch 41 (optimizer state restored)". Epochs took about 1,100 s on the AMD runner and 870 s on the Intel one. Epochs 0–23 ran in attempt 1 (run 38152208243, AMD EPYC 9V74), so training crossed three CPU families and a retrain will not reproduce the weights byte for byte. The committed weights (c52b628c…) are the artefact of record; re-scoring them is deterministic, and a reviewer reproduced this experiment folder byte for byte. Extracted log lines are in experiments/20261012-g1/training-logs/.

Training stopped after 50 epochs (best 46), stop reason early_stopping, against the instantaneous model's 60. Validation loss at the best epoch was 0.441 (attempt 1's best: 0.565 at epoch 23).

The epoch-48 spike. Train and validation loss jumped at epoch 48 (validation 0.447 → 0.535, then 0.448 at epoch 49). It is a training instability, not a resume artefact: the instantaneous model, never resumed, had one at epoch 40. The scored weights are from epoch 46, before it. The spike contributed to the patience stop at epoch 49, so the "equal epoch budget" was in effect 50 against 60; whether epoch 48 would otherwise have improved is unknown. Gradient clipping is an iteration-2 candidate.

At the gate point (zig + 300 s, test split)​

n = 9,784 scenarios (216 excluded), the same for every estimator, as in attempt 1.

G1 threshold: 0.5 × naive median = 19.2%. Classical non-converged rate at the gate point: 2.0%. Classical is still far ahead: its median is 3.5× lower than the instantaneous model's and 3.8× lower than the time-late model's.

Intervals, both attempts​

Paired bootstrap of model / naive median range error at zig + 300 s (2,000 resamples of the 9,784 test scenarios, seed 20261011). A G1 check passes only if the whole 95% interval is below 0.5. Medians are of range error; classical's ratio to naive is 0.119, a point value (no interval computed).

The time-late upper bound is 0.036 below 0.5, about 5 bootstrap standard errors; its point estimate is about 7 below. The statistics reviewer's independent bootstrap agreed (time-late [0.4377, 0.4649], instantaneous [0.4074, 0.4313]). The intervals reflect which test scenarios were drawn, not how a re-trained model would score: each variant was trained once.

Checks​

Machine verdict: inconclusive, from time-late-not-worse (tiered rule 2). verdict.json is left exactly as the code wrote it; the pass is the panel's application of rule 1.

Error against time since the zig (median range error, test split):

The continuation closed most of the gap to the instantaneous model (attempt 1's time-late model: 26.0%, 21.2% and 19.0% at one, five and ten minutes) but did not cross it.

By range band and zig angle​

Median range error at zig + 300 s, ratio to naive in brackets. Slices with n < 200 are not interpreted.

The time-late model's deficit is at 2–10 km; at 10–30 km it is level with, or slightly ahead of, the instantaneous model. Both models still follow the training prior's U-shape across range bands, whereas classical's error grows steadily with range.

Held-out geometry (20–30 km initial range, reported, not gated)​

n = 9,769 at zig + 300 s (231 excluded).

As in attempt 1, both models are held towards the training prior outside the training range.

What the panel decided (round 3)​

The same five specialist reviewers judged the continuation. The full record is in experiments/20261012-g1-r2/review.md.

All five said the evidence supports G1 pass under pre-registered rule 1. The reproducibility reviewer re-scored the committed weights locally and got verdict.json, bootstrap.json, metrics, curves and held-out results byte-identical; the red team found no protocol deviation, leakage or selection between attempts (one continuation, one re-score).

One major finding (M1, red team; geometry agreed). The SRD §5 G1 row still said a time-late consistency failure makes G1 inconclusive, and did not carry rule 1; as the SRD wins conflicts, the pass needed it there. The row was amended to carry rule 1 as the user decided it on 12 October before the run. Rule 1 was committed at 05:29:44Z and published in PR #28 at 05:30:16Z, 17 s after the run was triggered (05:29:59Z) but before any continuation epoch finished (about 05:49Z) and about 8 h before the test split was scored (13:25Z). The round-1 explanation of the consistency failure (undertraining) is withdrawn for the continued model.

Minor findings, answered in the record: the caveat now states the gap rather than only "not shown better"; the epoch-48 spike (above); a cross-process resume test for the bidirectional layer, which a reviewer ran (byte-identical weights) and which iteration 2 adds to the suite; the per-job record and both CPUs, now in the review; re-scored manifests lack cpu_model, which evaluate-g1.yml will record in iteration 2. The re-invitation corrected the wording of the gap, the margin and the spike. No reviewer rejects and no blocker or major finding is open; the verdict's decided_by names round 3.

