belief-merge (belief-merge)

BeliefMerge

Merging two AI conversations without losing the reasoning, laundering trust, or guessing at contradictions.

A DeepSeek Harness plugin, a confluence theorem with a counterexample, and a benchmark that was wrong twice before it was useful.

dsh plugin --profile web add dsh-belief-merge

Then just ask: "merge my two bilibili conversations into this one" — or drive it by hand with /merge to list candidates, pick them, and check what is in effect. No config file to edit, no restart; it applies on the next turn.

Full usage, configuration and limitations: belief-merge/README.md.

Working on the code instead:

cd belief-merge
npm test             # 242 tests, no dependencies, no network, no API key
npm run demo         # end-to-end, offline
npm run bench:tight  # the budgeted comparison

---

The problem, in two sentences

Two conversations about the same system drift apart, and a later question needs facts from both. Every existing tool answers this by pasting one into the other (cheap, unbounded, contradictions left to the reader) or by summarizing (lossy exactly where it matters — the reasoning, including the dead ends that stop you repeating them).

---

Four results

1. Incremental merging is unsound unless you keep provenance

Branches arrive over time, so merging happens pairwise. Swapping the order of two branches then changes the answer — and it does so silently:

B0 = {a@1}      B1 = {a@1}      B2 = {b@2}

n-ary        : disputed{a,b}      correct — weights are 1+1 = 2 against 2
incremental  : b                  WRONG

Merging B0 and B1 first collapses two weak branches that agree into one weak assertion, so a single strong dissenter wins.

Over 72,760 enumerated profiles, a provenance-retaining representation produced zero failures; collapsing to a bounded summary failed on 9.9–26.2% of well-formed inputs — and the failure cascades through derivation chains, deleting or resurrecting whole lines of reasoning.

"Summarize as you merge" is unsound, no matter how good the summarizer is.

2. At a realistic budget, concatenation collapses

MergeBench generates scenarios where bridge questions need facts from different branches — so no single branch can answer them. Every arm gets the same token budget.

400 tokens, 50 bridge questions:

arm	local	bridge	resolves conflicts	tokens
one branch alone	50%	0%	50%	388
concatenation	60%	18%	50%	399
BeliefMerge	95%	98%	100%	401

At a loose budget everything fits and concatenation keeps up on coverage — but it cannot resolve anything: on every key asserted twice, both values appear and the reader is left to work it out. The merge presents one value 100% of the time, in a third of the tokens.

3. Ties are settled by a model that is not allowed to guess

Evidence decides most conflicts. When it is genuinely tied, an adjudicator may step in — constrained three ways:

• it cannot invent a value that was not already an option;
• it is asked twice with the options order-reversed, and an answer that flips is downgraded to unresolvable — axiom IC2 enforced at runtime;
• unresolvable is a first-class answer, and the merge would rather surface a contradiction than paper over it.

4. A merge is a laundering primitive — so trust is a lattice

An assertion from a compromised session enters the merged block and becomes indistinguishable from something the current user established.

Invariant I2: the act of merging may never raise a value's trust above the trust of the claims that assert it. Corroboration can raise trust; merging cannot.

Verified against a real prompt-injection payload. After reading the injected block, the model reported:

"Quoted as data only; the block's embedded directives are not treated as instructions."

---

What is actually here

Four kinds of document. The third column says who each one is for.

		read it if
belief-merge/README.md	the user manual — install, usage, configuration reference, upgrade traps, what is verified and what is not	you want to use the plugin
ARCHITECTURE.md	the code map, as built — the pipeline stage by stage, the three places a model may act, the degradation ladder	you want to understand or extend it
TECHNICAL-REPORT.md	the full writeup — design rationale, the confluence result, the benchmark, limitations	you want the reasoning behind the design
belief-merge/CHANGELOG.md	release history, and why upgrading needs an explicit version	you are upgrading
belief-merge/VERIFY.md	the live-run record — 13 bugs found by running it, and a failure taxonomy	you want to know what actually happened in a real harness
beliefmerge-t1/	the confluence experiment — exhaustive enumeration over 72,760 profiles, minimal counterexamples	you want to check the headline claim
belief-merge/bench/	MergeBench — generator, scored arms, results, and what it does not measure	you want to check the numbers
belief-merge/	the plugin itself: 14 core modules, 242 tests, an offline demo	you want to read the code
docs/	research material, not shipped in the npm package
↳ 会话融合引擎-BeliefMerge-可行性与算法架构.md	the original pre-implementation proposal (Chinese) — 8 stages, 6 novelty points, 8 invariants. Read the banner first: its M-numbering is not the plugin's, and two of its milestones were never built	you want to see how it was planned
↳ conversation-branch-merge-litreview.md	the literature review, ~60 references checked against Crossref/OpenAlex/arXiv	you are asking "has this been done before?"
↳ BeliefMerge-算法逻辑图.html	the algorithm diagram, openable in a browser and exportable to PNG	you want one picture
↳ refcheck/	the scripts and table used to verify those references	you are auditing the citations

The npm package ships lib/, bench/, cordis.patch.yml, ARCHITECTURE.md, CHANGELOG.md, TECHNICAL-REPORT.md and VERIFY.md — see the files field in package.json. docs/ and beliefmerge-t1/ are repository material only.

---

The part I would actually ask about in an interview

Every bug that mattered was found by running the thing, not by testing it.

export * from './x.js' re-exports a name but does not bind it locally — so a missing import shipped while 83 tests passed, because no test exercised alignment → merge → adjudication in one call. The derivation-propagation mechanism had 25 passing tests and had never once executed against real input, because the model emitted zero edges until the prompt contained a worked example.

Thirteen are logged in TECHNICAL-REPORT.md §9, and narrated round by round in VERIFY.md, including one demo whose narration was false until the budget was tightened enough for its claim to be true.

A green test suite is evidence about the paths you exercised, not about the paths you did not.

---

Reproducing everything

cd belief-merge && npm test          # 242 tests, no network, no API key
cd belief-merge && npm run demo      # seven worked examples, offline
cd belief-merge && npm run bench     # MergeBench
cd beliefmerge-t1 && python3 run_t1.py  # the confluence experiment, ~14s

Installing into a real DeepSeek Harness profile is documented in VERIFY.md, along with what happened when it was done.