clay-borg/specs/FindingRegister.md

The finding register

Design findings about GROUND, with their reproductions. Governed by GameDesign.md (admissibility, kinds, states, metrics) and ADR-0012. Reported by make design.

Split out of GroundRules.md §Underdetermined on 2026-08-05 when that file crossed the ~400-line loadability limit. ADR-0012 D2 said "no new file" and this is a new file — but D2's substance was one register, not a second mechanism competing with the first, and that holds: this is §Underdetermined's register, moved, still driving off the same provisional/ruling machinery. D2 also named the awkwardness this resolves — a finding about the engine sitting in a document about the game.

The U-items themselves, with their defaults and rulings, stay in GroundRules.md §Underdetermined; this file tracks them as findings.

This section is the design-finding register (ADR-0012 D2). It was the register for dataset ambiguities already; CB-WP-0022 extended it to all five kinds rather than building a second one beside it. Admissibility, kinds, states and metrics: GameDesign.md. Reported by make design.

id	kind	state	reproduction	role	raised	owner
U1	underdetermined	applied	—	default	2026-07-31	ground-game
U2	underdetermined	applied	scenarios/ground/gr-d01-darvo-trigger.yaml	default	2026-07-31	ground-game
U3	underdetermined	applied	—	default	2026-07-31	ground-game
U4	underdetermined	applied	—	default	2026-07-31	ground-game
U5	underdetermined	applied	—	default	2026-07-31	ground-game
U6	underdetermined	applied	—	default	2026-07-31	ground-game
U7	underdetermined	applied	—	default	2026-07-31	ground-game
U8	underdetermined	applied	—	default	2026-07-31	ground-game
U9	underdetermined	applied	—	default	2026-07-31	ground-game
U10	underdetermined	applied	—	default	2026-07-31	ground-game
F11	inert	applied	scenarios/ground/gr-p05-solve-legality.yaml	counterexample	2026-08-02	clay-borg
F12	degenerate	note	—	—	2026-08-01	clay-borg
F13	inconsistent	withdrawn	scenarios/ground/gr-e01-threshold-reachable-2p.yaml	counterexample	2026-08-01	clay-borg
F14	unplayed	applied	games/ground/examples/attack-value.rs	counterexample	2026-08-01	clay-borg
F15	underdetermined	note	—	—	2026-08-05	clay-borg
F16	inconsistent	withdrawn	games/ground/examples/difficulty.rs	counterexample	2026-08-05	clay-borg
F17	degenerate	raised	games/ground/examples/attack-value.rs	counterexample	2026-08-06	ground-game
F18	inert	raised	games/ground/src/edition.rs::card_text_tests::the_engine_reads_only_part_of_what_it_vendored	counterexample	2026-08-06	clay-borg
F25	inert	raised	games/ground/src/edition.rs::card_text_tests::the_engine_agrees_with_the_editions_own_numbers	default	2026-08-08	clay-borg
F24	inert	raised	games/ground/src/edition.rs::card_text_tests::the_hardcoded_deck_still_matches_the_edition	default	2026-08-07	clay-borg
F19	degenerate	applied	crates/cb-render-html/src/lib.rs::overhead_table	counterexample	2026-08-06	clay-borg
F20	inert	applied	crates/cb-render-html/src/lib.rs::ending_page	counterexample	2026-08-06	clay-borg
F21	degenerate	note	—	—	2026-08-06	clay-borg
F23	inconsistent	applied	decisions/ADR-0017-chaos-window-2-verdict.md	counterexample	2026-08-07	clay-borg
F22	underdetermined	withdrawn	games_ground::edition::supply_tests::play_never_exceeds_the_components_the_box_holds	counterexample	2026-08-07	clay-borg
F26	inert	ruled	crates/cb-render-html/src/lib.rs::a_scoped_table_says_what_a_scope_does	counterexample	2026-08-08	ground-game
F27	unplayed	reported	games/ground/examples/scenario-panel.rs	counterexample	2026-08-08	clay-borg
F29	inconsistent	applied	games_ground::edition::card_text_tests::which_scenarios_are_mechanically_distinct	counterexample	2026-08-08	ground-game
F31	unplayed	reported	games/ground/examples/module-panel.rs	counterexample	2026-08-08	clay-borg
F30	degenerate	ruled	games/ground/examples/scenario-panel.rs	counterexample	2026-08-08	ground-game
F28	underdetermined	applied	games_ground::tests::the_mastery_rating_and_the_shared_score_count_different_things	counterexample	2026-08-08	ground-game

Prose for closed findings (withdrawn or applied) lives in FindingRegister-closed.md; the rows above are the whole register and make design reads them from here.

