005 — Anytime transfer metric: does scoring the fit trajectory give the instrument the resolution the endpoint metric never had?

Status: STOP (2026-07-27) — triggers #1 (still saturating) and #2 (the metric rewards any non-zero init, not retained structure) both fire at 3 of 3 seeds. G6 reproduced exactly, so the result is scorable: the anytime lever is not the fix. See "Result" at the bottom.

(Registered 2026-07-27. Pre-registered BEFORE any anytime number was measured. The checkpoint ladder, the primary statistic, the NF construction, the decidability rule, δ_rel, every GO/STOP/PARTIAL bar, the seed set, and the Gate-004 reproduction contract are frozen in this commit and may not move after a number is seen.)*

Rung: ROADMAP mid-horizon M2 — Gauntlet gate, but run before M1/M2 as an instrument gate, not a transfer claim. Gate 004 (S2, STOP 2026-07-21) left a standing methodological caveat that ROADMAP now carries into M2: it "must expect the NF=0.0 few-shot-fit washout to blunt the instrument unless the metric separates arms before the fit saturates." This gate tests exactly that — a different metric on the same data, at Gate 004's cost, before M1 builds new worlds or M2 spends the flagship budget.

What this gate is not. It is not a retain gate and it cannot overturn Gate 004. Gate 004's STOP was scored against its own pre-registered endpoint metric and stands. If the anytime metric turns out to resolve an arm difference the endpoint metric was blind to, that re-opens the retain question under a new pre-registered gate; it does not amend 004 (see "Re-opening clause").

Hypothesis

Scoring the trajectory of the few-shot fit — the mean real-trial distance-to-goal over a frozen log-spaced checkpoint ladder within the unchanged 30/29-iteration budget — rather than the fit's endpoint, yields an instrument that (a) desaturates: the cold arm's two independent fits no longer land on bitwise-identical scores, so NF > 0 and a majority of held-out goals become decidable; and (b) does not manufacture separation: it does not credit a warm start that is count-matched but structure-free (RAND-MATCHED), nor one drawn from the real retained repertoire but deliberately mis-retrieved (MISMATCH).

Failing (a) means the anytime lever is not the fix for the instrument. Failing (b) means the anytime lever fixes the instrument by measuring initialization non-zeroness instead of transfer — which would make M2's headline warm-vs-cold curve an unfalsifiable construction, the worse of the two failures.

The primary statistic (frozen; one choice, with the rejection recorded)

Checkpoint ladder `K` (frozen)

A checkpoint at step k is the fast-weight vector after exactly k ES steps of fit_operator. The ladder is log2-spaced:

K = {1, 2, 4, 8, 16, 29}      |K| = 6      K_max = 29 = min(FEW_ITERS, SEL_FIT_ITERS) = min(30, 29)

k = 29 is the post-loop state for the ITE arm (iters = 29) and the top-of-iteration-29 state for every other arm (iters = 30); it means "after exactly 29 ES steps," which is what makes the arms comparable at all. k = 0 (the retrieved seed itself, pre-fit) is snapshotted and reported as diagnostic H0[g], but deliberately excluded from the scored statistic — see below. Nothing in K may be added, dropped, or re-spaced after a number is seen.

`A` — anytime mean distance (PRIMARY)

For arm a and held-out goal g, with d_{a,g}(k) the existing policy_score real-trial distance (-r[0], the 64-tick deterministic rollout MSE to the goal BC — the identical scalar Gate 004 scored, just evaluated at more points):

A(a, g) = (1/|K|) * Σ_{k ∈ K} d_{a,g}(k)          (lower = better)

Uniform weight over the ladder, not trapezoid-in-iteration. This is load-bearing: the ladder is log-spaced, so a trapezoidal area would weight the 16→29 span 13× the 1→2 span and A would collapse back onto ~the terminal value — reintroducing the exact saturation this gate exists to escape. Uniform-over-K is log-uniform in iteration, the correct prior for a converging optimizer and the semantics the ladder already encodes: "the expected distance you get if the fit is interrupted uniformly at random on the rung schedule." That is a genuine anytime/interruptible statistic, not an arbitrary summary.

