0010: Tick quantum and constant rebaseline

Context

docs/dashboard-build-plan.md §1 prerequisite 4 flags this as blocking and non-cosmetic: the codebase has two clocks with no stated relationship. chronicle/fixtures/whiterun_schedule.py's ScheduleBlock.start_tick/ end_tick (int, chronicle/schedule.py) and chronicle/claims.py's / chronicle/social.py's gamets (float, used throughout for last_rehearsed, first_learned, decay elapsed-time math, and the rumor stage machine) have never been pinned to the same unit. Landing a quantum means deciding what a tick is, in real terms, and then re-deriving every time constant that was previously a dimensionless placeholder.

chronicle/claims.py:57-67 carries four such placeholders, explicitly marked as such:

CONFIDENCE_DECAY_HALF_LIFE = 500.0
VERBATIM_DECAY_HALF_LIFE = 200.0
GIST_DECAY_HALF_LIFE = 2000.0
RUMOR_DORMANT_AFTER = 5000.0

Their docstrings (and docs/decisions/open-questions.md's CHIM fork-threshold note) already say these are "tunable to be set empirically," not derived from any source report. That status doesn't change here — this ADR re-expresses the same placeholders in real units, it doesn't newly derive them from evidence. What changes is that "500" stops being an opaque number and becomes "~3 game-days," which is a number a human (or a scenario author) can sanity-check against intent.

The fixture is explicitly not evidence for the quantum and is out of scope for this change. chronicle/fixtures/whiterun_schedule.py's docstring already disclaims schedule realism — its blocks (0-200, 0-1000, etc.) are scaffolding sized to the tick ranges the existing scenarios/ suite exercises, not a daily routine. At the quantum this ADR adopts, those blocks read as an 8-day court gathering and a 41-day tavern shift — semantically odd, functionally harmless, since the scenarios that use them run fixed tick ranges the fixture already covers. No lane owns rewriting that fixture into a realistic daily routine; it stays as-is.

docs/v0.1-spec.md rules 5/6/16 set the qualitative intent this rebaseline must preserve, not numbers to hit:

• Rule 5: verbatim strength decays faster than gist strength (fuzzy-trace theory) — the source of VERBATIM_DECAY_HALF_LIFE < GIST_DECAY_HALF_LIFE.
• Rule 6: confidence decays with time since last rehearsal — the source of CONFIDENCE_DECAY_HALF_LIFE existing at all.
• Rule 16: the rumor stage machine unheard -> heard -> repeated -> dormant -> forgotten — the source of RUMOR_DORMANT_AFTER and RUMOR_FORGOTTEN_GIST_THRESHOLD.

docs/scenario-ladder.md (not yet implemented as pytest tests, but the forcing tiers this build works toward) supplies concrete anchors the new values should not contradict:

• T0.2 Decay — 30 quiet game-days, asserts stage is still heard, "not dormant yet."
• T2.1 Spread — 10 game-days, cast ≈ 25 (this is the run docs/ui-spec.md sizes its 10⁵-10⁶ trace-row estimate against).
• T2.5 Dormancy and reactivation — 90 quiet game-days after spread; the rumor must have migrated to dormant well before this window ends, so a retelling has something to reactivate.

Decision

One tick = one game-hour. Equivalently: tick (int, schedule.py) and gamets (float, claims.py/social.py) are the same clock, same unit — 1 tick = 1 gamets = 1 game-hour. This is the load-bearing part of this ADR: Lane 2's frame-log timestamps and Lane 4's driver tick loop advance one clock, not two that happen to correlate. A gamets value of 1050.0 and a tick value of 1050 refer to the same instant.

Derived cadence: 24 ticks = 1 game-day (matches the keyframe cadence docs/ui-spec.md already assumes: "K default one game-day"). 240 ticks = 10 game-days.

Rebaselined constants (chronicle/claims.py)

All four are still placeholders in the same "tunable, not derived" sense the original comments stated — this rebaseline changes their units, not their epistemic status. Each is chosen to preserve rule 5/6/16's qualitative shape and to sit sensibly against the T0.2/T2.5 anchors above, not to hit a specific research-derived number (none exists).

Constant	Old (unitless)	New (ticks)	Derivation
VERBATIM_DECAY_HALF_LIFE	200.0	72.0	~3 game-days (24×3). Exact wording should fade within a few days — fastest of the three, preserving rule 5's ordering.
CONFIDENCE_DECAY_HALF_LIFE	500.0	168.0	~7 game-days (24×7, one week). A belief nobody has rehearsed in a week has lost roughly half its confidence — rule 6, positioned between verbatim and gist per fuzzy-trace theory.
GIST_DECAY_HALF_LIFE	2000.0	1440.0	~60 game-days (24×60). Gist ("something bad happened to the Jarl") should outlive both verbatim and confidence by a wide margin — rule 5's central claim — while still eventually crossing RUMOR_FORGOTTEN_GIST_THRESHOLD over a multi-month horizon rather than never.
RUMOR_DORMANT_AFTER	5000.0	1080.0	~45 quiet game-days (24×45). Chosen to sit strictly between the T0.2 anchor (30 quiet days = 720 ticks, must still read as "not dormant yet") and the T2.5 anchor (90 quiet days = 2160 ticks, must have reached "dormant" well before the window closes so a later retelling has something to reactivate). The old value, 5000 ticks ≈ 208 days at this quantum, was answering the build plan's flagged question ("is that the intended dormancy horizon?") with "no" — 208 quiet days is far outside any scenario-ladder window this system is meant to exercise.

RUMOR_FORGOTTEN_GIST_THRESHOLD = 0.05 is a dimensionless strength floor, not a time constant — unchanged. RETELL_CONFIDENCE_DECAY, RETELL_VERBATIM_DECAY, RETELL_GIST_DECAY, and WITNESS_CONFIDENCE are per-event multiplicative factors (applied once per retelling, not against elapsed time) — also unchanged; they carry no tick unit to rebaseline.

Sanity check against the new values: at the T2.5 anchor (90 quiet days = 2160 ticks after spread), a belief with GIST_DECAY_HALF_LIFE = 1440 has decayed to 0.5^(2160/1440) ≈ 0.35 of its starting gist strength — comfortably above RUMOR_FORGOTTEN_GIST_THRESHOLD = 0.05 for any reasonable starting strength, so the scenario can exercise "dormant" and "forgotten" as genuinely distinct outcomes rather than always co-occurring.

Rule 16 ordering check: stage_at() checks "forgotten" before "dormant" (a belief that never got rehearsed can skip straight to forgotten even without a long quiet gap), so rule 16's stated sequence ... -> dormant -> forgotten is a typical path, not a hard invariant — but the constants should make the typical path typical. Solving for the starting gist_strength at which a belief crosses RUMOR_FORGOTTEN_GIST_THRESHOLD at the same elapsed time RUMOR_DORMANT_AFTER triggers: the old constants (2000/5000) cross over at starting gist ≈ 0.28 — over a quarter of beliefs by starting strength would skip "dormant" entirely. The new constants (1440/1080) cross over at starting gist ≈ 0.084 — only unusually weak beliefs skip "dormant," so the rebaseline incidentally makes the documented rule-16 sequence the common case rather than close to a coin-flip, on top of fitting the T0.2/T2.5 anchors.

No other chronicle/ module has time-denominated constants. Verified with grep -rn '^[A-Z][A-Z0-9_]* *[:=]' chronicle/, which catches every module-level constant assignment in the package (not just the narrower ^[A-Z_]* = pattern, which would miss a name containing a digit): schedule.py (ENCOUNTER_PROBABILITY = 0.5, dimensionless), social.py (GRUDGE_EMOTIONAL_WEIGHT, GRUDGE_EVIDENTIARY_WEIGHT, REPUTATION_PRIOR_ALPHA/BETA, REPUTATION_WEIGHT_BY_KIND, all dimensionless weights/priors — no grudge-cooling clock exists yet; that's Tier 3, out of scope here), claims.py's own RETELL_*/ WITNESS_CONFIDENCE (dimensionless, covered above), and no constants at all in propagate.py, events.py, or fixtures/.

Consequences

• Ticks-per-10-game-days = 240. This is the figure docs/ui-spec.md's 10⁵-10⁶ trace-row estimate for a 25-NPC, 10-game-day Tier-2 run should be checked against (Lane 2/M0's concern, not resolved here — just supplied).
• Any code or docs that referenced the old constants' magnitudes directly (none found outside chronicle/tests/test_claims.py, which imports the named constants symbolically rather than hardcoding their values) is unaffected.
• chronicle/fixtures/whiterun_schedule.py is unchanged and stays semantically odd (8-day court scene, 41-day tavern shift) under this quantum — a known, accepted consequence, not a defect this ADR should fix. A realistic daily-routine fixture is future work, unowned by any current lane.
• Future scenario-ladder tests (T0.2, T2.1, T2.5) should use game-day-based tick counts (24 * N) rather than bare tick literals, so their intent reads directly off the assertion.

Related

docs/dashboard-build-plan.md §1 prerequisite 4 (the forcing requirement). docs/v0.1-spec.md rules 5, 6, 16. docs/scenario-ladder.md T0.2/T2.1/T2.5. docs/decisions/open-questions.md's "tunable to be set empirically" note, which this ADR's constants inherit rather than override.