Skip to content

Canonical refused fits, PhaseSkill option memory, hypothesis prompts - #162

Merged
yichao-liang merged 4 commits into
masterfrom
bl-merge-11-prompts-fit
Sep 1, 2026
Merged

Canonical refused fits, PhaseSkill option memory, hypothesis prompts#162
yichao-liang merged 4 commits into
masterfrom
bl-merge-11-prompts-fit

Conversation

@yichao-liang

Copy link
Copy Markdown
Collaborator

A refused sysID fit is still the file's canonical fit and reused physics fits restore the margin gate, every PhaseSkill execution starts with clean option memory, dangling params[...] references name the missing parameter, and prompts ship goal-required mechanisms as labelled hypotheses with every cycle attempting the goal and running the top open question.

Commits:

  • params view: a dangling params[...] reference names the missing parameter
  • sim.fit: a refused sysID fit is still the file's canonical fit; reused physics fits restore the margin gate
  • skills: a PhaseSkill starts every execution with clean option memory
  • prompts: goal-required mechanisms ship as labelled hypotheses; every cycle attempts the goal and runs the top open question

Part of the bridge-learning merge stack.

🤖 Generated with Claude Code

https://claude.ai/code/session_01VQkk8ycwwwiSKraDw88FVm

@yichao-liang
yichao-liang force-pushed the bl-merge-11-prompts-fit branch from 44fef6b to acbbf0e Compare September 1, 2026 13:27
@yichao-liang
yichao-liang force-pushed the bl-merge-10-agent-tools branch 2 times, most recently from 5ba3975 to ee86a9e Compare September 1, 2026 13:56
@yichao-liang
yichao-liang force-pushed the bl-merge-11-prompts-fit branch 2 times, most recently from 1eef3cd to 8e97f58 Compare September 1, 2026 14:21
@yichao-liang
yichao-liang force-pushed the bl-merge-10-agent-tools branch from ee86a9e to 621fd8b Compare September 1, 2026 14:21
@yichao-liang
yichao-liang force-pushed the bl-merge-11-prompts-fit branch from 8e97f58 to 191662d Compare September 1, 2026 14:43
@yichao-liang
yichao-liang force-pushed the bl-merge-10-agent-tools branch from 592db38 to 386d6af Compare September 1, 2026 15:01
@yichao-liang
yichao-liang force-pushed the bl-merge-11-prompts-fit branch from 191662d to f2c481d Compare September 1, 2026 15:01
@yichao-liang
yichao-liang force-pushed the bl-merge-10-agent-tools branch from 386d6af to d073090 Compare September 1, 2026 15:26
@yichao-liang
yichao-liang force-pushed the bl-merge-11-prompts-fit branch from f2c481d to 0d929d2 Compare September 1, 2026 15:26
@yichao-liang
yichao-liang force-pushed the bl-merge-10-agent-tools branch from d073090 to e81c5f7 Compare September 1, 2026 15:42
@yichao-liang
yichao-liang force-pushed the bl-merge-11-prompts-fit branch from 0d929d2 to 80cae96 Compare September 1, 2026 15:42
@yichao-liang
yichao-liang changed the base branch from bl-merge-10-agent-tools to master September 1, 2026 15:55
…eter

The invented-predicate parameter view raised "accessed before any
parameter fit; call sim.fit()" for any missing key. In run_20260829
cycle 1 an agent deleted ParamSpec("butt_gap_tol") while a predicate
still read it, and the advice sent it to sim.fit() several times for
a name that no longer existed. The error now names the missing key,
lists the parameters that are available, and says what fixes it:
declare the ParamSpec or update the classifier (values refresh from
the current file at the next sim call; sim.fit() fits them).

Claude-Session: https://claude.ai/code/session_01Pbv9TvedRD2iKyoKMJKx6H
…d physics fits restore the margin gate

When the rollout-sysID trimming guard found no explainable trajectory,
sim.fit returned its report before publishing, so the probe kept
saying UNFITTED after the agent did fit and the deploy step re-ran the
same refused fit under a misleading FIT FALLBACK warning (both
sim_predicator runs, cycles 1-2). The refusal is now published as the
canonical fit of that file, pinned at the declared inits with the
planner's physical params left untouched, exactly as the harness
fallback would have ended.

Deploying a published physics fit also re-applies its physical values
and restores the cycle-applied snapshot and the physics-margin sigma
points (applying resets them; only the harness joint fit used to
rebuild them), so skipping the harness refit no longer empties the
capture gate's margin sweep. sim.fit computes the sigma points from
its own identifiability report and publishes them alongside the fit.

Claude-Session: https://claude.ai/code/session_01Pbv9TvedRD2iKyoKMJKx6H
_initiable reset only phase_idx, so a grounded option re-executed by
the explorer's certified-plan replay (or a suffix replan) ran its
phases from the previous execution's state: cached BiRRT waypoints
from a different start pose, a Place aim offset for a landing that
never happened, exhausted retries. In the 2026-08-29 bridge runs 10 of
11 replays of a plan that had just solved the task failed this way,
which is what kept the runs from early-stopping. Reproduced in a fresh
process: fresh grounding gives the solved episode's step counts, the
same objects give the failed replay's; with the reset both agree.

Claude-Session: https://claude.ai/code/session_01Pbv9TvedRD2iKyoKMJKx6H
…cycle attempts the goal and runs the top open question

In the 2026-08-29 bridge runs no cycle-0 episode consummated a bond
(glue on faces that touched nothing, spans placed near but not
against each other), so the learners had no event to model. The
plan-arm learners then followed the learn prompt's "never ship a
never-observed mechanism" rule, the belief had no bond, and the test
session proved the goal unreachable and submitted a partial plan.

Learn prompt: a hypothesis the goal can do without is still recorded
and not shipped; a mechanism the goal REQUIRES ships as a labelled
hypothesis with declared ParamSpecs and its confirming experiment as
the ledger's first entry, and a GO may rest on it if it says so.

Explore prompt: a cycle's data must contain at least one full-goal
attempt and the top open question's experiment as specified; the
first-cycle coverage rule now demands each named interaction be
carried to its consequence (contact, release, wait, probe), and the
explorer's ledger guidance makes the top entry mandatory.

Claude-Session: https://claude.ai/code/session_01Pbv9TvedRD2iKyoKMJKx6H
@yichao-liang
yichao-liang force-pushed the bl-merge-11-prompts-fit branch from 80cae96 to e02eef2 Compare September 1, 2026 15:55
@yichao-liang
yichao-liang merged commit 67868fc into master Sep 1, 2026
14 checks passed
@yichao-liang
yichao-liang deleted the bl-merge-11-prompts-fit branch September 1, 2026 16:42
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant