Objective feel metrics: what can pre-screen 'feel' before Chris plays (2026-10-17)

Objective feel metrics: what can pre-screen “feel” before Chris plays

Ask (Chris, 2026-10-16 23:40): “figure out how you can make tools that let you make ‘feel’ judgements; like we did for the sound, there may be more objective ways to identify when something is more or less likely to feel good.” The sound precedent is bh-synth-sfx + tools/sound_map/analyze_ref.py: measure a candidate against references Chris already liked, so Claude can throw out the obvious misses before his ear.

Outcome: tools/feel_probe/ (MVP, branch feel-probe) + scripts/feel-probe.sh + skill bh-feel-probe. It runs seeded headless autopilot runs with a read-only observer and prints nine run metrics, a per-weapon dead-pick table, and a static snap-vs-ramp scan, each with PASS / WARN / FAIL bands.

The uncomfortable part first: most bands below are not measured on players. They come from practitioner convention (Swink, Vlambeer, Valve) or from this game’s own history of Chris verdicts. The probe is a filter that says “this is likely to read as X”. It cannot say “this feels good”. Bands move only when a Chris verdict disagrees with them (see Calibration).

1. Literature: which feel proxies are measurable at all

A literature pass (one research agent, web search). Only four of these sources were re-checked against the web this session: the Eiserloh trauma model, Claypool’s latency work, Clark et al. on near-misses and the KOF XIII hit-stop table. The other citations are from memory of the literature. Verify them before quoting.

Topic	Metric	Reference band	Source	Confidence
Response	input → first visible response	< 100 ms reads as “direct”; ~240 ms is the human correction cycle	Swink, Game Feel (2008)	expert practice
Response	latency vs. precision/deadline	tight-deadline actions degrade from < 100 ms of local lag	Claypool & Claypool, “Latency Can Kill” (2010), https://sites.cs.ucsb.edu/~almeroth/classes/W10.290F/papers/claypool-10.pdf	measured (FPS games, not twin-stick)
Hit feedback	hit-stop frames	7 frames (light) to 11 (heavy) @60 fps in fighting games; 2-6 frames for rare events in a swarm game	KOF XIII frame data https://srk.shib.live/w/The_King_of_Fighters_XIII/Character_States/Hitstun_Blockstun_state	measured for fighting games, folklore for swarms
Hit feedback	flash duration	1-3 frames of white flash	Nijman, “The Art of Screenshake” (2013); Jonasson & Purho, “Juice It or Lose It” (GDC EU 2012)	expert practice
Shake	trauma model	shake = trauma² × max offset; linear decay 0.5-1.0/s; events add 0.2-0.6, clamp at 1	Eiserloh, “Juicing Your Cameras With Math” (GDC 2016)	model solid; numbers are implementers’ defaults
Shake	saturation	share of run time with trauma > 0.3; near 100% means the shake carries no information	derived	guess
Envelope	attack vs decay	attack < 100 ms (1-3 frames); decay 3-10× longer reads “punchy”	Thomas & Johnston, The Illusion of Life (1981); Swink	expert practice
Telegraph	warning before damage	≥ 240 ms floor, ≥ 400 ms comfortable	derived from Swink’s correction cycle	derived
Pacing	build-up / peak / relax	L4D Director: peak held 3-5 s, relax 30-45 s	Booth, “The AI Systems of Left 4 Dead” (AIIDE 2009)	expert practice
Pacing	flow channel	challenge tracks skill	Csikszentmihalyi (1990); Chen, “Flow in Games” (2006)	theory, no numbers
Reward	first reward latency	first reward within 10-20 s, then growing gaps (VS-likes)	practitioner heuristic	folklore
Reward	variable ratio	unpredictable timing sustains engagement	Ferster & Skinner (1957)	measured (in animals)
Choice	dead picks	an option picked far below uniform, or a dominant one above ~40%, is a smell	StS/Hades pick-rate vs. win-rate practice	thresholds are guesses
Readability	bullet vs. background contrast	≥ 3:1 (WCAG non-text), aim ≥ 4.5:1 for hostile bullets	WCAG 2.1 §1.4.3/1.4.11	measured for text; adaptation
Near miss	graze / near-miss rate	near-misses raise the urge to keep playing only when the player controls the risk	Clark et al., Neuron (2009), https://pmc.ncbi.nlm.nih.gov/articles/PMC2658737/	measured (gambling)

The most defensible proxies are response latency, the intensity series, and graze/near-miss rate. Shake, hit-stop and choice-variety numbers are practitioner folklore. Calibrate those on our own data or don’t use them.

