Skip to content

SameWrite 1.2.0

Choose a tag to compare

View all tags
@ipeterpetrus ipeterpetrus released this 14 Sep 21:00
· 41 commits to main since this release
v1.2.0
14eb217

Overall token-cost improvement: NOT_PROVEN.

That is the headline, and it is deliberate. 1.2.0 is released for the evidence layer, not for a
savings number. The observed effect was −1.7%, cheaper on 8 of 10 fixtures (p = 0.109) — and the
rig's own measured null-vs-null variance is larger than that, so no percentage saving is claimed
anywhere in this release.

What is measured and does hold:
default model-context overhead added by the observer 0 bytes — skills/ and hooks/ are byte-identical to v1.1.0, verified by sha256 per file
Hermes Agent skill support VERIFIED — 0.21.3 (1ad89ac), isolated profile, 11/11; 80 B listing, 0 B unused body
AI-VOS profile evidence-aligned — rewritten after reading the canonical governance material
quality non-inferior on correctness, safety, evidence completeness and authority compliance across 80 runs

What is new

The skill is byte-for-byte what 1.1.0 shipped. Everything here is the measurement layer learning to
survive a population of agents that never stops.

  • Scope isolation. --scope-id and --workload-class keep populations apart; records from
    different scopes are counted and named, never averaged into a population that describes nobody.
  • Evidence quality, failing closed. A sweep labels itself COMPLETE / PARTIAL / INVALID /
    EMPTY; a partial sweep cannot produce a candidate unless you say the bounded corpus was the
    target.
  • Opaque run identity. A random run_id is the only dedup key, so two agents with identical
    numbers stay two observations and a re-copied file stays one. History merges across hosts, in any
    order, twice, safely.
  • Time is never trusted. Record order is classified ok / reversed / ambiguous; skew yields
    no direction at all.
  • Safe to schedule. Deterministic candidate ids, a non-blocking exclusive lock, fsync +
    os.replace. A 30-day, 9,000-invocation simulation emits zero new proposals once the evidence
    stops moving. No daemon; nothing is installed.
  • Concurrency and crash safety. 32 and 64 parallel writers with every line intact; a torn tail
    no longer destroys the record appended after it.
  • Machine contract. Exit 0 for any valid run including NO_ACTION; --strict-exit maps status
    to code; output_schema_version and threshold_schema_version move independently.

Two candidates the optimizer proposed, and neither shipped

  • bash-output-shaping — REJECTED. Bash is 62% of carry, so bounding it looked like the best
    change available. Measuring 30,731 real Bash results ended it: median 449 B, p90 2,246 B, maximum
    28,989 B, none above 30 kB. The host already caps Bash output and spills the rest to a file.
    Bash dominates carry by accumulation across thousands of small results — a different problem.
  • listing-prune — DEFERRED. 70 of 82 listing entries never invoked, 24,002 B, but 2.30% of
    carry     — below the threshold fixed before the numbers were read, and those cold entries belong to
    other people's plugins, not to SameWrite.

The measurement finding worth more than the cost number

Because the skill file is identical at 1.1.0 and 1.2.0, two arms of the experiment show the model
the same bytes — which turns them into a ruler for the rig's own noise. Three such null-vs-null
comparisons read −5.6%, +15.1% and −6.4%. The middle one reached p = 0.021 between
treatments that cannot differ, and did not replicate.

Two methodological defects were found and fixed along the way: arm was perfectly confounded with
wall-clock time (jobs submitted arm-major drain in blocks), and the second-channel rig scored a
provider usage limit as ten model failures. Both are the kind that manufacture results.

Still untested, and named as such

  • GPT-5.6 Sol portability — UNTESTED. The attempt returned 10/10 infrastructure errors on an
    account usage limit. An untested channel is never estimated from a tested one.
  • Hermes observer — UNTESTED. It runs there, but has no reader for Hermes session files, so it
    measures nothing. Hermes skill support is VERIFIED; the two claims are kept apart on purpose.

Install

/plugin marketplace add ipeterpetrus/samewrite
/plugin install samewrite@samewrite

Upgrading from 1.0.0 or 1.1.0 is in-place and idempotent; the guard's 1.0.0 entry form is
recognised, so no duplicate is created.

Full notes: docs/RELEASE_NOTES_1.2.0.md ·
experiment: experiments/aivos/RESULTS.md ·
candidate verdicts: docs/CANDIDATES.md

Assets 2
Loading