LOOP 2026-10-10: verdict reuse to 100% -- ledger, weak spots, plan (epic #8095)

The owner started an autonomous improvement loop on 2026-10-10 (every 15 min, this session). It works on epic #8095: rebuild only what is not yet verified, all in .t27. Every tick reads this issue first, repairs anything red that a previous tick left behind, and only then starts new work. Every tick leaves one comment: what it did, PRs, anomalies, next.

Where we are (measured)

Remote silicon run: 2.5 h (die C, USB/IP) -> 42 s (bench agent, die A, verified: an agent next to each board -- bench_agent.t27 and t27c bench-agent (Closes #8153) #8155).
Repeat bitstream build: 39 s -> 4 s with bitstream reuse (verified: t27c silicon reuses a bitstream built from the same inputs (Closes #8290) #8291 spec; t27c silicon reuses its bitstream; a ported script pays for the glue (Closes #8295) #8296 glue, open).
frontier on master: 382 of 1739 reused (22%). With specs/base/types.t27: compile again at 4 trits per byte, seal v2 (Closes #8303) #8177 and math/constants: compile again and pass 36 of 36, seal v2 (Closes #8301) #8209 it is 423, and closure rebuilds drop from 381 to 79.

Weak spots (self-critical)

Coverage is limited by seals, not by the idea. 940 specs (1201 after the hubs land) have no valid v2 seal: their tests fail or they do not compile.
Undeclared inputs in the bitstream cache key. The key includes the tools but not the python venv packages that fasm2frames and prjxray use. Bazel's main poisoning cause is exactly this.
No audit of cache hits. Nothing ever rebuilds a hit to check that it is still byte-identical. Bazel's remedy is to rebuild without the cache and compare digests.
The cache is local only. A toolchain-less operator (die C) cannot load a sealed bitstream yet.
Hooks anomaly. core.hooksPath is the absolute path /Users/playom/t27/.githooks, so every worktree runs the main checkout's branch hooks, which are not master's.
Shrink-only rule vs glue. New I/O glue must be paid for in the same PR. t27 cannot express the I/O of thin glue: file read and hash, process run, file write #8298 asks for t27 I/O so glue can be generated.
t27b ledger cap is full (20 of 20). New blocked rows go OVER CAP; this needs the owner's decision.
Inherited reds. Non-required checks (t27b-native-ratchet, Corpus ratchet, test-ratchet) are red on master often enough that a new red stops being noticed.

What others do (2026 scan)

Bazel remote cache: an action cache plus content-addressed storage. Reuse across machines is safe only if the build is reproducible. Undeclared host inputs are the usual poisoning; rebuild-and-compare is the audit.
Nix content-addressed derivations: still experimental and opt-in. Deep realisations are hard to make sound, so input-addressing remains the fallback.
Open FPGA flows (nextpnr, F4PGA): no cross-run bitstream cache found. Vendor incremental compile exists (Quartus).
Verified compute: proofs of execution on Artix-7 (VAPE), FPGA self-attestation (SACHa), receipts (flexHEG). No deployed FPGA receipt standard was found.

Plan

P1 Land the hubs, math/constants: compile again and pass 36 of 36, seal v2 (Closes #8301) #8209 (constants.t27) and specs/base/types.t27: compile again at 4 trits per byte, seal v2 (Closes #8303) #8177 (types.t27): closure 381 -> 79
P2 Second re-mint after P1: reseal as v2 every spec whose tests now pass
P3 Land t27c silicon reuses its bitstream; a ported script pays for the glue (Closes #8295) #8296: the bitstream reuse glue (net -114 hand-written lines)
P4 Bitstream key: add the venv package set (undeclared input). The spec decides, the glue hashes
P5 Hit audit: the spec decides when a hit is rebuilt and compared byte for byte, and what a mismatch means (poisoned entry -> delete and report)
P6 Debt: chore: delete 57 hand-written originals already ported to .t27 (Closes #8311) #8312 deletes 57 ported originals (-5873 lines); 25 more need review
P7 ops.t27 (base/ops.t27: compile again (i8 index, tuple literal) and fix the 7 tests that then fail (#8095 closure hub) #8239) and the remaining missing seals, in batches the Queen can take
P8 Shared sealed bitstreams for toolchain-less agents (die C)
P9 Hooks anomaly: propose a relative core.hooksPath (the owner's call: it is shared config)

Rules for every tick

Only .t27 for logic. Hand-written code may only shrink, and new glue is paid for in the same PR.
Every PR gets auto-merge. A required red is fixed before new work starts.
Never touch hardware without the owner's word.
Never weaken a test or a gate to get green.

Boundary

specs/verified/bitstream_reuse.t27

User Scenarios

Given a spec has a valid v2 seal and its tests pass, when the verdict reuse check runs, then the spec's verdict is reused without re-execution
Given the bitstream cache key includes the python venv package set, when the venv changes, then the cache key changes and a fresh bitstream build is triggered
Given a cache hit occurs for a bitstream, when the hit audit runs, then the bitstream is rebuilt without cache and the digests match byte-for-byte

Requirements

FR-001: The bitstream reuse spec MUST declare the cache key including the python venv package set hash
FR-002: The hit audit mechanism MUST rebuild a cached bitstream without cache and compare digests byte-for-byte, deleting and reporting poisoned entries on mismatch
FR-003: The verdict reuse logic MUST achieve 100% reuse for all specs with valid v2 seals (940 → 1201 after hubs land)
FR-004: The bitstream reuse glue (net -114 hand-written lines) MUST be landed with generated I/O glue paid in the same PR per shrink-only rule
FR-005: The t27b ledger cap MUST be resolved (currently 20 of 20 full) before new blocked rows are added