2. What the repo already measures (inventory)

Tool	Measures	Feel relevance
tools/balance_probe	survival median/p25/p75, dead < 30 s, HP lost/s, damage share of watched weapons (Sim.damage_by_weapon), investment	dead picks (the Orbit 0.1% and Nova 1.4-1.7% numbers came from it), difficulty
tools/soak (SoakRun)	invariants, per-upgrade “dud”/“overpowered” impact (dealt/taken ±180 ticks), softlock, digests	dud upgrades are a choice-feel signal; on_tick is the read-only hook the feel probe uses
tools/economy_sim	runs-to-buy per shop item, career curves	meta reward cadence (across runs, not within one)
telemetry D1 (bh-telemetry)	runs (DPS, level, duration), samples (fps, render_ms, weapons), hitches, deaths	latency and frame pacing on real devices: the only real source for “response” metrics
tools/autopilot_bench	autopilot turn rate, snaps > 60°/s, boss damage per minute	humanised-play realism, which feel metrics depend on
tools/vfx_library	0-5 subjective scores per effect, MISSING cards for undrawn sim state	a reviewer’s labels, not a metric
tools/look_metrics	saturation/value/contrast traits on captured frames vs. Nova Drift	readability and “too bright” verdicts
tools/weapon_vfx/supernova_bench	clipped-pixel counts for stacked bursts	a direct proxy for “too white hot”
docs/research/2026-10-12-unremapped-alpha-audit.md	glow alpha vs. 8-bit (2.6-9× too bright)	the “hazy / glows soft” verdicts
polish categories §2 (bh-overnight-audit-loop)	snap-vs-ramp grep heuristic (manual)	the “felt sudden” class of verdicts

There was no per-second pacing series, no reward-cadence metric (one data point: first level-up panel at 18 s in the alpha audit), no automated snap scan, and no near-miss measure. The probe adds those.

3. Chris’s past verdicts: the labelled data

Mined from memory, release docs, specs and commits. 34 verdicts. The ones a metric can be checked against:

#	When	Subject	Verdict	Measured state then
V1	2026-07-16	low-HP border	“too much too quickly… should gradually get stronger from about half hp”	set_low_hp(hp <= max_hp*0.3), a boolean snap to 70-100% glow
V2	2025-07 → 2026-10	Orbit	ruled dead twice (“near-useless”)	damage share median 0.1% (Toby run: 79 dmg vs Blaster 68.7k)
V3	2026-10-01	Supernova	“increase its dmg… halve its rate of fire”	2.3% of run damage
V4	2026-10-13	Supernova	“doesn’t trigger until enemies close”	actually the 10 s cooldown: a dead-air gap, not range
V5	2026-10-16	App Clip 5× XP	“way too many upgrades”, “spent most of the time pressing upgrade cards”	settled on 3× XP plus a ≥ 10 s floor between level-ups
V6	2026-10-15	overall difficulty	“we have been tuning for Toby”	Hard median 58.7 s, 8/40 dead < 30 s
V7	2026-07	late snowball	Toby “became seriously overpowered”	DPS 321 at 31 min
V8	2026-09-30	evolved Nova / FX v2	“too white hot”, “makes the screen very white”	8 stacked additive bursts, core ×1.6
V9	2026-10-10…13	HDR	“washed out”, later “glows a bit soft”	saturation seeds 1.0; glow alphas 2.6-9× 8-bit
V10	2026-10-11	death sequence	hard cut → wanted “huge explosion… slow motion… zoom”	0.3× slow-mo 0.8 s, 1.3× zoom, 3.2 s total, “all good”
V11	2026-07	render timing	“effects don’t fire… then suddenly all fire together”	sim/render tick desync
V12	2026-10-07	elites	“too much health”	Orbiter 432 HP

There are no Chris verdicts yet on hit-stop, shake strength or enemy hit flash. Any band for those is uncalibrated.

4. The probe’s metrics: definition, band, provenance, which verdict it predicts

Run: scripts/feel-probe.sh (tools/feel_probe/feel_probe.gd). The observer (FeelObserver) is read-only (tests/test_feel_probe.gd pins a run bit-identical with it attached) and bins per-second series. Rows are medians across seeds.

