Why: Allow consumers to download an immutable module revision without nested source checkouts or local path replacements.
Evidence: SELF package tests and the prior canonical-path rejection are recorded in the Dev dependency-workspace task card; remote consumption is replayed against the committed candidate.
Impact: Change the module identity and its Go imports while preserving API behavior. Existing historical submodule revisions remain buildable.
Refs: Versioned dependencies U1/U3; feature/dependency-workspace
Why: Reminders are attached to the request copy and never reach the
session journal, so a transcript grep proves nothing about whether the
R4 trigger fired - the first verification attempt was invalid evidence.
The dual-speed nudge already logs for exactly this reason.
Evidence: go test ./internal/runtime green; the follow-up live probe
must show the [runtime] r4 escalation reminder injected line in the
instance log.
Impact: Two log lines, nothing else.
Refs: Dev task feature/agent-e4-escalation-ledger (third fix-up).
Why: The consecutive-failure counter never fired in the live suite (0
reminders despite 20 failures) because real failures interleave with
successes - each success reset the streak. The pattern to catch is
"spinning overall", not "N in a row".
Evidence: go test ./... green; the anchor test now covers scattered
failures (3 in the last 6 outcomes escalates, 2 does not, once per run);
the next live run must show the reminder actually appearing in the
transcript.
Impact: recentOutcomes ring (last 6) replaces consecToolFailures; the
zero-hit search trigger keeps its own threshold (2); message wording
matches the window semantics.
Refs: Dev task feature/agent-e4-escalation-ledger (second fix-up).
Why: The first E4 live comparison regressed badly (toolFailures 10->24,
toolCalls 63->166) and the instance transcript shows the loop: tagging
kernel read_file/grep results put a trailing [src_N] line on the
model's own workspace files, the model treated it as external
tampering, and re-read/re-edit cycles cascaded into the read-before-
write and file-changed guards.
Why: The ledger only ever needed SOURCES - searches always, and
workspace reads only under inputs/ (the upload material area). Grep
cannot distinguish materials from self-produced files, so it stays
untagged.
Evidence: go test ./... green; new anchor TestEvidenceLedger
OnlyTagsMaterialReads pins the inputs/-only rule; the re-run live
comparison follows on the Dev side.
Impact: evidenceTools loses read_file/grep; isEvidenceRead gates
workspace reads to inputs/; everything else (search tags, escalation
reminders, prompts) unchanged.
Refs: Dev task feature/agent-e4-escalation-ledger (fix-up of #131).
Why: The plan's E4 remainder under the MiniMax-only constraint: the
harness must detect the model going in circles and inject a
change-of-strategy instruction (R4 - the model cannot be trusted to
notice its own repeat-fail loop), and number sourcing needs addressable
evidence: tool results carry [src_N|tool] tags so "found in materials"
becomes "found in source src_N" for both the model and the fact-check
notice.
Why: Failure counting lives in one place (onEvent end, keyed by the
slot status) because tool execution is parallel; the ledger and tag
append are serialized by the same mutex, and escalation reminders fire
once per condition streak - following the dual-speed nudge finding that
MiniMax ignores vague reminders but follows concrete instructions.
Evidence: go test ./... green; new anchors TestEscalationOnConsecutive
ToolFailures, TestEscalationOnZeroHitSearches, TestEvidenceLedger
TagsSource, TestEvidenceLedgerSkipsNonEvidenceTools. Live before/
after comparison runs on the Dev side per ruling 6.
Impact: knowledge/web/tavily searches, fs_read_file and the workspace
read_file/grep results now end with a source tag and enter the run
ledger (2k-char head); two consecutive tool failures or two empty
search results inject a concrete change-of-strategy reminder;
base.deliverables v7 and agent-sub-research v3 explain the tags;
fact-check notice points at src results. Non-evidence tools and
failures are unchanged.
Refs: proposal 2026-10-10-agent-delegation-routing-sandbox.md (E4,
R4); Dev task feature/agent-e4-escalation-ledger.
Why: The E1 live comparison was the only metric that got worse
(unverifiedFacts 7 -> 12): the gate gave one correction round and then
shipped bare numbers. The design (E3) calls for number sourcing with a
mechanically checkable marking escape: numbers labeled as
targets/assumptions are honest planning quantities, not fabricated
facts, so the gate must let them pass - and the model must be told that
marking is the sanctioned way out.
Why: The marking decision is clause-scoped, not a fixed window - a
window wrongly blessed "3,500" because a neighboring clause said
"target 85%" (caught by the existing test). The second round only
fires when the unverified set actually changed, because the
continuation budget is shared across all final gates and a repeat
block on identical tokens starves the downstream gates (also caught by
an existing test).
Evidence: go test ./... green; new anchors TestMarkedAsEstimate,
TestUnverifiedFactsMarksPass, TestFactCheckGateRounds; the live eval
gains an unverifiedUnmarkedFacts field sharing the same marking rule.
Impact: base.deliverables v6 adds the hard number-sourcing rule;
research/explore/general sub-agent profiles v+1 require citations for
key numbers; fact gate: marked-in-clause numbers pass, round 2
escalates to "last interception" with the marking instruction, then
passes through with the check line. Eval report keeps the strict
series and adds the E3-series field.
Refs: proposal 2026-10-10-agent-delegation-routing-sandbox.md (E3,
D4); Dev task feature/agent-e3-evidence-gate.
Why: User rulings 4/5 plus the P2-P4 live evidence. Keyword scoring dimensions
(readability, actionability) kept failing valid documents - normal Chinese
section headings flagged as column titles, loops of 4 corrections - because
they grade writing preferences, not verifiable facts. Unverified numbers
(~10-13 per run) remained the largest quality leak through every phase.
And the verifier runs on the same model as the worker (lightweight falls back
to chat), so "independent" verification was self-review. Separately, svc
seeding after boot never registered the tavily tools until a restart because
keys were snapshotted at startup.
Changes: quality.HardFacts (lint actually ran + zero errors) replaces the
full ScoreBytes projection in CheckArtifacts - keyword dimensions remain in
ScoreBytes for eval visibility only. New fact-consistency check
(runtime.factCheck, default on): extract percentages/dates/large numbers from
deliverable text, trace each against the material corpus (attachment text +
tool evidence); misses get exactly one correction round (fix, drop, or mark
as target/estimate), never a loop. runVerifier skips same-source model
verification unless runtime.verifierSameModel is set (hard facts + fact check
still run). Search tools now register lazily, reading tavily keys per call
(settings/seeding take effect immediately).
Evidence: SELF - TestHardFactsIgnoresKeywordDimensions,
TestHardFactsKeepsLintUnavailableBlocking, TestUnverifiedFactsAndNotice,
TestFactCheckContinuationOnce (fact note fires, corrected run passes);
go vet + go test ./... pass; live before/after on the golden task set
recorded on the phase05 task card per the policy clause.
Impact: Gate stops rejecting valid documents on stylistic grounds; invented
figures get one forced correction round; same-model verification no longer
pretends to be independent; tavily config is hot. No contract or storage
changes.
weichen