• F31 — attack_relief.self_soothe_ge4 cannot fire, so it has never been measured. module-panel records zero ATTACKs in every cell, at every seat band, under both a mode-blind and a module-aware bot. The module only acts on an uncancelled ATTACK by a seat at Stress ≥ 4, so it has had no opportunity in any game this repo has run — including every game behind CB-EV-0030's H1 measurement, where it was half the package.

  The mechanism is legible: GreedyPolicy ranks Ground at 100 when the stress gate bites and Attack at 10, so at exactly the position where self-soothe would pay, GROUND wins the comparison. ModuleAwarePolicy raises ATTACK by 55 at the gate and that is still not enough, which is a fact about the ranking rather than about the game.

  unplayed, emphatically not inert. The kernel implements the module correctly; nothing has ever given it an opportunity. Reporting it as a null result is the error the panel now refuses to make: its row shows real numbers with ← never fired: UNMEASURED attached, because the numbers belong to the other modules in the row.

  Sensitivity: vary only the ATTACK rank. The falsifier is a policy that attacks at the gate — if group success then moves under attack_relief.self_soothe_ge4 and not under the baseline, the module is real and only our seats were hiding it. Until such a seat exists, no claim about this module is supported in either direction, and CB-EV-0030's H1 verdict rests on its flat-pressure half alone.

  FALSIFIED, 2026-08-08, and the answer has two parts. GateAttackPolicy — a probe that ranks ATTACK above GROUND at the stress gate, delegating everything else — was built for exactly this.

  (a) The module is unreachable in the printed game, not merely unchosen. With attack_relief.self_soothe_ge4 selected alone the probe still never attacks and the row still reads never fired: peak Stress never exceeds 2, so the gate never bites and no ATTACK is ever gated. The baseline has no inbound Stress (F17's observation from the other side), so this module needs a problem_stress module before it can act at all. A module can be unreachable because of an aspect it does not name.

  (b) Once reachable it is real. Holding the probe fixed and varying only the module — the only comparison that isolates it:

  seats	scoped	scoped_plus_attack_soothe	group wins	DARVO arms
  3p	12	21	+9	299 → 221
  4p	19	28	+9	294 → 213
  6p	50	57	+7	334 → 164

  Under flat pressure the same shape appears in DARVO alone: h1 against problem_stress.flat_any_open arms 200 against 300 / 400 / 600 at 3/4/6p, with group wins 0 in both.

  Stated as the conditional it is: given seats that attack at the gate, self-soothe raises group success by ~7–9 per 100 and cuts DARVO arming by a quarter to a half. The probe is deliberately poor play — it wins 12/100 where greedy wins 60/100 under the same module — so this is not a recommendation to attack, and the row is not evidence against F17.

  What this changes about H1. CB-EV-0030 rejected H1 with policies that never attacked, so H1-B was inert for every game behind that verdict. The rejection stands on flat pressure alone — group wins are 0 with or without self-soothe — but "H1-B does nothing" was never established and is now known to be false.

  Sensitivity: vary only the ATTACK rank. At Attack = 10 (greedy) and at +55 (module-aware) the module never fires; at 110 it fires in every game. Nothing about the module changed — only whether any seat ever gave it a turn.

• F30 — SCN_04 is materially harder at 2 players. 52/100 group success against 67 and 73, with every other parameter held by the edition itself: same deal shape, same 6 available points, same threshold 5, same starting Stress. The lone difference is that SCN_04 is the only 2p deal needing two of one suit (Repair, Clarify, Repair).

  Causation is not established and the finding does not claim it. The falsifier is a deck with a doubled suit that does not lose group success.

  Sensitivity: vary only the seat band and the gap closes — at 4p SCN_04 is 97 against 94/99, so it is not the hard board there. The effect is specific to the 2p deal, which is the band where the deck has fewest cards to route around a suit it cannot match.