Aggregation across goals is unchanged Route A/C discipline: median of paired per-goal ratios, never a mean of ratios, with Gate 004's EPS = 1e-9 denominator clamp and the clamp count printed.

R^A_1 = median_g  A(worse ONE-SRC, g) / A(TWO-SRC, g)      (>1 ⇒ TWO-SRC better)
R^A_c = median_g  A(RAND-MATCHED, g) / A(TWO-SRC, g)
W(a)  = median_g  A(cold_a, g)        / A(a, g)             (warm-over-cold anytime speedup)

Why not iterations-to-threshold `τ`

Rejected, and the rejection is frozen so it cannot be revisited mid-run:

τ is a free knob with no principled pre-registration. Distances are per-goal-scale; any τ picked before the data is either off-scale (everything censors → universal ties → the same washout, merely quantized) or would be calibrated from the data, which is scaffolding.
It takes only 7 values (6 rungs + "never"), so its tie probability is high — the very failure mode under test.
Censoring needs a second convention (what value for "never reached"), a third knob.

A has zero free parameters beyond the frozen ladder, never censors, is continuous, and uses every rung. Iterations-to-τ (at any τ) may be reported as a diagnostic but the verdict is scored on A alone; naming a different primary after the run is a documented negative.

Why `k = 0` is excluded from the scored statistic

Two independent reasons, both anti-manufacture:

The ITE arm's k=0 is an order statistic, not a transfer signal. pool_trials returns the argmin over SEL_POOL = 9 real-trial scores. Even a pool of pure noise gets a k=0 head start over cold zero-init purely by taking a min over 9 draws. Scoring k=0 would hand every warm arm a guaranteed advantage with no transfer content.
NF would be structurally deflated. The cold arm is memset_zero in both draws, so d_{cold_a}(0) ≡ d_{cold_b}(0) bitwise and the k=0 term cancels exactly in A(cold_a) − A(cold_b) — while it does not cancel in an arm-vs-arm difference (different retrieved seeds). Including k=0 would give the numerator 7 non-cancelling terms and the noise floor only 6: a silent ~17% deflation of the instrument's own noise estimate.

Instrument statistics (frozen)

Noise floor. The cold arm is fit twice per goal at different fit-RNG stream positions (cold_a = the cold fit inside the TWO-SRC call, cold_b = the cold fit inside the ONE-SRC-A call — Gate 004's existing, unchanged construction), traced identically:

NF_A     = median_g | A(cold_a, g) − A(cold_b, g) |
NF_rel   = NF_A / median_g A(cold_a, g)                (scale-free instrument dispersion)
TIE_frac = fraction of g with A(cold_a, g) == A(cold_b, g) exactly (bitwise Float32)

Decidability (redefined, with a float-dust floor). A goal g is DECIDABLE for the arm pair (TWO-SRC vs worse ONE-SRC) iff

| A(worse ONE-SRC, g) − A(TWO-SRC, g) | > max( NF_A , δ_rel · A(cold_a, g) )
      with  δ_rel = 0.02

The δ_rel = 0.02 floor is new in this gate and frozen now. It exists because the anytime metric's failure mode is the mirror image of the endpoint metric's: where NF = 0.0 made everything undecidable, an NF that collapses to numerical dust (1e-7) would make everything decidable for free — manufacturing a fake GO on the very question this gate asks. δ_rel makes the gate strictly harder and can only be moved by re-registering.

The calibrated null (the sharpest instrument statistic here). Apply the identical decidability rule to the null pair (cold_a, cold_b):

DEC_null = fraction of g with | A(cold_a,g) − A(cold_b,g) | > max( NF_A , δ_rel · A(cold_a,g) )
DEC      = n_dec / ng                                        (the arm pair)

By construction NF_A is the median of the null differences, so DEC_null ≈ 0.50 is the signature of a metric with honest dispersion, and a DEC_null far below 0.50 is the signature of pervasive ties, i.e. residual saturation. This gives DEC a known null of 0.50 and makes it a proper test statistic rather than a bare count. Under Gate 004's endpoint metric the machinery returns DEC_null ≈ 0.25 at best (NF = 0.0 with >50% exact ties by construction) — that is the number to beat.