Outcome: G1 passes, with a caveat​

G1 passes under pre-registered rule 1 (SRD §5 as amended 2026-10-12). Both ratio intervals lie wholly below 0.5, instantaneous 0.419 [0.407, 0.431] and time-late (continued to 50 epochs) 0.450 [0.438, 0.464], and the shuffled-label model sits at naive. A small network does extract range information from bearings alone, in real time and in reconstruction, though at the gate point it is still about 3.5× (instantaneous) and 3.8× (time-late) worse than the classical solver.

Caveat, recorded with the pass:

• Time-late not shown better than instantaneous. It is measurably worse: at zig+300s, 0.173 vs 0.161, a ratio of 1.075 [1.046, 1.108]. Its median is higher at 18 of the 20 gated points, by more than 5% (up to 9%) at 14 of them (120–480, 600 and 720–1020 s). It is lower only at zig+60s (0.984 [0.964, 1.006]) and zig+1200s (0.973 [0.864, 1.126], n 516); neither difference is significant.
• Not explained by undertraining. The continued model trained to early stopping.
• The BiLSTM under-uses the future. Its error profile is shaped like a causal estimator's. Masking and the backward pass are verified correct, and truncating the future does change outputs, but weakly.
• One training draw per variant. Training-seed variance is unquantified.

What G1 cannot tell us​

• The targets move at constant velocity and the ownship zig is instantaneous; nothing here says how the models cope with manoeuvres or gradual turns.
• σ = 0.5° and no dropouts only; noise and gaps are iteration 2's question (G2).
• The time-late model's estimate at zig + 300 s uses bearings recorded after that time. It is a post-hoc reconstruction, not a real-time solution.
• The interval covers test sampling, not training variation: each model was trained once, and weights differ across runner CPU families.
• The time-late model was trained in three stretches on three CPU families with one optimizer restart; the instantaneous model in one. The protocol equalised the epoch budget, not the training path, and early stopping ended the time-late model at 50 epochs against 60.
• The shuffled-label test rules out leakage through labels and the training harness; truth hidden in the inputs is ruled out by a separate counterfactual test.

The full list, with mitigations, is in specs/002-iteration-1-signal/g1-limits.md.

Next: iteration 2​

Iteration 2 (spec 003) runs the same models on σ = 1–2° with 10% dropouts, judged at G2: median range error below 20% and course below 10° five minutes after the zig, speed below 1 kt ten minutes after, at σ = 1°. At σ = 0.5° the instantaneous model is at 16.1% and 9.4° (zig + 300 s) and 1.83 kt (zig + 600 s), against classical's 4.6%, 1.9° and 0.31 kt, so speed is furthest off. Leads from this gate:

• Pre-zig rows carry 64% of the time-late model's training loss but are never scored.
• Gradient clipping, against spikes like the one at epoch 48.
• Both models still follow the training prior by range band, and are held to the training prior on the held-out 20–30 km set; G2 is to include a range-extrapolation check.
• Why the bidirectional model under-uses the future, before time-late is relied on.
• Engineering: training budgets in epochs rather than wall-clock time; the runner CPU recorded in re-scored manifests; a cross-process bidirectional resume test; a git commit in model manifests.

Verdict: G0 pass (machine verdict; team review pending). Experiment: experiments/20261011-g0/.

• 100,000 noise-free scenarios generated (seed 42); the 10,000-scenario test split was scored.
• The least-squares solver recovered truth on 10,000 / 10,000 scenarios. Worst cases: 2.9×10⁻⁴ relative range, 0.016° course, 0.002 kt speed (tolerances 10⁻³, 0.1°, 0.01 kt). Median range error 1.9×10⁻⁷.
• The naive baseline reproduced exactly across two independent fits.

Yardsticks for iteration 1​

On the same noise-free test set, the naive baseline's median range error is 40% (P90 139%), its median course error 90° and its gross-error rate 37%. Any model must beat these to pass G1.

On noisy data (σ 0.5–2°, up to 10% dropouts) the classical solver's median range error is about 5%, with P90 ≈ 43% and gross-error rate ≈ 9%. None of those gross errors were optimiser failures: starting from the true state never found a lower cost. The tail comes from the information in the bearings, not from a weak solver.

Caveat​

Passing G0 shows that generator and solver agree. On its own it would not catch an error shared by both; hand-computed geometry tests cover that separately.

Generation plus the G0 run took about 70 seconds on a laptop-class machine.