Metric	How measured (headless sim)	Band (PASS / WARN)	Band source	Predicts	Confidence
dead air share	seconds with zero kills after 10 s / counted seconds	≤ 0.10 / ≤ 0.25	L4D build-up rhythm + VS-likes’ constant kill stream; Hard main = 0.04	V4 (Supernova “doesn’t trigger”) at run scale; untested on a Chris verdict	low: boss fights legitimately have few kills
dead air longest	longest zero-kill streak	≤ 6 s / ≤ 12 s	same	same	low
intensity build	mean threat near pilot (bodies + bullets within 360 px), last third / first third of 10 s windows	≥ 1.5× / ≥ 1.1×	L4D Director build-up	V7 (snowball: build collapses when DPS outruns spawns)	medium
intensity variation	CV of 10 s window means	≥ 0.20 / ≥ 0.10	L4D peak/relax	none yet	low
first level-up	run seconds to level 2	≤ 15 s / ≤ 25 s	practitioner 10-20 s; alpha-audit capture 18 s	none yet	low-medium
median level-up gap	median gap between level-ups, two-sided	10-25 s / 7-40 s	low side calibrated on V5 (5× XP → 6.5 s “way too many”; 3× → 8.7 s accepted with a 10 s floor)	V5	medium (one verdict)
longest reward drought	longest gap incl. the open tail to run end	≤ 45 s / ≤ 75 s	folklore	none yet	low
worst 3 s HP loss	max HP lost in any 3 s / max HP	≤ 0.35 / ≤ 0.60	“came out of nowhere” deaths	V6 partly; does NOT separate Hard (0.36) from Normal (0.40)	low
graze-s per hit	seconds with a hostile within 32 px of contact and no hard hit ÷ hard hits taken (Sim.player_hard_hits delta; DoT ticks count as HP loss, never as hits)	≥ 3 / ≥ 1	danmaku graze; Clark 2009 near-miss	none yet	low: autopilot dodging is not human dodging, so compare runs and don’t quote the absolute value
dmg share per weapon	damage_by_weapon[w] ÷ all damage dealt since w was first owned (≥ 60 s owned); weapons with no damage archetype are util	≥ 5% / ≥ 2%	B1 patch targets (Orbit/Nova → ~12%)	V2, V3 directly	high for the direction, medium for the cut
unattributed damage	damage_by_weapon["unattributed"] ÷ total	≤ 2% / ≤ 8%	guards the share rows	n/a	high
snap-vs-ramp scan	static regex over render/ ui/ fx/ audio/ main.gd: a visual property set by A if <qty> < x else B, or a set_*() fed a comparison on HP/charge/danger/distance/time; zero-compares and lines with lerp/clamp/ease skipped	0 findings	polish categories §2	V1 directly (the unit test reproduces the original line)	high for recall on that pattern; it misses snaps written any other way

Deliberately NOT in the MVP:

• Hit-feedback latency (damage → fx event, in ticks). The sim emits death and other fx events in the same tick by construction, so a headless measure is always 0. The real latency is in the render path and device frame pacing. It belongs to telemetry (render_cpu_ms, hitches) and bh-diagnose-render-timing.
• Shake/flash/hit-stop budgets. These need a render capture, and there are no Chris verdicts to calibrate them against.
• Bullet contrast. It needs frames (tools/look_metrics style). This is a good next metric: hostile-bullet pixels vs. the local background luminance, with a ≥ 3:1 band.

5. Current main, measured (2026-10-17, origin/main 43a88350)

10 seeds (5000-5009), cap 600 s, cobalt, fresh save, autopilot (SkilledAutoPlayer + skill profile):

Metric	Normal / median	Normal / weak	Hard / median
run length median	262 s	202 s	64 s
kills/s	0.95	0.97	4.22
dead air share	0.51 FAIL	0.42 FAIL	0.04 PASS
dead air longest	19 s FAIL	13 s FAIL	1 s PASS
intensity build	2.01× PASS	1.76× PASS	3.85× PASS
intensity CV	0.60 PASS	0.51 PASS	0.48 PASS
first level-up	27.8 s FAIL	28.3 s FAIL	18.5 s WARN
median level-up gap	17.5 s PASS	16.4 s PASS	8.0 s WARN (card spam side)
longest drought	50 s WARN	37 s PASS	19 s PASS
worst 3 s HP loss	0.40 WARN	0.18 PASS	0.36 WARN
graze-s per hit	1.12 WARN	0.83 FAIL	0.98 FAIL
unattributed	0.0% (max 1.7%)	0.0% (max 3.5%)	0.0% (max 1.7%)