Goal budget escalation, inherited unchanged. NUM_GOALS = 24 per (world, seed) cell; if a cell yields < N_DEC_MIN = 8 decidable, regenerate once at NUM_GOALS_ESC = 48. Still < 8 after escalation ⇒ that cell is NOT-TESTABLE and is reported as such, never rounded up.

Metric + consumer (no proxy)

Consumer. The number this gate decides is which per-goal scalar M2's transfer curve is built from. Concretely, the consumer is the NF / decidability / median-of-ratios block that today lives in tests/test_cbr_retain.mojo::eval_cell (median_of, ratio_median, the NF loop, the decidable/sign-consistency loop) and that the M2 Gauntlet driver will inherit alongside src/transfer.mojo::run_family_select and src/ued.mojo. That block reads exactly one scalar per (arm, goal). Today that scalar is written in src/transfer.mojo by d[g] = -r[0] — the terminal policy_score after fit_operator returns.

A GO replaces that scalar with A[g]: M2's d_* buffers get filled by fit_operator_traced + post-fit checkpoint scoring instead of one terminal policy_score, and M2 additionally ships the RAND-MATCHED-vs-cold control curve beside its headline.

Rejected proxies (each cheaper, each not the consumer's number): fast-weight L2 distance between final weights; ES fitness on the demonstrations (calculate_fitness inside the fit); fit-loss decay rate; BC-space distance without a rollout. All are fit-side quantities a warm arm could look good on while generalizing no better. The scored quantity stays the held-out real-trial rollout distance.

Arms (frozen) — Gate 004's five, plus one

Exactly Gate 004's five arm maps, byte-identical in their builds and constants: TWO-SRC, ONE-SRC-A, ONE-SRC-B, RAND-MATCHED (count-matched to N_AB, source-A real elites held fixed, B's contribution replaced by Fisher-Yates-scrambled, open-world-re-rolled fillers), RAND2X (build-compute-matched, all-random). cold remains not-a-map (d_cold from run_family_select), read twice for NF/G4.

MISMATCH (new, sixth, appended LAST). The TWO-SRC map, retrieved for the wrong goal: goal g's fit is warm-started from the elite that BC-nearest/ITE returns for goal (g + 1) mod ng, then fitted and scored against goal g's true BC. Identical map, identical count, identical elite provenance (real retained structure from both real sources), identical pool_trials/fit_operator machinery, identical initialization scale — the only thing destroyed is the goal-relevance of the retrieval. The permutation is a fixed cyclic shift by 1 over the goal index: deterministic, zero RNG draws, the least hand-choosable permutation available.

MISMATCH must be appended after the five existing arms in per-cell stream order. Inserting it anywhere else advances the shared fit-RNG stream and would invalidate the Gate-004 reproduction contract (G6). This is the single most likely implementation mistake in this gate.

Refuting controls

There are two distinct ways A could separate arms without measuring transfer.

Control 1 — RAND-MATCHED vs cold: does a structure-free warm start capture the advantage?

Note what Gate 004 established: under the endpoint metric RAND-MATCHED tied TWO-SRC exactly (Rc = 1.0, all three seeds). RAND-MATCHED is a count-matched, BC-retrieved, non-zero warm start carrying zero retained structure from source B. So it is precisely the arm that a metric merely rewarding "any plausible non-zero initialization" would credit.

F_triv = ( W(RAND-MATCHED) − 1 ) / ( W(TWO-SRC) − 1 )

the fraction of the anytime warm-over-cold advantage that is reproduced by a structure-free warm start. Evaluated only when W(TWO-SRC) − 1 ≥ 0.05 (otherwise there is no advantage to decompose, and it is reported as such — the honest-null branch).

