Repository navigation
RFC 0002 wave 0: freeze (RFC draft, bench/body fixtures, seeds and bars) - #37
Merged
Merged
Conversation
…asurement docs/rfcs/0002-neuromodulators-and-body.md: the RFC 0002 draft (needs and the omission let-down, surprise, pace, dispositions, valence; part 2 rest and load) in RFC 0001's shape, references by author and year. bench/body/: the benchmark that defines "done" for it, frozen: 9 fixture personas and 13 derived twins, 11 seeded streams and their generator (fixtures/gen.py --verify regenerates them byte for byte), the bars (README tables, bars.md, bars.json), run.py and the per-part scripts, and fixtures/MANIFEST (50 sha256 entries, manifest_check.py). arith_0002.py: 40 checks, 0 mismatches. No engine has the blocks yet: every part reports "setup error: this engine does not support <block>" and run.py exits 2. No engine change; every bar unmeasured by design.
An independent review graded the 40 bench/body bars (12 sound, 10 weak, 18 vacuous) by naming a degenerate policy or null that passes each. No engine with an RFC 0002 block has produced a number for any of them, so they are repaired under the freeze rule (a bar is never edited after a number is seen). No pass value relaxed. - Two-sided correctness and coverage: exact wins on a 15-cell needs1 grid (n1, n2); success-weighted order, odds magnitudes and released goals on a second probe (n3); contact ceiling also while the principal is away and every answered report paying (n8); non-positive wins, coverage and a deficit bound (n9); a positive baseline and recomputed deficits (n10); contact before the absence and after the return (n11); learning that must happen (s2) and an exact opening step (s3); E trajectories exact at every event (o1, o2); residual at lags 1-5 (v3); exact contraction plus a below-threshold control (r0); the lesson kept after the gap (r1); a non-saturated lineage probe with a 20 % reduction (r2); conflicts that bind (r4); each load channel's amplitude and time course (l1-l3). - Nulls that fail inside their own fixture: the env control through the refused sources' transports (n7), an admissible burst above the pace cap (p6), constant-gain engine twins at the gate's movement budget (s1), planted corruptions for the replay bars, habit-free twins gated. - Effects on decisions: released odds for v1/v2, changed pursue goals for o3, P(careful) for l1. - nulls.py: a competent plain-Python null for every bar, judged by the parts' own gate functions; run.py stops with exit 2 when one passes. 51 null rows over 40 bars, 0 pass. The RFC's own rules run as non-gating references: they pass every bar they cover except n10 and n11 (kept as frozen; flagged). - New streams return, bursts, burst, two-point-gaps; variants needs2-s02, gate-g1.25..3.0, load-effects; every v1 stream and variant regenerates byte for byte. bars.json/bars.md at version 2; v1 kept verbatim as bars-v1.json/bars-v1.md; MANIFEST regenerated (64 entries). arith_0002.py: 66 checks, 0 mismatches. - run.py on an engine without the blocks still exits 2, one setup line per part.
At 1 h the deficit is 0.1 and the amount 0.2 overshoots it back to 0.1, so a correct engine shows no change and would fail n7. Found by a dry run of needs.py against a Python model of RFC 0002 A.3-A.4 (not the engine), before any engine measurement; pass value and bar text unchanged; logged under 'Changes after the freeze'.
…, v2's null within rounding Found by dry runs of every part against Python models of RFC 0002 (not the engine), before any engine measurement: - r1 is measured on the reversal draws without the lock turns: with them the stand-in learner re-learns in 6 turns after rest instead of 5 (6/11 = 0.545 against 0.5) because r4's lock habit shares the stream (rule 2); without them 5/11 = 0.4545, the number the bar's reachability check cites. The locked ratio is reported. - s3 checks the cap where the RFC puts it, on the step before recentring: the recentred changes come from a step within +-g_max x step_cap exactly when their range is <= 2 x g_max x step_cap (v1's script compared each recentred change with the cap, which a correct step exceeds in 968 of 1,800 lives of the dry run). - v2's no-block twin moves by rounding residue (1e-6), so its move is gated at <= 2e-6, not == 0. Pass values and bar text unchanged; logged under 'Changes after the freeze'.
…d events, v1's and v2's mood odds) Verdicts unchanged (51 null rows, 0 pass a v2 bar); the rows printed illustrative numbers where the null's outcome follows from its definition, and now print the computed ones.
BitmapAsset
marked this pull request as ready for review
October 17, 2026 06:44
Merged
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
RFC 0002 wave 0: the freeze. The RFC draft and the benchmark that defines "done" for it, committed before any engine has
the blocks.
The two rules (bench/body/README.md):
fixtures/MANIFESTin one commit before any measurement on the engine. A bar is never edited after a number is seen. Abar that fails is a finding: the design is fixed or the part is reverted, and the failed result stays in the history.
about that opposition.
What is frozen
docs/rfcs/0002-neuromodulators-and-body.md: the draft in RFC 0001's shape (Status: draft for review), references byauthor and year.
bench/body/: 9 fixture personas and 13 derived twins (fixtures/variants/), 11 seeded streams (fixtures/streams/,regenerated byte for byte by
python3 bench/body/fixtures/gen.py --verify), 40 bars (the README's tables, frozen asfixtures/bars.mdsha2560b2e95f8…; their pass values infixtures/bars.json),run.pyand the per-part scripts, andfixtures/MANIFEST(50 entries;python3 bench/body/manifest_check.py).CHANGELOG.md(0.10.0 - unreleased) andBENCHMARKS.md§12 (frozen, not yet run).No engine change; every bar unmeasured by design. No engine has the blocks yet: on today's binaries
persona lintrefuses each fixture's block as an unknown field, every part prints
setup error: this engine does not support <block>,and
run.pyexits 2 (checked with probbit 0.8.2: 10 parts reported, exit 2).Checks
python3 bench/body/arith_0002.py: 40 checks, 0 mismatches (the closed forms behind the bars).gen.py --verify: 24 files regenerated, 0 differ;py_compileon every script.Where the spec left a choice before the engine exists (scripted hosts that read the stance, a second unpaid goal in the
one-goal fixtures, field names), the README's last section records it. One finding while freezing: the persona reader
refuses a bare YAML key
n, so the fixtures writedrive: {"n": 4, "m": 3}; the RFC's schema example needs the samequoting, or the build renames the key.