Dead-pick rows, Normal/median (the share while owned; n = runs owning it ≥ 60 s): blaster 24.3% (10), railgun 9.9% (2), pulse 18.9% (1), momentum_core 18.5% (1), thunder_wake 17.6% (1), beam 11.5% (1), blade 4.9% WARN (3), turret 4.3% WARN (3), scythe 2.1% WARN (1), scatter 2.0% WARN (1); chalice and scar_tissue print n/a: their damage is credited to no weapon on main (see §6.2), so their 0.0% is an attribution gap, not a measured share. An earlier 12-seed run with the plain autopilot (no profile) had orbit 0.1% FAIL and chronowell 0.0% (now util). The Normal/median shares do not add up to 100% because Rift Bolt damage is credited to aim, which the table does not list.

Snap scan on main: 2 findings, ui/hud.gd:271 and :323. The HP bar fill (P1 and P2) switches to a fast 9 rad/s alpha pulse the instant HP crosses 30%. That is the same pattern as V1’s border, on the bar instead of the screen edge.

What-if runs (measurement only, nothing changed on main)

• App Clip XP (--force=Difficulty.NORMAL_XP_MULT=x,…BOOST_LEVELS=99, 8 seeds, 180 s): median gap 14.5 s at 1.5×, 8.7 s at 3×, 6.5 s at 5×. This is the calibration for the low side of the median-gap band.
• Denser Normal start (--force=Difficulty.NORMAL_BIOMASS_START_FRAC=0.35, 10 seeds, 600 s): survival unchanged (264 s vs. 262 s), kills/s 0.95 → 1.56, dead air 0.51 → 0.36, first level-up 27.8 → 20.3 s. This is a candidate dial for Chris, not a proposal to merge.

6. Top feel issues the probe flags on main now

1. Normal mode’s first minutes are probably empty. Half of all seconds have no kill and the first level-up takes ~28 s, against 18.5 s on Hard. The cause is the design knob: NORMAL_BIOMASS_START_FRAC 0.15 ramping over 600 s. Normal is the 1.1 default, so this is the first thing a new player feels. Confidence: medium. The direction is clear, but no Chris verdict exists on Normal pacing yet, and the autopilot kills less efficiently than a human. The 0.35 what-if fixes a third of it without moving survival. DECISION (Chris).
2. Damage attribution blind spot. Two weapons’ damage is credited to no weapon on main, so they read 0 on the death screen, in telemetry and in this probe. Chalice pool damage is applied in Sim._update_chalice_zones, outside the per-weapon damage_context loop. Scar Tissue’s discharge runs inside Sim._hurt_player (sim.gd ~3373 → weapon_scar_tissue.gd) under whatever context is current, never scar_tissue. In the two Chalice runs unattributed was ≤ 1.7% of all damage over 130-217 s of ownership, so Chalice is probably near-dead too, but that bound is loose. Scar Tissue’s real share is unknown. The probe marks both n/a. Branch damage-attribution (W10, 2026-10-17) is fixing the crediting; drop them from UNATTRIBUTED_WEAPONS in feel_probe.gd once it lands, then re-run. Orbit was still 0.1% in the one run that took it (B1 is only live on dev).
3. HP-bar 30% snap (ui/hud.gd:271/323): a fast pulse switches on at a hard threshold. This is the exact class Chris called “too much too quickly” for the border. The fix pattern is a ramp from ~50% HP with eased pulse depth.

Weaker flags: graze-per-hit is about 1 grazing second per hard hit (WARN on Normal/median, FAIL on Normal/weak and Hard), well below the ≥ 3 band. The band is a guess and the number is autopilot-dependent, so compare branches rather than reading the absolute value. (Round 1 of this doc counted every DoT tick as a hit and reported 0.24-0.73; fixed after review.) Hard’s median level-up gap of 8.0 s sits on the card-spam side of the V5-calibrated band.

7. Calibration log (move a band only with a verdict)

Date	Band	Change	Verdict that forced it
2026-10-17	median level-up gap	made two-sided, low side 10 s pass / 7 s warn	V5 App Clip 5× XP (probe: 6.5 s)

8. Next steps

• Bullet-contrast metric from captured frames (tools/look_metrics pattern, band ≥ 3:1).
• Exclude boss-alive seconds from dead air (a boss fight is not dead air).
• A humanised-autopilot run (humanize) as the default for graze metrics.
• A shake/flash budget once Chris gives a verdict on either.
• Ask Chris for a verdict on Normal’s opening minute. That one answer calibrates the dead-air and first-level bands.