F_triv ≥ 0.50 ⇒ STOP. This control directly guards M2's headline: if half the warm-vs-cold anytime curve survives destroying the retained structure, then M2's warm-vs-cold claim would "pass" for trivial reasons and the metric is worse than useless — it is unfalsifiable.

Control 2 — MISMATCH: does the metric know a relevant retrieval from an irrelevant one?

Control 1 has one blind spot: if RAND-MATCHED's scrambled policies are degenerate enough to be worse than cold, F_triv ≤ 0 and the control stays silent while A is still merely rewarding "a plausible-looking policy beats a zero vector." MISMATCH closes it: its warm starts are real retained elites — as plausible as TWO-SRC's, drawn from the same map — with only their relevance destroyed.

| W(MISMATCH) / W(TWO-SRC) − 1 | ≤ 0.05      (while W(TWO-SRC) − 1 ≥ 0.05)   ⇒ STOP

i.e. the metric cannot distinguish a goal-relevant retrieval from a deliberately mis-retrieved one out of the same repertoire; whatever it is measuring, it is not transfer.

MISMATCH precondition (G-MM). Report median_g ||bc(goal g) − bc(goal g+1)|| against median_g ||bc(goal g) − bc(retrieved elite for goal g)||. If the mismatched retrieval target is not at least 2× farther from the true goal than the matched one, the control is toothless and STOP #3 is unreadable ⇒ that cell is NOT-TESTABLE for control 2.

If neither control can produce a "no" — e.g. every arm ties on A at every rung so W ≡ 1 everywhere while TIE_frac is also high — the hypothesis is not testable on this data: declare NOT-TESTABLE for the manufacture clause and score only the desaturation clause. Do not manufacture a GO.

GO / STOP / PARTIAL / NOT-TESTABLE (committed now)

Judged on the PRIMARY world room; scatter is the reported contrast. GO requires every GO condition at ALL THREE seeds (0, 1, 2). A STOP trigger fires if it holds at ≥ 2 of 3 seeds. The asymmetry is deliberate and in the conservative direction.

NOT-TESTABLE (checked first — nothing else is scored if this trips):

G6 reproduction contract fails. fit_operator_traced must reproduce fit_operator's final fast weights bitwise on ≥ 3 independent (seed, init) draws, and the five-arm run must reproduce Gate 004's published room numbers exactly: R1 = Rc = 1.0 at seeds 0/1/2; NF = 0.0 in every cell; n_dec = 11 / 14 / 10 at seeds 0 / 1 / 2 under the 48-goal escalation. If tracing or the appended MISMATCH arm moved any of these, the instrument perturbed the experiment and no anytime number is scored.

GO — all of:

Desaturation: TIE_frac ≤ 0.05 and NF_A > 0.
Resolution floor: n_dec ≥ 0.40 · ng and n_dec ≥ 12. (Gate 004 got DEC = 11/48 = 0.23, 14/48 = 0.29, 10/48 = 0.21.)
Null calibration: DEC_null ∈ [0.40, 0.60] — the NF construction still defines a ≈50% null, so DEC can be read against a coin instead of against nothing.
Not float dust: NF_rel ≥ 0.005 and NF_rel ≤ 0.50.
Not manufactured: neither STOP #2 nor STOP #3 fires.
Held-out discipline (inherited G0/G4): leaks == 0 in every arm map; every arm map holds ≥ SEL_K = 8 elites; cold cross-arm gap |median_g A(cold_a)/A(cold_b) − 1| ≤ 0.05.

STOP — any one triggers (at ≥ 2 of 3 room seeds):

Still saturating. TIE_frac > 0.20 or NF_A == 0. Verdict: the anytime lever does not fix the instrument. See "What a STOP costs M2."
The metric rewards any non-zero init, not retained structure (THE refuting control). F_triv ≥ 0.50 while W(TWO-SRC) − 1 ≥ 0.05. A count-matched, structure-free warm start reproduces ≥ half the anytime warm-over-cold advantage. This is the single most important "no" — adopting A would make M2's flagship warm-vs-cold curve pass for trivial reasons.
Relevance-blind. | W(MISMATCH)/W(TWO-SRC) − 1 | ≤ 0.05 while W(TWO-SRC) − 1 ≥ 0.05.
Dispersive, not resolving. NF_rel > 0.50 — the metric's own test–retest spread exceeds half the cold scale, so an M2-sized arm effect could never clear it.