Success Criteria

t27c specs/verified/bitstream_reuse.t27 compiles and t27 test specs/verified/bitstream_reuse.t27 exits 0
t27 verdict-reuse --check reports 100% reuse for specs with valid v2 seals (exit 0)
grep -q "venv" specs/verified/bitstream_reuse.t27 exits 0 (venv package set in cache key)
Hit audit rebuilds 5 cached bitstreams and all compare byte-identical (exit 0)
wc -l specs/verified/bitstream_reuse.t27 shows net -114 hand-written lines or better vs baseline
t27b ledger --cap reports cap resolved (exit 0)

bitstream reuse: the venv in the key, and every 10th hit is audited (C…

Tick 1 (2026-10-10, ~03:15 +07)

Repaired:

t27c silicon reuses its bitstream; a ported script pays for the glue (Closes #8295) #8296 failed own-language twice.
- First, net growth: fixed by paying for the glue with the deletion of scripts/gen_w662.py, which has a .t27 port.
- Second, math_compare.rs is not on the exception list: that edit is withdrawn.
Net hand-written change of t27c silicon reuses its bitstream; a ported script pays for the glue (Closes #8295) #8296 is now -103.

New (PRs, all on auto-merge):

t27c silicon reuses its bitstream; a ported script pays for the glue (Closes #8295) #8296: the bitstream key now includes the venv's pip freeze (P4), and every 10th hit is audited (P5).
chore: delete 57 hand-written originals already ported to .t27 (Closes #8311) #8312: deletes 57 hand-written originals that were already ported to .t27 (-5873 lines) (P6).
math/constants: compile again and pass 36 of 36, seal v2 (Closes #8301) #8209 and specs/base/types.t27: compile again at 4 trits per byte, seal v2 (Closes #8303) #8177 taken over (P1): master merged, resealed v2, Closes specs/math/constants.t27 does not compile, so 223 specs rebuild on closure (#8095) #8301 and specs/base/types.t27 does not compile, so 340 specs rebuild on closure (#8095) #8303.

Anomaly found by the new audit.

Two builds of the same inputs differed in 3 bytes (offsets 162..165): the .bit header's build time.
Fixed by comparing from the sync word 0xAA995566.
Re-measured without hardware: audit run byte-identical, then a reuse in 5 s.

Measured: frontier on master 382 of 1739; projected 423 after P1, with closure 381 -> 79.

Skill: verdict-reuse-loop written, with the tick order, the CI rules and the traps.

Next tick: P1 or P3 merge status, then P2 (second re-mint) if P1 has landed.

Tick 2 (started ~03:30 +07)

P1 done: math/constants: compile again and pass 36 of 36, seal v2 (Closes #8301) #8209 and specs/base/types.t27: compile again at 4 trits per byte, seal v2 (Closes #8303) #8177 merged. frontier on master: 424 of 1743 (243 permille), closure 381 -> 79, missing 1203.
LOCK, P2 running: second re-mint of the 1203 missing specs in worktree $TMPDIR/wt-remint3 (branch claude/seal-v2-remint-3). Do not start another re-mint until this one reports "remint3 finished".
Anomaly, self-inflicted: math/constants: compile again and pass 36 of 36, seal v2 (Closes #8301) #8209 made constants.t27 parse cleanly (top-level-resync 37 -> 0). The Corpus ratchet reads that as a moved ledger entry, so it is red for every PR. Blessing docs/reports/suite_expectations.json on branch chore/suite-ledger-constants; only the constants.t27 entry will be kept.
t27c silicon reuses its bitstream; a ported script pays for the glue (Closes #8295) #8296 and chore: delete 57 hand-written originals already ported to .t27 (Closes #8311) #8312: waiting on CI, auto-merge on.

t27c silicon reuses its bitstream; a ported script pays for the glue (C…

t27b: the node scans of lower.rs move to specs/tri/t27b/ast_scan.t27 (C…

Tick 2, done (~04:45 +07). LOCK released: the re-mint finished.

P2: second re-mint. 1172 of 1203 sealed and 31 refused (failing tests). After filtering, 727 specs keep honest v2 seals: no blocked, no zero-test.
- PR seals: second v2 re-mint, 727 specs -- frontier reuses 1109 of 1743 (Closes #8371) #8372, Closes Second v2 re-mint after the hubs: every spec whose tests now pass gets a v2 seal (#8095) #8371.
- frontier on the branch: 1115 of 1751 (637 permille), up from 424 (243). Closure 73, missing 521.
Anomaly 1, self-inflicted: math/constants: compile again and pass 36 of 36, seal v2 (Closes #8301) #8209 moved a Corpus-ratchet pin (constants.t27, top-level-resync 37 -> 0). Fixed by PR suite ledger: re-record constants.t27 after #8209 (Closes #8345) #8355 (Closes Corpus ratchet red after #8209: constants.t27 discard channel moved (top-level-resync 37 -> 0) #8345), which re-records the entry and restores the history notes the writer drops. Re-checked: RATCHET CLEAN.
Anomaly 2: frontier reused specs with no checks at all (vacuous verdicts). Fixed by PR frontier: a verdict needs at least one check that ran (Closes #8354) #8374 (Closes frontier reuses a spec with no test, invariant or bench (vacuous verdict, L4) #8354): checks_pass needs tests + invariants > 0. Glue +1/-1.
Disk self-repair: removed the worktrees of merged PRs (free 9.4 GB).

Next tick:

watch t27c silicon reuses its bitstream; a ported script pays for the glue (Closes #8295) #8296, chore: delete 57 hand-written originals already ported to .t27 (Closes #8311) #8312, suite ledger: re-record constants.t27 after #8209 (Closes #8345) #8355, seals: second v2 re-mint, 727 specs -- frontier reuses 1109 of 1743 (Closes #8371) #8372, frontier: a verdict needs at least one check that ran (Closes #8354) #8374 to merge;
P7: the 31 specs with failing tests and the 521 still missing go to Queen-format issues in batches;
P8: shared sealed bitstreams;
tri command for the loop, spec first.

Tick 3 (~05:00 +07)

Merged since tick 2: t27c silicon reuses its bitstream; a ported script pays for the glue (Closes #8295) #8296 (bitstream reuse, net -103), chore: delete 57 hand-written originals already ported to .t27 (Closes #8311) #8312 (-5873 hand-written lines), suite ledger: re-record constants.t27 after #8209 (Closes #8345) #8355 (Corpus ratchet green), seals: second v2 re-mint, 727 specs -- frontier reuses 1109 of 1743 (Closes #8371) #8372 (727 v2 seals).
frontier on master: 1112 of 1751 (635 permille), measured with the frontier: a verdict needs at least one check that ran (Closes #8354) #8374 rule; closure 73, missing 521.
frontier: a verdict needs at least one check that ran (Closes #8354) #8374 (no vacuous verdicts) waits on CI. Its duplicate-bodies red is inherited from master: has_at in t27b/ast_walk.t27 duplicates starts_with in t27b/source_text.t27. Filed as duplicate-bodies red on master: ast_walk.t27 has_at duplicates source_text.t27 starts_with #8376.
LOCK, P7 running: classifying the 524 unsealed specs (blocked, failing, no checks, passes) in $TMPDIR/wt-classify, output $TMPDIR/classify.tsv. Next tick turns the classes into Queen-format issues, about 20 specs each, with exact failure lines.

Tick 3, done; tick 4 under way (05:56 +07)

Merged since the last comment

frontier: a verdict needs at least one check that ran (Closes #8354) #8374 (no vacuous verdicts).
frontier: resolve item-level use and module-less seals (Closes #8384) #8385: frontier resolves item-level use and module-less seals. 1115 -> 1151.
seals: v2 records for 59 specs the re-mint filter dropped (Closes #8386) #8387: v2 seals for 59 specs the second re-mint dropped (15 invariant-only). -> 1163.
seal: a use of an item names its module's spec in the closure (Closes #8351) #8391: one resolver for seal_closure and frontier (Closes seal v2 closure: use module::item is recorded as a missing spec module/item.t27, and the imported module is not in the closure #8351). 180 imports of actors.t27 were recorded as 180 missing entries; now they are one real one.

frontier on master: 1163 of 1756 (662 permille). rebuild: spec 47, closure 108, toolchain 3, verdict 3, missing 432.

Open, auto-merge on unless noted

frontier --reseal: the re-mint as one command (Closes #8389) #8392 t27c frontier --reseal: the re-mint as one command. The keep rule is in seal_identity.t27 (remint_wants, remint_keeps, REMINT_CACHE_EVERY). Paid for by deleting 2 dead generators. Auto-merge goes on once the full run below finishes clean.
gen-zig: an empty typed array &[_]T{} lowers to .{} (Closes #8390) #8415 gen-zig: &[_]T{} lowered to &.{ _ }, which blocked 17 specs (Closes gen-zig: empty typed array &[_]T{} lowers to &.{ _ } (blocks 17 specs) #8390). The regression spec is BLOCKED on master and 2/2 with the fix. Corpus ratchet CLEAN.

P7 done: epic #8393 lists 524 unsealed specs by class, across 21 Queen-format batches (#8394..#8414). 185 undeclared identifiers, 189 other compile errors, 21 failing, 40 with no checks. All were labelled queen-shaped automatically.

LOCK: t27c frontier --reseal is running on master in $TMPDIR/wt-remint5 (branch claude/seal-v2-stale-reseal), over 161 stale specs. It is both the full end-to-end test of #8392 and the next seal PR. A partial run (31 specs) wrote 22 files, all v2 with failed 0. Do not start another re-mint until this one reports.

Anomalies, self-critical

I ran the re-mint filter by hand three times and got it wrong once (seals: v2 records for 59 specs the re-mint filter dropped (invariant-only and late passes) #8386). The rule now lives in a spec.
run_seal calls process::exit on any refusal, so a loop cannot call it in-process; the glue runs it as a child.
A zig cache wiped after every spec costs about 20 s per spec. It is now shared and wiped every 25. FPGA testbench specs still take about 2 min each.

What others do (scan 2, 2026-10-10)

Bazel never caches a failing test. Old Buck reported test results from a cache hit until it bound the result to the rule key. Our seal already binds the verdict to spec, closure and toolchain.
Develocity PTS excludes flaky outcomes and re-selects tests that were skipped under the same inputs. We have no flake check: a verdict sealed once is reused forever. Proposal P10: sample a few REUSE specs per run, rerun them, and treat a mismatch as a poisoned seal, the same audit bitstream_reuse.t27 does.
Lake records a fetched artifact as a stand-in trace, not as a compile. P8 (shared bitstreams) should say the same in the receipt: "loaded from cache key K".
Dafny/Boogie cache per method along the call graph. Ours is per spec, so one edited test voids the whole spec's verdict. Finer granularity is possible later.

Next

Finish the re-mint and open its PR.
P10, verdict audit: spec first.
P8.
Rebase frontier --reseal: the re-mint as one command (Closes #8389) #8392 if it conflicts.

Tick 4 (19:10 +07). The session stalled for about 12 h after 06:30, so the cron queued about 50 ticks; this comment covers them as one.

Merged: #8387 (59 seals), #8391 (one use resolver for seal and frontier, Closes #8351), #8415 (gen-zig &[_]T{}, Closes #8390).

Anomaly, self-inflicted: #8415 left master red. It auto-merged on the required checks with three non-required checks red (#8417):

its regression spec declared struct Pair, which another spec also declares (types ratchet);
the codegen change drifted 45 specs' seals (Seal Coverage, spec-guards).

Repair, #8418, auto-merge on:

the type renamed;
tri seals drift --fix resealed 42 seals;
html, xml and terminal now compile and fail their own tests, so they are sealed --force and moved stale -> tests-fail, 6 baseline lines only;
it also carries gen-zig: a shadowing local's _lv rename reaches its own initializer (const PI_lv = PI_lv) #8416: the _lv rename reached the shadowing local's own initializer (const PI_lv = PI_lv).

On today's master, #8418's coverage still shows 56 [stale] seals. Those are other sessions' spec edits; no gen-drift is left.

Anomaly found by the end-to-end run of #8392 frontier --reseal: seal file collisions. Two specs can derive one seal file name, and seal --save overwrites the other spec's seal; in one case the glue then deleted it.

remint_keeps now takes file_is_its_own, so such a file is always written back.
The store-level defect is filed as seal store: two specs derive one seal file name, and seal --save overwrites the other spec's seal #8540 (5 files).
frontier --reseal: the re-mint as one command (Closes #8389) #8392 waits for a clean full run (in progress in $TMPDIR/wt-remint5) before auto-merge.

frontier on master: 1144 of 1777 (644 permille). Down from 1167 of 1756, because 86 specs were edited today without a reseal (spec 86). This is why --reseal has to run on every tick, not once.

P7: epic #8393, 21 batches. Ranking the hubs behind closure rebuilds:

hub	specs held behind it
`math/sacred_physics.t27`	26
`base/ops.t27` (#8239)	24
`isa/registers.t27`	10
`compiler/lexer.t27`	9

sacred_physics stops at sin2_theta12_trinity(). Nothing declares it, and its comment formula 8*phi^4/3/pi evaluates to 5.818, not 0.309. That needs the author (#8396).

P10: spec verdict_audit.t27, #8543 (Closes #8542). Glue next.

Inherited reds, not mine:

specs/memory/graph/ingest.t27: unterminated string. It fails emit-bitexact and the Corpus ratchet.
8 UNLISTED t27b rows (port specs).
5 parse-no-discard entries.

Next:

finish the re-mint, then its PR;
auto-merge frontier --reseal: the re-mint as one command (Closes #8389) #8392;
glue for frontier --audit;
hub base/ops.t27.

Tick 4, continued (19:35 +07)

Merged: frontier --reseal: the re-mint as one command (Closes #8389) #8392 frontier --reseal, spec: verdict_audit.t27 -- rerun a sample of reused verdicts (Closes #8542) #8543 verdict_audit.t27, base/ops.t27: compiles; 1 wrong function and 4 wrong tests fixed (Closes #8239) #8547 base/ops.t27 (95/95, Closes base/ops.t27: compile again (i8 index, tuple literal) and fix the 7 tests that then fail (#8095 closure hub) #8239), types ledger: drop Pair, resolved by #8418 (Closes #8549) #8550 (Pair left the types ledger, Closes types ledger: Pair was blessed in while master was red, and #8418 resolved it #8549).
seals: the first frontier --reseal run on master, 41 specs (Closes #8551) #8554, auto-merge on: the first --reseal run on master. 194 stale specs: 41 resealed, 67 files, 0 bad, 0 collisions. frontier on master 1146 -> 1173 of 1780.
- Anomaly caught before commit: tri seals drift --fix resealed two specs that PASSED on master and stopped compiling today (compiler/optimizer.t27 after Census the optimizer: 4 of 7 passes have no precondition, and dead_store_elim deletes calls #8502, port/tools/gft_deep_demo.t27 after Port part 1 of 2 of tools/gft_deep_demo.py to specs/port/tools/gft_deep_demo.t27 (8 functions) #8494). Sealing them would have recorded the regression as truth. I reverted those seals and filed regression: compiler/optimizer.t27 and port/tools/gft_deep_demo.t27 stopped compiling on 2026-10-10 #8552.
frontier --audit: rerun a sample of reused verdicts (Closes #8556) #8559, auto-merge on: t27c frontier --audit <round> (Closes t27c frontier --audit: rerun a sample of reused verdicts (glue for verdict_audit.t27) #8556).
- Round 3 on master: 65 reused verdicts rerun, 65 agree, 0 poisoned, 146 s.
- Negative control: a planted total 15 for 14 tests reads POISONED and exits 1.
- It also adds the t27b row CI measured for zig_shadow_initializer.t27 (pass), which gen-zig: a shadowing local reads the outer name in its initializer; master green after #8415 (Closes #8416, Closes #8417) #8418 had left UNLISTED.
frontier note: verdict went 1 -> 27 and missing 439 -> 417 on master because gen-zig: a shadowing local reads the outer name in its initializer; master green after #8415 (Closes #8416, Closes #8417) #8418's drift fix gave blocked specs v2 records. Same specs, different column.

Next tick:

--reseal and --audit on master (both now in the tick order of the skill);
hub isa/registers.t27 (10 specs);
generics: 21 specs fail on an undeclared T.

Tick 5 (19:55 +07)

Merged: #8554 (first --reseal run, Closes #8551), #8559 (frontier --audit, Closes #8556), #8561 (stale Pair classification row, Closes #8560).

frontier on master: 1177-1178 of 1780 (662 permille).

Audit, round 5: 65 reused verdicts rerun, 65 agree, 0 poisoned, 166 s. Round 3 earlier: 65/65.