  RULED 2026-08-08: acceptable variety for now; SCN_04 stands as a harder board and is monitored. If table feel proves too swingy the first lever is priority-2's suit (dropping the double Repair), not the global thresholds — which also confirms our reading that the suit multiset is the operative difference, without our having measured it.

• F27 — the two competitive modes are scoring lenses over cooperative play. scenario-panel finds group success exactly equal across SHARED GROUND, COMMON PROBLEM and BONDED COALITIONS in all 36 cells. Correct arithmetic: GreedyPolicy maximises the group outcome and never reads state.mode, so the same games are played and only the winner set is carved differently. Whether a mode changes how GROUND is played is therefore untested, and cannot be tested by any panel built from the current policies — it needs one that plays for personal score. unplayed rather than inert: the modes score correctly, they have simply never faced a seat that wanted to win alone.

  Re-measured 2026-08-08 with ObjectivePolicy (CB-WP-0049 T03), which plays its seat's own objective. The finding splits in two:

  • Group success does not move. 34 of 36 cells identical; SCN_03 at 4p goes 99→100 in all three modes, which is the SOLVE-by-value refinement and not a mode effect. Whether the table survives is the same game in all three modes.
  • Who wins does move. Under BONDED COALITIONS at 4p, winning seats per game go 2.04 → 2.98 (SCN_01/02), 2.12 → 3.29 (SCN_03), 2.05 → 3.01 (SCN_04): a seat whose score is its coalition's sum bonds more, and the coalitions get about half again as large. COMMON PROBLEM moves at 6p (1.10 → 1.17).

  So the modes do reach decisions — the original claim that they are "scoring lenses over cooperative play" was too strong, and is withdrawn in favour of the sharper one: GROUND's scoring modes change who wins, not whether the group succeeds.

  Sensitivity: vary only the seat band and the effect appears and disappears — 2p shows no divergence in any mode, 4p shows the largest, 6p shows none under BONDED COALITIONS. A candidate explanation is that two relation slots per seat cap network growth, so at 6p the incentive exists and cannot be acted on; that is untested and is the next thing to vary.

  Still open, because the part that motivated it is untested: no policy models a rival playing their objective, so a competitive mode in which nobody anticipates an opponent remains a weak test of that mode (CB-WP-0049 "Not done here").

  The seat-band explanation is answered, 2026-08-08. ground-game: the two-relation-slot limit is an intentional cap on complexity and play time; a third slot stays an extension rather than core r0. So the 6p flatness is designed, not a defect — and our sensitivity run is welcome but would be measuring an extension, not the shipped game.

  The aspect half is closed, 2026-08-08 (CB-WP-0049 T04). ModuleAwarePolicy reads the resolved Rules, so "a bot that does not attend to a mechanism cannot test claims about it" no longer holds for modules. The mode half stands: no policy models a rival playing their objective, so a competitive mode in which nobody anticipates an opponent remains weakly tested. That is the whole of what keeps this open.

  A second limit surfaced while closing the first: the scope term is inert at round-one positions, because one Problem is face up and a ranking preference needs two candidates. Any measurement of problem_stress.scoped weighted toward early rounds is measuring a mechanism that has not started.

• F26 — a package that adds a FILE is invisible, where a package that adds a column is not. h2-scoped-problem-stress ships Rules_Text.csv — twenty-two passages of player-facing rules, including the one passage that says what a stress_scope does. No reader in clay-borg mentioned the file, so edition-check could not call it stale (a file nothing reads is not out of date, it is unseen) and no gate could call it unread. The rule the table's whole layout encodes sat in the repo, reachable by no player, for the entire measured life of H2.

  Inert rather than underdetermined: nothing is ambiguous about the rule and the engine implements it correctly — the text simply could not fire. Third instance of F18's shape and the first where the unit is a file, which is why it is filed separately instead of folded in.

  Raised against ground-game because the actionable half is theirs: a package's manifest should name its own files, so that a consumer reading none of them is a detectable state. clay-borg now reads the one passage it needed (CB-WP-0046); the other twenty-one remain unread and that is the standing evidence.

  RULED 2026-08-08: accepted as a process ask. Catalog and modules will declare consumed_files / overlays so a consumer reading none of a package's files is a detectable state. Tracked on their side under GROUND-WP-0008.