PARTIAL:

Desaturation (G1–G4) passes but F_triv ∈ [0.25, 0.50): A is adopted only on the condition that M2 ships the RAND-MATCHED-vs-cold control curve alongside the headline; the headline alone would overstate.
Desaturation passes but DEC ∈ [0.40, 0.60] and R^A_1, R^A_c both within 1.00 ± 0.05 — see the honest-null hatch, which upgrades this to a GO on the instrument question.
Any GO condition splits across seeds (holds at 2 of 3): PARTIAL, never rounded up to GO.

The honest-null hatch (pre-distinguished, so a null cannot be spun)

These two outcomes look similar in a results table and have opposite implications for M2. Both readings are committed now.

Branch A — "the instrument is fine; the walls family simply has no signal." Signature: G1–G4 and G6 pass (TIE_frac ≤ 0.05, NF_A > 0, NF_rel ∈ [0.005, 0.50], DEC_null ∈ [0.40, 0.60]), and R^A_1 and R^A_c both land within 1.00 ± 0.05, and DEC ∈ [0.40, 0.60] — i.e. the arm comparison lands exactly on its calibrated coin. Reading: the metric resolves per-goal differences and reports the arms genuinely equal. This is a GO — the gate's question was about the instrument, and the instrument passed. It also corroborates Gate 004's STOP on a working instrument, which Gate 004 could not itself claim. M2 adopts A and proceeds on M1's built-for-purpose family.

Branch B — "the metric still saturates." Signature: STOP #1 — TIE_frac > 0.20 or NF_A == 0. Reading: Gate 004's arm ties remain uninformative (we still cannot tell a true null from a blind instrument), and M2's design must change before flagship budget is spent. The distinguishing number is TIE_frac with DEC_null, not the arm ratios.

Re-opening clause. If GO and R^A_1 lands outside 1.00 ± 0.05 — i.e. the anytime metric sees a retain difference the endpoint metric was blind to — that is not a retain result. It licenses a new pre-registered retain gate scored on A. Amending 004 post-hoc on this gate's numbers is scaffolding.

What a STOP costs M2 (committed now, so a STOP is actionable)

If STOP #1 fires, the resolution problem moves onto M1's world/budget design, and the following levers — not further metric shopping — are what M2 must then pre-register: per-goal difficulty calibrated so the fit does not converge within budget; a shortened FEW_ITERS chosen and frozen as an M1-time world-design parameter (never a post-hoc metric knob); a continuous per-goal score that does not bottom out on an attractor shared by all arms.

If STOP #2 or #3 fires, M2 must additionally treat any warm-vs-cold curve as unfalsifiable without a structure-free warm control shipped beside it.

Seeds / independent-draw plan

Seeds 0, 1, 2 — three required, not "two primary plus a tiebreak." This is a claim about an instrument, and instruments must hold on every draw; a two-of-three instrument is an instrument with an unexplained failure mode. seed(0) alone is repeatability, not robustness — the standing house lesson.

Each seed independently redraws every stochastic ingredient: the source me_emitter_run builds (shelves, columns), the RAND-MATCHED Fisher-Yates fillers and open-world re-rolls, the RAND2X double-budget build, the held-out goal synthesis (gen_family_disjoint), and all fit RNG. Each draw is therefore an independent repertoire + goal set, not merely fit noise — so TIE_frac, DEC, and DEC_null are three independent goal-set draws, not three reads of one.