Re-mint, tick 5:

Under the claim tri loop claim verdict-reuse-remint, now the lock between ticks and sessions. 12 resealed, 830 s.
Rebase conflict: Reseal the #8415 FROZEN_HASH wave (56 seals) + trinet_node j_go test observes the accept edge #8564, another session's 56-seal reseal, had touched the same files. After keeping master's version, only built_by and sealed_at differed, so seals: tick-5 frontier --reseal run, 12 specs (Closes #8567) #8568 was closed as churn.
Follow-up frontier --reseal: a record that differs only in built_by is not a change (Closes #8571) #8572 (Closes frontier --reseal: a record that differs only in built_by is not a change #8571): --reseal treats a built_by-only difference as no change.

Compiler, #8566 (Closes #8565), auto-merge on:

The dead-const pass ended a block at an unindented break; and discarded consts that were used after a loop.
Measured over every spec: 10 specs change, only by removed discard lines. All were blocked and stay blocked; lexer (a hub), router and hybrid_bigint get past this error.
Corpus ratchet: master's 5 inherited failures, none new.

Anomalies:

Bees duplicate issues I file for my own commits. queen-8389 -> PR t27c frontier --reseal: the re-mint as one command, its keep rule in seal_identity.t27 #8505 (closed as a duplicate of frontier --reseal: the re-mint as one command (Closes #8389) #8392), plus queen-8416, queen-8417, queen-8549, and queen-8314 (this ledger). The Queen does not skip issues with an open closing PR: filed trios#1736.
Master Corpus ratchet: 5 unexpected failures from other merges today, filed with diagnoses as master Corpus Ratchet: 5 unexpected failures from 2026-10-10 merges (ingest.t27 apostrophe, a/b_c, optimizer, orchestrator_bench) #8562. memory/graph/ingest.t27:414 is a ; comment containing "let's", whose apostrophe opens a char literal, and the comment describes a deliberately wrong assert.
I pushed during a stopped rebase once. The pushed commit was the old one; the PR was then closed as churn anyway. Now a rule in the skill.

Next:

re-mint (claim first) and audit round 6;
the lexer.t27 hub's next error;
P8 shared bitstreams.

Tick 6 (20:02 +07)

Audit, round 6: 75 reused verdicts rerun, 75 agree, 0 poisoned, 170 s.
P8 spec, spec: bitstream_share.t27 -- shared sealed bitstreams for agents without a toolchain (Closes #8573) #8574 (Closes P8: shared sealed bitstreams for agents without a toolchain -- spec bitstream_share.t27 #8573): bitstream_share.t27.
- bit_source: BUILD / SHARED / WAIT / REFUSE. A toolchain-less agent loads a shared bitstream only when the fetched bytes hash to the sha256 named by the owner-approved job. A mismatch is refused, never built around.
- may_publish and receipt_complete: a SHARED receipt names the key and the bit hash (Lake's stand-in trace).
- 4/4, 0 vacuous. Glue next.
Self-repair as a tool, seal --save refuses to record a regression without --force (Closes #8576) #8577 (Closes seal --save: refuse to record a regression (a passing seal replaced by blocked) without --force #8576): seal --save refuses to replace a passing seal with blocked unless --force (blocked_reseal_allowed). tri seals drift --fix now reports such a spec as REFUSED instead of writing the regression. Negative control: a planted undeclared call is refused with exit 1 and the seal is untouched.
Waiting in CI, auto-merge on: gen-zig: the dead-const pass no longer ends a block at an unindented break; (Closes #8565) #8566 (dead-const pass), frontier --reseal: a record that differs only in built_by is not a change (Closes #8571) #8572 (built_by-only reseal is no change).
Lexer hub, next error: check_keyword takes [256]u8 and the 8 keyword tests pass string literals. A spec-test fix (lexeme_buf("fn")) waits for gen-zig: the dead-const pass no longer ends a block at an unindented break; (Closes #8565) #8566, which touches lexer's seals.
Bees: queen-8560 (already delivered by types classification: drop the stale Pair row (Closes #8560) #8561; any PR from it will be closed as a duplicate) and queen-8562 (master's reds, a legit task).
Inherited reds: compiler/optimizer.t27 also fails spec-guards' duplicate-declaration check on every PR. Added to regression: compiler/optimizer.t27 and port/tools/gft_deep_demo.t27 stopped compiling on 2026-10-10 #8552.

Re-mint this tick: skipped. The last run was 15 minutes ago and its records duplicated #8564's.

Tick 7 (20:10 +07)

Merged: gen-zig: the dead-const pass no longer ends a block at an unindented break; (Closes #8565) #8566 (dead-const pass, Closes gen-zig: the dead-const pass ends a block at an unindented break; and discards consts that are used #8565), frontier --reseal: a record that differs only in built_by is not a change (Closes #8571) #8572 (built_by-only reseal is no change, Closes frontier --reseal: a record that differs only in built_by is not a change #8571).
Lexer hub, compiler/lexer.t27 compiles: keyword tests use the lexeme buffer, invariants compare tags (Closes #8578) #8583 (Closes compiler/lexer.t27: the keyword tests pass string literals where check_keyword takes [256]u8 (hub, 9 dependents) #8578), auto-merge on: compiler/lexer.t27 compiles for the first time, 16/16, 0 vacuous.
- The keyword tests now build the [256]u8 lexeme buffer with a spec helper, lexeme_buf.
- The invariants compare enum tags.
- frontier +1 (1180). The 9 dependents now show their own defects (verdict 32 -> 35).
Lesson (self-critical): a hub's dependent count overstates what fixing it frees. ops gave +3, lexer +1. The closure walk counts specs that also have their own blockers. Rank hubs by dependents whose only blocker is the hub next time.
P8 glue parked: a toolchain-less agent cannot compute the cache key, so t27c silicon needs a load-entry path verified against the job's sha256, and that needs one SRAM hardware check on die A. It waits for the owner's word.
In CI, auto-merge on: spec: bitstream_share.t27 -- shared sealed bitstreams for agents without a toolchain (Closes #8573) #8574 (P8 spec), seal --save refuses to record a regression without --force (Closes #8576) #8577 (seal --save refuses a pass -> blocked regression).