• F17 — ATTACK cannot affect whether the table succeeds. Reported from play as "there is no incentive to play attacks as long as I have positive cards", and the artifact now exists: games/ground/examples/attack-value.rs, 200 games per cell, one number varied — ATTACK's rank in an otherwise identical policy.

  ATTACK ranked	wins (2/3/4/5/6p)	attacks	DARVO armed
  10 (below all)	132 / 165 / 190 / 200 / 200	0	0
  75 (above SUPPORT)	132 / 165 / 190 / 200 / 200	315–923	13–218
  95 (above SOLVE)	0 / 0 / 0 / 0 / 0	1400–5170	400–1000

  The middle row is the finding: identical win counts at every seat count, while attacking hundreds of times and arming DARVO. Attacking is not punished — it is inert with respect to the goal. Group success is a function of SOLVE alone, and ATTACK only costs anything when it ranks above SOLVE and displaces it.

  So the maintainer was right and the reason is sharper than the phrasing. There is no incentive because there is no path: ATTACK's effects (Stress, Rivalry, DARVO) feed nothing that decides group_success.

  Asked in all three modes 2026-08-07, and ATTACK earns its place in none of them:

  mode	never attack	attack sometimes
  SHARED GROUND	132/165/190/200	identical — free but pointless
  COMMON PROBLEM	59/52/48/44	59/52/48/34 — a cost at six seats
  BONDED COALITIONS	131/134/132/116	59/52/48/34 — roughly halved

  The coalitions row has a mechanism the data confirms. GR-A07 flips a Bond to a Rivalry on Attack, and GR-E04 scores Bond networks — so attacking destroys the thing that scores. And the attacking numbers in E04 are identical to E03's, which is exactly what that predicts: break every Bond and each seat becomes a coalition of one, so GR-E04 degenerates into GR-E03.

  Not a claim that the game is broken. DARVO is the pattern the game is about not falling into; a self-destructive ATTACK may be the design. The question for ground-game is whether the namesake mechanic being unreachable in competent co-op play is intended.

• F18 — the engine imports one of nineteen edition files, so the cards cannot say what they do. Reported as "I don't understand the GROUND card" — which is not a design gap: Actions.csv carries that card's tagline ("Regulate. Restore the frame. Decide.") and full rules text, and clay-borg never imported it. Every other card is the same: a player sees Clarify where the card reads "Ask What Happened — Invite a concrete account before judging." inert: the data exists and cannot fire, because nothing reads it. Ours, and CB-WP-0028 fixes it.

• F25 — the engine hardcodes numbers the edition states. GroundState::threshold is a match returning 5/7/9 while Scenarios.csv carries threshold_2_players, threshold_3_4_players and threshold_5_6_players; setup writes stress: 2 against "All players start at Stress 2."; the engine plays five rounds against a printed 1–5 track. All three agree, on all four scenarios, and that is the finding rather than the reassurance: these are the most contested numbers in the project — the whole 4/6/9 versus 5/7/9 episode turned on them — and the engine has been right by maintenance coincidence rather than by reading the file that owns them. inert, role default: green because they agree, red the moment either side moves. Ours. Same shape as F24.

• F24 — the draw pile is a Rust literal. solution_deck() builds six of each suit from an array and never opens Solutions.csv, whose suit and quantity columns say the same thing. They agree today — 24 rows, six per suit — so nothing is wrong now, and that is the point: the engine is right by maintenance coincidence rather than by reading. inert: the data exists and cannot fire. Role default, not counterexample — the reproduction is green because the two agree, which is the state GameDesign §1.3 says to expect from a documented provisional choice, and it turns red the moment either side moves. Ours. CB-WP-0037 T02 deletes the literal.

• F15 — the rules define one game, not a series. OutcomeView gives personal (per seat), group_success (per table) and winners. Summing the first and counting the third answer different questions, and GROUND says nothing about how several games combine. CB-WP-0024 T04 shows both, labelled, rather than picking one and letting it become the score by default. note: no artifact demonstrates that this harms play — a unit test showing the two tallies point at different seats demonstrates only that they can differ, which is arithmetic, not a design defect. Under GameDesign §3.1 it may not be reported until one exists.