Within a seed, all six arms are CRN-paired: identical goal set, one continuous fit-RNG stream, arms back-to-back, every ratio a paired per-goal difference (Gate 001's lesson). The RNG draw order per cell is frozen and load-bearing (the B-POC-3 stream-position lesson):

seed(s) → build shelves → build columns → build RAND-MATCHED fillers → build RAND2X
        → [room:    goals → TWO-SRC, ONE-A, ONE-B, RAND-MATCHED, RAND2X, MISMATCH]
        → [scatter: goals → TWO-SRC, ONE-A, ONE-B, RAND-MATCHED, RAND2X, MISMATCH]

AND rule: GO requires every GO condition at all three seeds on room. A STOP trigger fires at ≥ 2 of 3 seeds. Seed-splits on GO conditions are PARTIAL, never rounded to GO.

Instrumentation contract (how the trace may be taken)

The S3 forward_with_learning_traced precedent, tightened one notch:

fit_operator_traced[M] is a pure sibling in src/esper_evolution.mojo. Production fit_operator is not touched. The traced sibling replicates the schedule exactly — same alpha_rate/sigma_rate derivation, same exp annealing, same evolve_fast_weights[M] call with the same arguments in the same order.
It snapshots the weight vector only, never a score. At the top of iteration t for t ∈ K, plus one final post-loop snapshot, it memcpys POLICY_DIM fast weights into a caller-provided, pre-allocated |K| × POLICY_DIM buffer. No allocation inside the loop, no rollout inside the loop, nothing inside the SIMD hot path.
Scoring happens after the fit returns, in a plain loop over the snapshot buffer via the existing policy_score (which draws zero RNG). This is strictly stronger than the S3 precedent: the measurement cannot perturb the ES stream even in principle.
The k = 29 snapshot is "after exactly 29 ES steps" for every arm — post-loop for the ITE arm (iters = 29), top-of-iteration-29 for the rest (iters = 30).
Budget: ~6 measurement rollouts × 4 fits × 6 arms × 48 goals × 2 worlds × 3 seeds ≈ 0.4% of the fit's own rollout cost. The count is reported. These rollouts are measurement-only, excluded from any budget-parity claim, and may never feed a fit.

Expected wall-clock: Gate 004's ~3 min plus the sixth arm.

Anti-stone-soup clause

The fit stays a single COLD fit per arm at the unchanged budget: no per-goal hand-staging, no mid-fit boosts, no bespoke tuning. Specifically, the following are scaffolding, and a pass obtained with any of them is a documented negative regardless of the raw number:

Metric shopping. Computing A, several τs, trapezoid AUC, and a slope fit, then reporting whichever separates. Alternatives may be printed as diagnostics; the verdict is A alone. Naming a different primary after the run is a documented negative.
Ladder shopping. Trying {1,2,3,5,8,13,29}, {1,4,16}, or any denser/sparser ladder and keeping the resolving one. Per-arm, per-world, or per-seed ladders; giving the treatment arm more checkpoints than the controls.
Shortening FEW_ITERS/SEL_FIT_ITERS to manufacture resolution. A shorter fit desaturates the endpoint metric too, which would make this gate's comparison meaningless. The run uses the unchanged 30/29-iteration budget so the comparison is strictly metric-vs-metric on identical fits. (Budget reduction is an M1 world-design lever, pre-registered there, never here.)
Touching fit_operator, evolve_fast_weights, run_family_select, pool_trials, me_emitter_run, EliteMap, gen_family_disjoint, the BUILD_*/FEW_*/SEL_* constants, or the RNG draw order — including "harmless" refactors. If Gate 004's published room numbers (R1 = Rc = 1.0, NF = 0.0, n_dec = 11 / 14 / 10) do not reproduce byte-for-byte, the run is NOT-TESTABLE and must be re-run, never re-baselined and never explained away.
Inserting MISMATCH anywhere but last, or hand-choosing its permutation instead of the fixed cyclic shift by 1.
Feeding a checkpoint rollout back into a fit, counting measurement rollouts toward any arm's budget, or spending extra ES iterations on any arm.
Reporting mean per-goal ratios instead of median (the standing Route A/C one-outlier lesson).

Frozen constants: ladder K = {1,2,4,8,16,29}; δ_rel = 0.02; NUM_GOALS = 24 with a single 48-goal escalation at N_DEC_MIN = 8; seed set {0,1,2}; arm order with MISMATCH last; cyclic-shift-by-1 permutation; FEW_ITERS = 30 / SEL_FIT_ITERS = 29 unchanged.

Scope of new code: fit_operator_traced, the post-fit checkpoint scoring loop, the A / NF_A / TIE_frac / DEC_null / W / F_triv computations in the harness, the MISMATCH arm's retrieval-target indirection, and un-discarding the already-computed d_unif trace. No new ES architecture, no new world, no new scoring primitive — policy_score and calculate_fitness are untouched.

Result — STOP (2026-07-27)

Run: mojo run -I src tests/test_anytime_metric.mojo, seeds 0/1/2, both worlds, ~13 min. Instrumentation: fit_operator_traced (src/esper_evolution.mojo) + run_family_select_traced (src/transfer.mojo); harness tests/test_anytime_metric.mojo. Scored on the PRIMARY world room, at the escalated 48-goal count (every seed escalated, as in Gate 004). Budget unchanged: FEW_ITERS = 30 / SEL_FIT_ITERS = 29.

NOT-TESTABLE check (G6) — PASSED, so the numbers below are scorable

Bitwise sibling: tests/test_fit_traced.mojo — fit_operator_traced reproduced fit_operator's final weights with 0 / 294 differing floats on 3 / 3 independent (seed, init) draws. The k == iters snapshot equals the post-loop weights exactly, so the ladder means "after exactly k ES steps" as registered.
Five-arm reproduction: R1 = 1.0 and Rc = 1.0 at seeds 0/1/2; NF = 0.0 in every cell; n_dec = 11 / 14 / 10 at seeds 0/1/2 under the 48-goal escalation. Byte-for-byte the Gate-004 table.

Scored numbers (`room`, 48 goals)

seed	TIE_frac	NF_A	NF_rel	DEC	DEC_null	W(TWO)	W(RAND-M)	F_triv	\|W(MM)/W(TWO)−1\|	G-MM
0	0.354 (17/48)	2.28e-4	0.029	0.208 (10/48)	0.438	5.73	5.43	0.936	0.893	16.1×
1	0.313 (15/48)	3.50e-4	0.043	0.229 (11/48)	0.438	4.49	4.98	1.142	0.788	11.2×
2	0.354 (17/48)	3.34e-4	0.030	0.250 (12/48)	0.479	2.39	2.59	1.145	0.737	15.4×

R^A_1 = 1.000 and R^A_c = 1.000 at all three seeds. G4 cold cross-arm gap on A = 0.000 (bar: ≤ 0.05). leaks == 0; every arm map ≫ SEL_K = 8 elites.

Verdict against the frozen bars

STOP #1 — still saturating: FIRES at 3/3. TIE_frac = 0.354 / 0.313 / 0.354, all above the 0.20 trigger and ~7× the 0.05 GO bar. Roughly a third of held-out goals still produce bitwise-equal A for two independent cold fits.

STOP #2 — rewards any non-zero init, not retained structure: FIRES at 3/3. F_triv = 0.936 / 1.142 / 1.145, all ≥ 0.50, with W(TWO-SRC) − 1 = 4.73 / 3.49 / 1.39 (well past the 0.05 liveness guard). At seeds 1 and 2 the count-matched scrambled repertoire is better than the genuinely retained one. This is the gate's "single most important no".

GO conditions 1 and 2 fail; 3 (DEC_null = 0.438 / 0.438 / 0.479), 4 (NF_rel = 0.029 / 0.043 / 0.030) and 6 pass. Two failed GO conditions plus two STOP triggers at 3/3 ⇒ STOP, not PARTIAL: the PARTIAL clauses all require desaturation (G1–G4) to have passed.

Honest-null hatch: Branch B, not Branch A. Branch A required G1–G4 to pass; G1 failed. So the flat R^A_1 = R^A_c = 1.000 is not evidence that the walls family has no signal — it is the residual saturation showing through, exactly as the gate pre-committed.