Re-ranking, self-correction. Counted strictly (the dependent's own seal passes AND the hub is its only non-reusable import), the hubs free very little: sacred_physics 3, then notebooklm, parser, top_level, formal, placement, vcd_trace and runner 1 each. My earlier "26 / 24 / 10 / 9" counted dependents that have blockers of their own, which is why ops gave +3 and lexer +1. Consequence for the plan: the remaining ~600 non-reused specs fail on their OWN defects (missing 418, verdict 35). Coverage grows per spec fixed, so the engine is the swarm on #8393's 21 batches, once trios#1736 stops duplicate dispatch, plus shared compiler defects. Hub hunting is closed as a strategy.

Tick 8 (20:45 +07)

Merged: spec: bitstream_share.t27 -- shared sealed bitstreams for agents without a toolchain (Closes #8573) #8574 (P8 spec bitstream_share.t27), seal --save refuses to record a regression without --force (Closes #8576) #8577 (seal --save refuses pass -> blocked), compiler/lexer.t27 compiles: keyword tests use the lexeme buffer, invariants compare tags (Closes #8578) #8583 (lexer.t27 compiles, 16/16), gen-zig: a local declared []T from an array literal lowers to @constCast(&[_]T{...}) (Closes #8602) #8617 (Closes gen-zig: a local declared []T and initialized with an array literal lowers to .{...}, which is no slice #8602).
gen-zig: a local declared []T from an array literal lowers to @constCast(&[_]T{...}) (Closes #8602) #8617: gen-zig lowers a slice-typed local initialized from an array literal to @constCast(&[_]T{...}).
- Measured over every tracked .t27: 28 specs change, each only at such locals. None passed before; 2 move past this error.
- 56 seal files refreshed, blocked -> blocked or none -> blocked.
- Corpus ratchet: master's 5 inherited, none new. Paid for by deleting a dead generator.
Self-correction (posted earlier): the strict hub ranking shows each hub frees 1 to 3 specs, not 9 to 26. The remaining ~600 specs fail on their own defects. The levers now are the swarm on Unsealed specs: 524 classified, in batches the Queen can take (#8095 P7) #8393 and shared compiler defects.
Anomaly: the disk hit 0 bytes at 20:40 (other sessions' worktrees hold about 7 GB). The Bash tool failed with ENOSPC before running anything. I recovered through the app's storage report and the terminal panel, deleting only my own temp data; 2.3 GB free now. The P7 re-classification (1 h at today's load) was stopped and dropped. Recovery and prevention steps are now in the skill.
Evidence for regression: compiler/optimizer.t27 and port/tools/gft_deep_demo.t27 stopped compiling on 2026-10-10 #8552: Reseal the #8415 FROZEN_HASH wave (56 seals) + trinet_node j_go test observes the accept edge #8564 resealed gft_deep_demo as blocked on master, the regression seal --save refuses to record a regression without --force (Closes #8576) #8577 now refuses.

Tick 9 (21:05 +07)

frontier on master: 1184 of 1786 (663 permille). Disk: 2.0 GB free at the start (checked first, per the new rule).
gen-zig: escape a keyword enum literal in switch prongs and enum values (Closes #8621) #8636, auto-merge on (Closes gen-zig: an enum literal that is a Zig keyword (.error) is not escaped in switch prongs and enum values #8621): gen-zig escapes a keyword enum literal (.error -> .@"error") in switch prongs and enum values.
- Over every tracked .t27, 9 specs change, only at such literals. All 9 were blocked on exactly that syntax error, and all 9 get past it.
- 18 seal files refreshed blocked -> blocked. Corpus ratchet: master's 5 inherited, none new.
Error classes, from the batch issues (Unsealed specs, do not compile: undeclared identifiers: batch 1 of 9 (21 specs) #8394..Unsealed specs, no test or invariant (L4): batch 2 of 2 (20 specs) #8414), excluding undeclared identifiers:
- expected ',' after field (12): angle-bracket generics (List<Int>, Box<OctNode>, map<K,V>), the Result<T,E> is Zig's E!T: 182 angle-bracket generics, 163 with exact counterparts #2610 family.
- expected ')', found ',' (8): Python-style tuple types in ported specs.
- Both are foreign syntax in the specs themselves, which is batch work for the swarm, not codegen bugs.
- The codegen bugs left are small classes: 7 local variable is never mutated, 7 invalid builtin.
Plan note: the shared compiler defects that move whole classes are nearly exhausted: gen-zig: the dead-const pass no longer ends a block at an unindented break; (Closes #8565) #8566, gen-zig: a local declared []T from an array literal lowers to @constCast(&[_]T{...}) (Closes #8602) #8617 and gen-zig: escape a keyword enum literal in switch prongs and enum values (Closes #8621) #8636 each moved 2 to 9 specs past one error. What remains is per-spec work in the 21 batches.

