Posterior-Credible Monotone Floor¶
TransitionShield supports shield_opponent_mode=credible_floor as an
experimental alternative to the default MAP-based monotone_floor. It uses
the opponent-model posterior to discard only a controlled lower tail, while
still enforcing safety robustly over every retained level.
Floor rule¶
Let alpha_t(l) be the normalized posterior over ordered opponent levels
l = 0, ..., M - 1. At observation count t, the candidate floor is the
largest z satisfying
The effective floor is monotone: it is at least the previous floor. Runtime
action risks and successor-budget constraints take the worst case over the
whole retained suffix {z, ..., M - 1}. The posterior is used only to select
the suffix and to rank rewards; safety is not averaged inside that suffix.
The existing patience gate also applies. It can delay a candidate raise and
therefore retain extra lower levels, which is conservative. Runtime inspection
uses that patience-filtered floor exactly.
Fixed and summable schedules¶
shield_credible_delta is the posterior-tail error budget. The default is
0.05. shield_credible_schedule controls how it is allocated:
fixedusesdelta_t = deltaat every decision. This is a pointwise posterior-credible suffix. Reusing a fixed 5% allowance indefinitely does not itself give a 5% infinite-horizon error bound.summableusesdelta_t = 6 * delta / (pi^2 * (t + 1)^2). Since the allocations sum todelta, a union bound limits the probability of ever excluding the true level to at mostdelta, subject to the modelling assumptions below.
The observation count belongs to the shield instance and survives episode
resets, matching the intended continuing-learning interpretation. Per-step
info reports the allowance used for that decision. After the opponent update,
the wrapper's credible_delta_t property can already reflect the next
observation-indexed allowance.
Safety accounting and assumptions¶
The credible-floor claim is conditional on all of the following:
mix_alphais a calibrated posterior over the relevant ordered level hypotheses;- the true opponent path respects the monotone-level assumption, so an excluded lower level cannot later become the true active level;
- the learned world model, opponent policies, graph coverage, unsafe labels, and numerical reachability bounds satisfy the ordinary shield assumptions.
Under those assumptions, the summable schedule supports the union-bound accounting
Notebook factories interpret max_risk as the common total risk target. For
credible_floor, they reserve the tail error automatically and initialize the
reachability budget at max_risk - shield_credible_delta; monotone and other
opponent modes receive the full max_risk. This makes a nominal 0.20
OmSh-Credible run with delta=0.05 enforce a conditional model budget of
0.15 and report an additive nominal quantity of 0.20. Direct construction
of TransitionShield remains lower-level: its init_safety_bound is the
conditional reachability component and its
nominal_posterior_tail_risk_quantity adds the tail allowance only for the
summable schedule. A credible delta larger than the total notebook budget is
rejected.
This is not a proof that the learned posterior is calibrated. It is also not a
safe response to arbitrary non-monotonic switching: without a restriction on
switching, all levels remain possible and all_levels is the appropriate
robust mode.
Posterior reversals and telemetry¶
Because the floor is irreversible, a later posterior can place more than
delta_t mass below it. The implementation does not lower the floor. Instead,
it exposes the mismatch so experiments can test the assumption:
- decision
info: total and pointwise delta, schedule, excluded and retained mass, posterior-tail constraint status, observation count, and the nominal additive risk quantity; - traces: the same credible-floor fields;
EpisodeLog: credible steps, mean pointwise delta, mean/max excluded mass, mean retained mass, and posterior-tail constraint-miss count/rate.
shield_posterior_tail_constraint_held=false means the current posterior no longer
assigns the promised mass to the already-retained suffix. It is an empirical
warning about the irreversible monotonic assumption, not a repair or a new
safety certificate.
Experiment configuration¶
The standard and parameterized notebook factories accept the same settings:
--shield-opponent-mode credible_floor
--shield-credible-delta 0.05
--shield-credible-schedule summable
The launcher exposes the two primary conditions as distinct algorithms:
omsh_monotone selects monotone_floor, while omsh_credible selects
credible_floor. Their published labels are OmSh-Monotone and
OmSh-Credible, and their histories remain collision-free. Credible-floor
runs remain conditional on posterior calibration and monotonicity diagnostics.
Reward-frontier interpretation and primary-suite status¶
For the current M = 3 experiments, every summable allowance satisfies
delta_t <= 0.05 < 1 / M. At a fixed posterior, the credible-tail candidate
floor cannot exceed the MAP candidate floor: excluding the MAP level would
discard at least 1 / M posterior mass. The one-level patience operator
preserves this ordering on a shared posterior history. Whenever the credible
rule requests a raise from a common floor, the MAP rule requests the same next
one-level raise, so credible evidence cannot accumulate while MAP evidence
resets. By induction, the MAP floor is never below the credible floor when
both see the same posterior sequence.
The primary comparison also matches the user-facing total risk target. A
credible run reserves delta = 0.05 from that target, whereas monotone uses the
full target as its model-controlled reachability budget. Consequently,
OmSh-Credible has no intrinsic feasible-set or optimal-reward advantage under
matched histories: it retains at least the MAP suffix and also has the smaller
model-controlled budget. Separate learning runs can still produce different
state and posterior histories after their actions diverge, and the extra
conservatism can act as an optimisation regulariser. A higher credible return
is therefore possible as a finite-training empirical outcome, but it does not
follow from the credible-set construction and cannot improve the best
attainable reward frontier relative to a monotone controller that is free to
imitate the same actions.
For that reason, Credible is retained as an implemented diagnostic/ablation
but retired from the primary launcher and final comparison. The default suite
contains IPPO, IPPO-Lagrangian, CPO, and OmSh-Monotone. An explicit
algorithms=omsh_credible request remains supported so the posterior-tail
assumptions and finite-training regularisation effect can still be studied.