B · FA Reference Parameters (V1)

1.112026-08-29

All FA-related parameters below are the V1 production values: calibrated by the expert group and engineering tuning, frozen at release under the tag fa-v1 with a frozen copy hyperparams_v1.0.frozen.yaml, and governed by the registry's tri-state parameter discipline (frozen / initial-production-effective / disabled). The status column uses “frozen” for permanently settled values and “initial” for values that are initial-production-effective design values of V1; in this registry's governance, “initial” does not mean “to be calibrated later” — the unified calibration experiment (Section 12.2) is a planned sensitivity study, not a prerequisite for these values.

Table 12. FA-1. Trichotomy group and block weights.

Layer	Weight	Status
Capability / ecosystem composite (top level)	$0.60 / 0.40$	frozen
Capability block / edge block (within capability)	$0.55 / 0.05$	frozen
Ecosystem block / commerce block (within ecosystem)	$0.25 / 0.15$	frozen
Developer block / academic block (within ecosystem)	$0.50 / 0.50$	frozen
Ripple / ledger (within academic block)	$0.50 / 0.50$	frozen

Closed form: with the V1 values, (11)–(13) reduce to $\mathrm{FA}_i = 100\,(0.55\,\mathrm{cap}_i + 0.05\,\mathrm{edge}_i + 0.25\,\mathrm{eco}_i + 0.15\,\mathrm{com}_i)$ , and the FA value reads as the percentage of the best-weighted combination in the included set.

Table 13. FA-2. Capability block: seven benchmark sources with expert weights.

Benchmark source	Weight	Status
Composite model-capability evaluation	$0.40$	frozen
Dimensional model-capability evaluation	$0.15$	frozen
SWE-bench solve rate	$0.15$	frozen
Aider edit accuracy	$0.05$	frozen
LiveCodeBench	$0.05$	frozen
HF Open LLM Leaderboard composite	$0.10$	frozen
LMArena Elo	$0.10$	frozen

Table 14. FA-3. Developer block: source weights (heavy-tail scale).

Source	Weight
GitHub stars	$0.35$
HF downloads	$0.30$
HF likes	$0.15$
GitHub repos	$0.10$
HF models	$0.10$

Table 15. FA-4. Commerce block: four-factor weights and pricing rule.

Factor	Weight	Rule
value	$0.20$	flagship capability / flagship blended price (same-model pairing)
price	$0.20$	inverse blended price, heavy-tail
free	$0.10$	boolean presence of official free API access (true-zero discipline)
momentum	$0.50$	token adoption, linear direct division, leaderboard floor

The blended price is $\mathrm{blend} = (3\,\mathrm{prompt} + 1\,\mathrm{completion})/4$ (3:1 input–output mix); special price tiers ( $:batch$ , $:contributor$ , $:free$ ) are excluded from value/price.

Table 16. FA-5. Edge block: parameter-band boundaries (total parameters, billions).

Band	Total parameters (B)
1	$(0,\,2]$
2	$(2,\,10]$
3	$(10,\,100]$
4	$(100,\,300]$
$> 300$ B: not counted in the edge block.

Within a band, a unit scores its best-model capability divided by the band-wise cross-organization maximum over the included set; a missing band contributes $0$ and is always present in the mean (true-zero discipline, Section 6.4).

Table 17. FA-6. Unified scale contract (factor-level).

Class	Formula	Applies to
Linear	$v/\max$	benchmark counts, edge, momentum
Heavy-tail	$\log_{10}(1+v)/\log_{10}(1+\max)$	ecosystem sources, ripple, value
Inverse	$1/v$ , then linear or heavy-tail	price
Boolean	$0/1$ as-is	free

Table 18. FA-7. Missing-value classification (four-state). States $P$ (true zero), $Z$ (numerical absence), $A$ (leaderboard floor), $F$ (technical unavailability) match Section 6.4.

State	Class	Handling
$P$	True zero (participation with no observable activity)	weight zero, kept present; renormalization forbidden
$Z$	Numerical absence (source does not cover the unit)	re-normalize over present dimensions
$A$	Leaderboard floor (below the printed cutoff)	value by the last readable rank, taken dynamically
$F$	Technical unavailability (source outage / schema failure)	carry most recent valid value forward with alert

The classification of each factor for each unit is registered and versioned, so that the handling of every missing value is auditable (Section 6.4).

Table 19. FA-8. Sustained-leadership ( $D$ ) parameters.

Parameter	Value
Flagship thresholds	$\{1,2,3,5,10\}$
Edge-side thresholds	$\{1,3,5\}$
Weight function	$f(n) = \log_2(1+n)$
Tier weights $w_y$	$[1.00,\,0.85,\,0.70,\,0.50,\,0.30]$
Leaderboard weights $\beta_d$	MCE $0.35$ ; SWB $0.25$ ; LMA $0.20$ ; AID $0.10$ ; HFL $0.10$
Cap $D_{\mathrm{cap}}$	$+10\%$ of FA (frozen)
Reset	immediate to zero on falling out of all thresholds; no decay tail

Board codes: MCE = model-capability evaluation (Table 13); SWB = SWE-bench; LMA = LMArena; AID = Aider; HFL = the HF Open LLM Leaderboard.

Table 20. FA-9. Adjustment caps.

Cap	Value	Status
$D_{\mathrm{cap}}$	$+10\%$ of FA	frozen
$R_{\mathrm{cap}}$	$\le 5\%$ of FA	initial

Table 21. FA-10. Ledger artifact tiers (adjudicated research artifacts).

Tier	Meaning	Value
$T_1$	record-level	$1.0$
$T_2$	milestone-level	$2.0$
$T_3$	first-level	$3.5$
$T_4$	paradigm-level	$6.0$
$T_5$	frontier-opening	$10.0$

Verification factor $V \in \{0.7 \text{ (public-list)}, 0.9 \text{ (committee)}, 1.0 \text{ (independent)}\}$ ; the ledger value is $\sum_e T_e V_e$ , normalized linearly.

Table 22. FA-11. Presentation-layer display-noise parameters (non-RI). These are display-layer parameters of the nominal line (Section 6.8); they are not RI calibration values.

Parameter	Value
Relative amplitude $\sigma_0$	$20$ bps
Amplitude cap	$75$ bps
Noise half-life	$15$ min
Taper window	$15$ min
Jump probability	$0.15\%$ per min per unit
Jump magnitude	uniform $(3,5)\sigma$
Display jitter (sub-minute)	$\sigma_j = 10$ bps

The live-signal coefficient $k_{\mathrm{ri}}$ and all RI calibration values are reserved and are not listed.

Back to contents | ← A · Symbols, Notation, and the Code–Symbol Binding | C · First-Assessment Roster and Inclusion →