Tick 10 (21:30 +07)

Disk: 3.1 GB free. gen-zig: escape a keyword enum literal in switch prongs and enum values (Closes #8621) #8636 (escape keyword enum literals) in CI; its reds are the inherited duplicate-bodies (bits, clear) and ingest.t27.
Next compiler defect, gen-zig: a pointer or slice parameter written through (p.*, s[i]) is re-bound as var and Zig says never mutated #8641: local variable is never mutated (7 specs, mostly isa/*).
- Cause: collect_mutable_names counts p.* = ... and s[i] = ... as reassigning a pointer or slice parameter, so the parameter gets a var x = x_arg; shadow that Zig refuses.
- Fix, committed locally: such a parameter is re-bound only on direct reassignment.
Self-critical: my "only failing specs change" reasoning was wrong. The fix moves the gen output of 181 specs, 74 of which pass today, for example base/types.t27. Zig compiles lazily, so a function no test calls is never analyzed, and a passing spec can hold a never-mutated re-binding. That means:
- the PR must test all 74 passing specs before and after, and reseal every moved spec (about 400 seal files with twins);
- my "before" binary carried the unmerged gen-zig: escape a keyword enum literal in switch prongs and enum values (Closes #8621) #8636, so the before binary is rebuilt from exactly master once gen-zig: escape a keyword enum literal in switch prongs and enum values (Closes #8621) #8636 merges.
- Both lessons are now in the skill.
Not compiler bugs: invalid builtin (@tanh, @sigmoid, @trim, ...) is builtins the specs invented. That is batch work.

Next tick: after #8636 merges:

rebase gen-zig: a pointer or slice parameter written through (p.*, s[i]) is re-bound as var and Zig says never mutated #8641 and rebuild before/after from master;
test the 74 passing specs before and after (abort if any outcome changes);
reseal the moved specs; run the Corpus suite; open the PR.

Tick 10, finished (22:15 +07)

Merged: gen-zig: escape a keyword enum literal in switch prongs and enum values (Closes #8621) #8636 (keyword enum literals).
gen-zig: a pointer or slice parameter written through is not re-bound as var (Closes #8641) #8667, auto-merge on (Closes gen-zig: a pointer or slice parameter written through (p.*, s[i]) is re-bound as var and Zig says never mutated #8641): pointer and slice parameters written through are no longer re-bound as var.
- 173 specs change, only by removed re-bindings.
- All 74 that pass today pass identically after (tested before and after with a master build).
- The 7 targets get past "never mutated".
- 274 seal files refreshed (96 pass -> pass, 143 blocked -> blocked, 35 none -> blocked, 2 forced -> forced).
- Local checks: coverage OK, drift 0, types CLEAN; Corpus ratchet shows master's 5 inherited, none new.
Anomalies:
- 2 more seal-name collisions (seal store: two specs derive one seal file name, and seal --save overwrites the other spec's seal #8540) were caught and restored. A bulk reseal must always diff spec_path against master.
- A peer session asked me to fix ingest.t27 as "my commit". It is a bee's (queen-8425 via memory/graph slice 4: ingest.t27 -- the extraction state machine and the LLM effect slot #8493), already filed as master Corpus Ratchet: 5 unexpected failures from 2026-10-10 merges (ingest.t27 apostrophe, a/b_c, optimizer, orchestrator_bench) #8562, and bee branch queen-8562 holds an unpublished fix. I replied and left it alone to avoid a third copy.
Disk: 7.1 GB free (others cleaned up).

Tick 11 (22:55 +07)

Merged:
- gen-zig: a pointer or slice parameter written through is not re-bound as var (Closes #8641) #8667 (Closes gen-zig: a pointer or slice parameter written through (p.*, s[i]) is re-bound as var and Zig says never mutated #8641): pointer and slice parameters are no longer re-bound. 173 specs moved; all 74 that pass today pass identically.
- seals: let_reassigned.t27 after #8667 (Closes #8671) #8672 (Closes master seal coverage: let_reassigned.t27 drifted after #8667 #8671): the seal of let_reassigned.t27, which landed during gen-zig: a pointer or slice parameter written through is not re-bound as var (Closes #8641) #8667. gen-zig: a pointer or slice parameter written through is not re-bound as var (Closes #8641) #8667 had auto-merged on its required checks with seal coverage red for it.
seals: tick-11 frontier --reseal run, 3 seal files (Closes #8687) #8688, auto-merge on (Closes seals: tick-11 frontier --reseal run (4 seal files) #8687): the routine re-mint on the gen-zig: a pointer or slice parameter written through is not re-bound as var (Closes #8641) #8667 compiler, under the claim.
- 13 resealed, of which 3 seal files carry real changes; the rest were timestamp-only and written back (frontier --reseal: a record that differs only in built_by is not a change (Closes #8571) #8572).
- Drift 0 after the final rebase.
Audit, round 11: 54 rerun, 54 agree, 0 poisoned.
frontier on master: see the line below, measured on the seals: tick-11 frontier --reseal run, 3 seal files (Closes #8687) #8688 branch.
Lessons, self-critical, now in the skill:
- Run tri seals drift after the FINAL rebase. Specs land during a PR, and a codegen change moves them too.
- When a non-required check goes red because of MY change, disable auto-merge until it is fixed. This happened twice (gen-zig: an empty typed array &[_]T{} lowers to .{} (Closes #8390) #8415, gen-zig: a pointer or slice parameter written through is not re-bound as var (Closes #8641) #8667).
- The first tick-11 re-mint ran on the pre-gen-zig: a pointer or slice parameter written through is not re-bound as var (Closes #8641) #8667 compiler. It was stopped and rerun, because its gen hashes would have drifted.
Competitor scan, posted to trios#1736: Copilot coding agent claims by assignment plus an immediately linked draft PR. Gitworthy's preflight skips work with an active PR or an unlinked branch.
frontier on the branch: frontier: reused 1197 of 1806 (663 permille); rebuild: spec 48 closure 102 toolchain 0 config 0 verdict 111 missing 348

Tick 12 (23:50 +07). The biggest single gain of the day.

