martin/meeting-lab
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Document controlled rejection V1 baseline

Browse Source
This commit is contained in:
admin
2026-08-20 13:19:54 +02:00
parent 0d4b426021
commit 8ca62fbd92
5 changed files with 264 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/experiments.md
+62
View File
@@ -1962,6 +1962,68 @@ rejection, production integration or beginning cross-pattern reconciliation.
Artifacts are preserved under
`artifacts/experiments/negative_act_form_v0/20260820_qwen35_9b_single_run/`.
## EXP-0036 — Controlled Rejection Derivation V1
Status: Experimental; architecturally unsuccessful
Date: 2026-08-20
This isolated experiment followed the failed binary rejection baseline
(EXP-0034) and successful Negative Act Form classification (EXP-0035). Its V1
hypothesis was to classify the negative act first, resolve its local target in
a separate semantic call, and only then derive `explicitly_rejected`
deterministically. It did not modify either predecessor or any accepted Stage-2
pattern, and it has no production integration.
The target recognizer emitted exactly `candidate_observation_id`,
`target_observation_id`, and `normalized_target_text`. Only
`explicit_non_pursuit` was deterministically eligible. Personal preference,
recommendation, temporary non-action, and `none` could never derive rejection,
even with a valid target. Provenance, local membership, ordering, non-empty
target text, and strict non-normative output were additional gates.
`explicitly_rejected` means rejected by the cited evidence only, not a final
decision, topic outcome, permanent state, or closure.
The run reused five exact accepted Negative Act Form outputs and made three new
Negative Act calls plus eight target-resolution calls. All calls used
`qwen3.5:9B`, temperature 0, `think=false`, `num_ctx=16384`,
`num_predict=1024`, no retries, voting, or prompt changes.
| Case | Negative Act expected / actual | Target result | Verdict |
| --- | --- | --- | --- |
| CR-01 | `explicit_non_pursuit` / same | Model returned the string `"null"` as an unknown ID; self-contained target was not linked | FAIL |
| CR-02 | `explicit_non_pursuit` / same | `obs_1`, external solution and continuation preserved | PASS |
| CR-03 | `personal_preference` / same | `obs_1`; eligibility gate prevented rejection | PASS |
| CR-04 | `recommendation` / same | `obs_1`; eligibility gate prevented rejection | PASS |
| CR-05 | `temporary_non_action` / same | `obs_1`; eligibility gate prevented rejection | PASS |
| CR-06 | `none` / same | Null target; final non-rejection was correct, but expected local target was unresolved | FAIL |
| CR-07 | `explicit_non_pursuit` / same | `obs_1`; real-plant and pressure-test scope survived, but normalization remained proposition-like | PARTIAL |
| CR-08 | `explicit_non_pursuit` / same | `obs_1`; real-plant scope preserved and Technikum alternative excluded | PASS |
Result: five PASS, one PARTIAL, two FAIL. All eight Negative Act forms were
correct. There were no false-positive rejections: the valid targets in CR-03,
CR-04, and CR-05 could not override their ineligible forms. There was one
false-negative rejection, CR-01, caused by invalid target output. Target
resolution missed two expected links (invalid CR-01 and null CR-06), so the
wrong/unresolved-target count was two. CR-06 exposed a strategy flaw: a
non-eligible Negative Act form should not be required to pass target resolution
when it cannot derive rejection. Qualifier-loss count was zero. CR-08 isolated
the positive alternative successfully. No individual owner, responsibility,
decision, outcome, topic-closure, or LLM-emitted rejection status appeared.
There were 3 new Negative Act calls, 8 target calls, 5 accepted classification
reuses, zero technical call failures, and one structural target-validation
failure. Aggregate runner time was 11.951 seconds.
Conclusion: negative-act-form gating is promising and successfully contains
the semantic false positives that defeated EXP-0034, but the experiment is not
architecturally successful. The current target-resolution strategy failed the
required self-contained positive CR-01 and unnecessarily evaluated the
ineligible CR-06 path; it is not reliable enough for rejection derivation.
Artifacts are preserved under
`artifacts/experiments/controlled_rejection_v1/20260820_qwen35_9b_single_run/`.
## EXP-0026 — Topic-oriented Discussion Subject reconstruction V2 prototype
Date: 2026-08-11