What the controls bought (the reason this STOP is informative and not just a "no")

The two controls disagree, and the disagreement is the finding:

Control 1 (RAND-MATCHED) fires. W(RAND-MATCHED) ≈ W(TWO-SRC) on every seed. Nearly 100% of the ~2.4–5.7× anytime warm-over-cold advantage survives replacing the retained structure with Fisher-Yates-scrambled weights re-rolled in the open world.
Control 2 (MISMATCH) does NOT fire. |W(MISMATCH)/W(TWO-SRC) − 1| = 0.89 / 0.79 / 0.74, far above the 0.05 relevance-blind trigger, with the G-MM precondition satisfied 11–16× over (bar: 2×). W(MISMATCH) = 0.62 / 0.95 / 0.63 — a mis-retrieved real elite is worse than a cold start.

Together: A is sensitive to where in BC space the warm start sits (aiming at the wrong goal hurts), but blind to whether the policy carries retained structure (a scrambled policy landing in roughly the right BC region does just as well). What the anytime metric measures is initialization placement, not transfer. Had only Control 1 been registered, the picture would have been the flat "warm starts are trivially good" that the gate-designer warned Control 1 alone could not distinguish.

Anti-stone-soup confirmation

Verdict scored on A alone. Ladder unchanged (K = {1,2,4,8,16,29}); δ_rel = 0.02 unchanged; no τ/AUC/slope alternative was computed, let alone substituted; FEW_ITERS / SEL_FIT_ITERS unchanged; fit_operator, evolve_fast_weights, run_family_select, pool_trials, me_emitter_run, EliteMap and every BUILD_*/FEW_*/SEL_* constant untouched (G6 proves it). All aggregates are medians of paired per-goal ratios. Checkpoint rollouts (6 per fit, vs 1920 fit rollouts) never fed a fit and are excluded from every budget claim. Ratio clamps fired 12 / 33 / 20 across the endpoint ratios — reported, not hidden; they are exact-reach denominators (d = 0), themselves a symptom of the saturation.

One registered detail that needed care (recorded, since it nearly voided the gate)

The gate froze "MISMATCH appended LAST" to protect the shared fit-RNG stream, but its stream diagram did not cover the 48-goal escape hatch re-running the cell inside the same stream. Placing MISMATCH at the end of the 24-goal pass still puts its 24 draws before the 48-goal pass draws its goals — and the 48-goal pass is the scored one. Measured, not theorised: that placement moved seed 0's room count from 11 to 12, i.e. G6 fails and nothing is scorable. The fix keeps the registered intent exactly: MISMATCH runs only in the pass that is actually scored, still last, and the scored-pass decision is made from the five arms' endpoint decidable count, which is already in hand and costs no RNG. All Gate-004 draws then sit at identical stream positions, which is what the 11 / 14 / 10 reproduction confirms.

scatter is reported in the log as a contrast only. Because MISMATCH consumes RNG at the end of the scored room cell, scatter's goals are drawn from a different stream position than in Gate 004, so its counts are not comparable to the published Gate-004 scatter numbers. Its NF = 0.0 (a structural property, independent of which goals are drawn) does still hold. The verdict above uses room only, as registered.

What this costs M2 (per the pre-committed "What a STOP costs M2")

STOP #1 moves the resolution problem onto M1's world/budget design — not onto further metric shopping. The levers M2 must pre-register: per-goal difficulty calibrated so the fit does not converge within budget; a shortened FEW_ITERS chosen and frozen at M1 time as a world-design parameter (never as a post-hoc metric knob); and a continuous per-goal score that does not bottom out on an attractor shared by all arms.

STOP #2 additionally binds M2: any warm-vs-cold curve is unfalsifiable unless a structure-free warm control ships beside it. Control 2's result sharpens this — the control that belongs beside the headline is a BC-placement-matched one (RAND-MATCHED), because a merely mis-aimed warm start is already separable and would flatter the headline.

Gate 004's STOP stands, untouched and now independently reproduced. The re-opening clause is not triggered: R^A_1 did not land outside 1.00 ± 0.05.

FilesExpand file tree

005-anytime-transfer-metric.md

Latest commit

History