Finding: of 348 MISSING specs, 101 had no seal at all, and 44 of those passed their tests. --reseal never tried them, because remint_wants skipped every MISSING spec. The assumption "MISSING needs fixing first" was mine (frontier --reseal: the re-mint as one command (Closes #8389) #8392) and wrong for never-sealed specs.
frontier --reseal tries a never-sealed spec; 39 pass and are sealed (Closes #8691) #8700, auto-merge on (Closes frontier --reseal never tries a never-sealed spec: 44 of 101 pass and have no seal #8691): remint_wants(decision, has_seal) tries never-sealed specs. The first run on master sealed 39 of them.
- 2 fully vacuous suites held back: checks_pass accepts them, filed as checks_pass accepts a fully vacuous suite (every test ran 0 asserts) #8693.
- frontier 1197 -> 1240 of 1810 (685 permille), missing 348 -> 308. Drift 0 after the final rebase.
Merged: seals: tick-11 frontier --reseal run, 3 seal files (Closes #8687) #8688 (tick-11 re-mint).
New weak spot, checks_pass accepts a fully vacuous suite (every test ran 0 asserts) #8693: checks_pass accepts a suite whose every test ran 0 asserts. Proposal: record tests.vacuous in the seal and require (total - vacuous) + invariants > 0.
Disk: 16 GB free.

Loop stopped by the owner (2026-10-11 00:50 +07, cron 52ab9d25 deleted). Final state.

frontier on master: reused 1242 of 1813 (685 permille); rebuild: spec 49 closure 102 toolchain 0 config 0 verdict 111 missing 309. It was 2 of 1712 on 2026-10-09.

Landed, by PR:

Reuse and its tools:
Guards:
- no verdict from no checks: frontier: a verdict needs at least one check that ran (Closes #8354) #8374;
- none from fully vacuous suites: checks_pass refuses a fully vacuous suite; seals record tests.vacuous (Closes #8693) #8709, audit: a reused verdict whose rerun checked nothing is poisoned (Closes #8710) #8719;
- seal --save refuses a regression: seal --save refuses to record a regression without --force (Closes #8576) #8577.
Compiler (t27 -> Zig): gen-zig: an empty typed array &[_]T{} lowers to .{} (Closes #8390) #8415, gen-zig: a shadowing local reads the outer name in its initializer; master green after #8415 (Closes #8416, Closes #8417) #8418, gen-zig: the dead-const pass no longer ends a block at an unindented break; (Closes #8565) #8566, gen-zig: a local declared []T from an array literal lowers to @constCast(&[_]T{...}) (Closes #8602) #8617, gen-zig: escape a keyword enum literal in switch prongs and enum values (Closes #8621) #8636, gen-zig: a pointer or slice parameter written through is not re-bound as var (Closes #8641) #8667. Plus base/ops.t27 (base/ops.t27: compiles; 1 wrong function and 4 wrong tests fixed (Closes #8239) #8547) and compiler/lexer.t27 (compiler/lexer.t27 compiles: keyword tests use the lexeme buffer, invariants compare tags (Closes #8578) #8583), both compiling for the first time.
Silicon: the bench agent (verified: an agent next to each board -- bench_agent.t27 and t27c bench-agent (Closes #8153) #8155) and bitstream reuse with an audit (t27c silicon reuses its bitstream; a ported script pays for the glue (Closes #8295) #8296); P8 spec bitstream_share.t27 (spec: bitstream_share.t27 -- shared sealed bitstreams for agents without a toolchain (Closes #8573) #8574).
Blog: https://t27.ai/blog/rebuild-only-what-changed/ (trinity#1595), published by hand. X goes via the blog-share desk after the owner's Telegram yes.

Last finding: the audit with #8719 found 9 reused verdicts that check nothing in 2 of 16 rounds (#8721). The other 14 rounds will name the rest.

Open, needs the owner:

Unsealed specs, do not compile: batch 2 of 9 (21 specs) #8404: register format (10 specs);
Unsealed specs, do not compile: undeclared identifiers: batch 3 of 9 (21 specs) #8396: the sacred_physics formula;
trios#1736: the Queen should skip issues that already have an open PR;
the P8 hardware run.

Open for the swarm: the 21 batches of #8393, and #8721.

Lessons: in the verdict-reuse-loop skill.

Navigation Menu

Repository navigation

LOOP 2026-10-10: verdict reuse to 100% -- ledger, weak spots, plan (epic #8095) #8314

Where we are (measured)

Weak spots (self-critical)

What others do (2026 scan)

Plan

Rules for every tick

Boundary

User Scenarios

Requirements

Success Criteria

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

LOOP 2026-10-10: verdict reuse to 100% -- ledger, weak spots, plan (epic #8095) #8314

Description

Where we are (measured)

Weak spots (self-critical)

What others do (2026 scan)

Plan

Rules for every tick

Boundary

User Scenarios

Requirements

Success Criteria

Activity

gHashTag commented on Oct 9, 2026

gHashTag commented on Oct 9, 2026

gHashTag commented on Oct 9, 2026

gHashTag commented on Oct 9, 2026

gHashTag commented on Oct 9, 2026

gHashTag commented on Oct 10, 2026

gHashTag commented on Oct 10, 2026

gHashTag commented on Oct 10, 2026

gHashTag commented on Oct 10, 2026

gHashTag commented on Oct 10, 2026

gHashTag commented on Oct 10, 2026

gHashTag commented on Oct 10, 2026

gHashTag commented on Oct 10, 2026

gHashTag commented on Oct 10, 2026

gHashTag commented on Oct 10, 2026

gHashTag commented on Oct 10, 2026

gHashTag commented on Oct 10, 2026

gHashTag commented on Oct 10, 2026

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions