Initial observation: holdout line 282
Machine: 1RB0LD_1RC0RD_1LA0RC_1RE1LB_1LF0LA_---0LE. Byte-exact string read from the current 740-machine list, line 282. Root claimed this directory for Astra and supplied the preflight in astra/bouncer-triage/preflight.json; no existing proof or open-PR conjugacy was found there. The saved community class is “Bouncer + Chaotic Edge”.
Result and stopping point
No halt through 1,000,000,000 direct steps from blank tape. This is a finite observation, not a theorem or a kill. The mechanism confidence is low. No complete macro return map, invariant, or Coq proof has been proposed. Possible chaotic-edge behavior is flagged; no pseudo-random arithmetic map was extracted, so this is not a confirmed cryptid classification.
The visited span expands mainly left. The right record moves from 154 at one million steps to 228 at one billion. Long left/right sweeps rewrite a word with many short blocks; a bounded list of counter exponents has not been identified.
Relative priority within this pair: 1/2. More promising of these two for a bounded word-transducer investigation: the sampled 12-bit-factor count stabilizes at 102, and three fixed local windows cover 1,024 of the first 1,282 left-record gaps. The long return gaps and the changing whole word are still unexplained.
Stop here for a priority decision before deeper rule/invariant work. The useful next question is whether the entire evolving word has a closed finite transducer description and whether that description protects F0. Longer simulation alone will not answer it.
Concrete measurements
| Direct steps | Visited interval | Ones | Distinct length-12 factors | K=12 DP text characters |
|---|---|---|---|---|
| 1,000,000 | [-893, 154] | 795 | 90 | 617 |
| 10,000,000 | [-2918, 192] | 2,052 | 101 | 1,912 |
| 100,000,000 | [-9313, 217] | 6,614 | 102 | 5,749 |
| 1,000,000,000 | [-29278, 228] | 17,521 | 102 | 17,304 |
The factor count is measured in the whole raw visited tape at the stated instant; it is not an entropy/randomness claim. Compression uses the existing dynamic-programming compressor with maximum period 12 and nested folding disabled. It is a finite diagnostic, not evidence that no better representation exists. Full compressed strings and 1-run histograms are in work/finite-diagnostics.json.
The bounded accelerated run stopped at 30,000 simulator loops, corresponding to 324,434 exact base steps. It found 10 rules and made 10 rule moves; it did not establish nonhalting. Its 30-second cap was not reached. Full output is work/qsim-30s.jsonl.
An independently implemented sparse-tape simulator agrees byte for byte with the C simulator at step 1,000,000, including head, state and visited bounds. All 78,075 observed F entries there read 1. There are 71 distinct 13-cell neighborhoods at those entries. E0 writes 1 and moves left into F; the cell reached must be 1 to avoid F0. No proof that this safety condition is preserved has been found.
Exact finite local windows
Head coordinates are relative to the initial head at 0. The input/output words occupy the same listed window. Each run was independently executed transition by transition; every read stays inside that window until the final configuration, so exterior cells are irrelevant to the finite run. These are local finite calculations, not a global invariant or formal Coq lemmas.
| Initial state | Window offsets | Before | Steps | Final head/state | After | Matching observed record gaps |
|---|---|---|---|---|---|---|---|
| A | 0..2 | 000 | 5 | B at -1 | 101 | 513 |
| B | 0..4 | 01000 | 11 | A at -1 | 00101 | 256 |
| B | 0..12 | 0101101100100 | 81 | A at -1 | 0010110110101 | 255 |
The three windows account for 1,024 of 1,282 consecutive left-record gaps in the first 2,000,000 steps. The remaining 258 gaps include the long returns and are not covered by these rules.
work/local-window-runs.json includes every transition of each local run. These windows do not close the family of whole-tape configurations; the unknown return map is the main gap.
Artifacts and reproduction
index.html: standard machine page with observation, ranking, finite-window rules, and labeled space-time diagram.img/space-time-labeled.svg: axes and sampling labels around the existing renderer's PNG; tape coordinate is absolute, time increases downwards. White cells are 1; dark cells are 0 or unvisited; colored head cells show state. This is a sampled diagram, not every transition.work/direct-1e8.jsonlandwork/direct-1e9.jsonl: exact C snapshots every 1,000,000 and 10,000,000 steps, respectively, plus the final snapshot.work/records-2e6.jsonl: every new span record through 2,000,000 steps; the 2,500-event cap was not reached.observe.json,status.json: machine-readable outcome and stopping condition.
Reproduce the initial runs with python3 -B astra/observe_bouncers_282_465.py, diagnostics with python3 -B astra/analyze_bouncers_282_465.py, and packaging/window checks with python3 -B astra/finalize_bouncers_282_465.py.
The extra direct run is python3 -B tools/sim.py '1RB0LD_1RC0RD_1LA0RC_1RE1LB_1LF0LA_---0LE' --engine c --max-steps 1e9 --every 1e7 --snapshot --end-snapshot. The acceleration parameters were --max-loops 30000 --max-seconds 30 --every 1000 --bf-loops 3000. The diagram uses tools/diagram.py with --mode every:1000 --steps 2e6 --style explore --frame abs.
Runs used nice -n 10, ulimit -v 6000000, process timeouts, and one worker at a time. At start MemAvailable was 19,330,164 kB and 1-minute load was 6.31, within the shared guards; guards were checked again before the later simulation batches. No shared tools/vendor files were changed, and no Coq or formal gate was run.
Priority decision — 2026-10-14
The user said: “Follow your recommendations for the 5.” Recorded study for this family; a mechanism study is funded for at most 30 minutes.
Funded mechanism study completed — 2026-10-14
The approved study is complete. See study.md and study-20261014/result.json. Two exact parameterized left returns were derived: R1 takes 20n+41 steps for n≥1; R2 takes 12n+69 for n≥0. Their ten finite components and 263 parameter samples were checked; root also independently replayed both families. The exact B/E word model reproduces every saved C left record through two million steps, including the previously unexplained long gaps.
The two families plus the prior short rules cover 1,027/1,282 gaps. The remaining gaps require repeated word rewriting. A fitted whole-word language fails at its endpoint, and adding a length residue modulo6 does not fix it. No closed global invariant, Coq proof or kill was produced. The stage stops here with overall mechanism confidence low. Any follow-up should address the word/endpoint relation, not merely extend simulation. Do not rerun the old initial-observation finalizer over these funded-stage results.
Funded stage closed — 2026-10-14
The approved mechanism study is complete. Its answered decision is archived in decisions/2026-10-14-study.json. No research job remains running. The new priority card concerns a separate stage and is unanswered. Full comparison: coord/audits/bouncer-studies-20261014.md.
Further research deferred — 2026-10-14
The user selected defer on the follow-up priority card. The answer was relayed in coord/mailbox/to-astra/20261014-154300-claude-dennis-answers-6-282.md. The completed study and evidence are retained; no additional research is scheduled. The answered card remains in decision.json.